Showing 223 of 223 papers.
001 · ICLR 2026 oral · Full text + selected visuals inspected
Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer, Duc Le
Writing lesson: Separate efficiency evidence into equal-compute and equal-data comparisons, and keep the same task columns across both. Define the computational unit before presenting a speedup.
Scientific selection: Included for writing analysis only. High-value writing example for controlled efficiency claims; technical fit is latent speech/text sequence compression, not knowledge distillation or WSI analysis. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: Text / sequence models · LLM pretraining
002 · ICLR 2026 oral · Full text + selected visuals inspected
Philipp Becker, Niklas Freymuth, Serge Thilges, Fabian Otto, Gerhard Neumann
Writing lesson: Explain a drop-in method by naming exactly which component changes, then hold surrounding components fixed. Pair a high-level mechanism diagram with an equation-level computation graph and controlled cost measurements.
Scientific selection: Included for writing analysis only. Strong writing and technical-fit candidate for principled RL post-training objectives, but catalog promotion remains a separate decision. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: LLM post-training · AI / reinforcement learning · Loss / objectives
003 · ICLR 2026 oral · Full text + selected visuals inspected
Nina Vesseron, Elsa Cazelles, Alice Le Brigant, Klein Thierry
Writing lesson: Teach an abstract method through a solvable special case, then carry the same geometric vocabulary into the general algorithm. Include cases where the simple baseline already works and where the proposed objective has undesirable behavior.
Scientific selection: Included for writing analysis only. Useful advanced optimal-transport/statistical representation reading; optional priority for the broad ML curriculum rather than direct pathology or KD. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: Math / statistics · Generative / representation · Graph / geometric ML
004 · ICLR 2026 oral · Full text + selected visuals inspected
Kejing Lu, Chuan Xiao, Yoshiharu Ishikawa
Writing lesson: When introducing multiple related operators, preserve a one-to-one mapping between problem, definition, theorem, application and experiment. Put comparison protocol in the table caption, especially when gains are small.
Scientific selection: Included for writing analysis only. Fits probabilistic similarity-search foundations and retrieval infrastructure; medium technical priority, not a learned RAG or pathology model. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: Math / statistics · NLP tasks / evaluation · RAG / tools / agents
005 · ICLR 2026 oral · Full text + selected visuals inspected
Fang Wu, Xu Huang, Weihao Xuan, Zhiwei Zhang, Yijia Xiao, Frank Wan, Xiaomin Li, Bing Hu, Peng Xia, Jure Leskovec, Yejin Choi
Writing lesson: Use a comparison matrix to make a unifying framework falsifiable: specify which settings recover each baseline. Keep theoretical guarantees attached to the assumptions that support them, especially when introducing a practical extension.
Scientific selection: Included for writing analysis only. Relevant multi-opponent preference objective, but prioritize mathematical-claim checking before promoting; external-teacher KD is an appendix extension, not the main demonstrated result. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: LLM post-training · AI / reinforcement learning · Loss / objectives
006 · ICLR 2026 oral · Full text + selected visuals inspected
Gabriel Grand, Valerio Pepe, Joshua B Tenenbaum, Jacob Andreas
Writing lesson: Decompose an agent claim into the ability to obtain information and the ability to use it. Pair formal equations with plain-language intuition, and let results that break the intuitive story remain visible.
Scientific selection: Included for writing analysis only. High-priority writing exemplar for an agent benchmark with human grounding and decomposed mechanisms; technical fit is Bayesian information seeking, not general VLM visual perception. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: RAG / tools / agents · AI / reinforcement learning · NLP tasks / evaluation
007 · ICLR 2026 oral · Full text + selected visuals inspected
Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alex Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, Omar Khattab
Writing lesson: Pair each headline advantage with its comparison unit, an exception-preserving table and a worked artifact showing what the optimizer actually changed.
Scientific selection: Included for writing analysis only. A useful prompt-optimization counterpart to weight-space RL, with module-level feedback, explicit Pareto selection and reproducible search details. Retained for the complete oral writing study; outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: LLM post-training · RAG / tools / agents · Reasoning / verification
008 · ICLR 2026 oral · Full text + selected visuals inspected
Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, Ashima Suvarna, Benjamin Feuer, Leon Liangyu Chen, Zaid Khan, Eric Frankel, Sachin Grover, Caroline Choi, Niklas Muennighoff, Shiye Su, Wanjia Zhao, John Yang, Shreyas Pimpalgaonkar, Kartik sharma, Charlie Ji, Yichuan Deng, Sarah Pratt, Vivek Ramanujan, Jon Saad-Falcon, Stutee Acharya, Jeffrey Li, Achal Dave, Alon Albalak, Kushal Arora, Blake Wulfe, Chinmay Hegde, Greg Durrett, Sewoong Oh, Mohit Bansal, Saadia Gabriel, Aditya Grover, Kai-Wei Chang, Vaishaal Shankar, Aaron Gokaslan, Mike Merrill, Tatsunori Hashimoto, Yejin Choi, Jenia Jitsev, Reinhard Heckel, Maheswaran Sathiamoorthy, Alex Dimakis, Ludwig Schmidt
Writing lesson: Organize a large empirical paper around reproducible decisions, with compact takeaways linked to complete ablations, explicit held-out evaluation and safety checks.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority KD fit: sequence-level reasoning-trace distillation studies data quality, teacher suitability, repeated sampling and verification rather than proposing feature-based KD.
Relevant reading routes: Knowledge distillation · Reasoning / verification · Evaluation / reproducibility
009 · ICLR 2026 oral · Full text + selected visuals inspected
Tal Daniel, Carl Qi, Dan Haramati, Amir Zadeh, Chuan Li, Aviv Tamar, Deepak Pathak, David Held
Writing lesson: For a modular method, align notation, architecture colors, subsection names and implementation snippets, and state what information each evaluation mode may access.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority SSL/CV fit: self-supervised object-centric representation and stochastic video world modeling, with a precisely delimited robotic-control application.
Relevant reading routes: Visual representations · Video / 3D / embodied · Generative / representation
010 · ICLR 2026 oral · Full text + selected visuals inspected
Yaoyu Wang, Hankun Dai, Zhidong Yang, Junmin Xiao, Guangming Tan
Writing lesson: Use a concrete task-level example, a stage-aligned training diagram, and separate correctness/runtime evidence; surface matched controls and deployment costs early enough to support the central claim.
Scientific selection: Included for writing analysis only. Useful systems/RL writing case, but not a priority KD, CV/VLM, SSL, or pathology method; substantial internal inconsistencies need resolution before scientific promotion.
Relevant reading routes: Code / software engineering · AI / reinforcement learning · LLM systems
011 · ICLR 2026 oral · Full text + selected visuals inspected
Nikita Kornilov, David Li, Tikhon Mavrin, Aleksei Leonov, Nikita Gushchin, Evgeny Burnaev, Iaroslav Koshelev, Aleksandr Korotin
Writing lesson: For a unification paper, establish common notation, identify the precise mathematical obstacle, and test a theory-derived hyperparameter prediction under fixed architecture before citing heterogeneous leaderboard results.
Scientific selection: Added to the scientific catalog with a full review. Strong priority fit connecting generative knowledge distillation, flow/diffusion objectives, real-data supervision, and one-step image generation without adding a GAN discriminator.
Relevant reading routes: Knowledge distillation · Generative / representation · Loss / objectives · Generative vision
012 · ICLR 2026 oral · Full text + selected visuals inspected
Justin Deschenaux, Lan Tran, Caglar Gulcehre
Writing lesson: Show the baseline's wasted computation, the exact information-flow change that removes it, and the resulting quality–speed trade-off; preserve training overhead, sampler settings, and diversity checks.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-relevant generative vision architecture with explicit compatibility and experiments for self-distillation through time; useful for efficient masked modeling across images and text.
Relevant reading routes: Generative / representation · Generative vision · Knowledge distillation · Generation / MT evaluation
013 · ICLR 2026 oral · Full text + selected visuals inspected
Zeyi Liao, Jaylen Jones, Linxi Jiang, Yuting Ning, Eric Fosler-Lussier, Yu Su, ZHIQIANG LIN, Huan Sun
Writing lesson: For an agent benchmark, explicitly separate exposure, attempted behavior, successful execution, and environmental navigation; accompany headline rates with denominators, selection rules, and concrete failure categories.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong VLM-agent evaluation fit: screenshot-driven computer use, multimodal attack-surface analysis, and controlled cross-interface robustness testing.
Relevant reading routes: Generation / video / agents · V&L evaluation / safety · LLM evaluation / safety
014 · ICLR 2026 oral · Full text + selected visuals inspected
Peiyan Hu, Haodong Feng, Hongyuan Liu, Tongtong Yan, Wenhao Deng, Tianrun Gao, Rong Zheng, Haoren Zheng, Chenglei Yu, Chuanrui Wang, Kaiwen Li, Zhi-Ming Ma, Dezhi Zhou, Xingcai Lu, Dixia Fan, Tailin Wu
Writing lesson: For a real-data benchmark, make data provenance and shared evaluation splits first-class contributions, and pair ordinary prediction errors with domain-specific diagnostics and long-horizon tests.
Scientific selection: Included for writing analysis only. A strong scientific-ML benchmark-writing case, but fluid/combustion forecasting is not a direct priority KD, CV/VLM, SSL, or computational-pathology addition. Do not classify these physical fields as WSI data.
Relevant reading routes: Evaluation / reproducibility · Transfer / adaptation / continual
015 · ICLR 2026 oral · Full text + selected visuals inspected
Shifeng Xie, Vasilii Feofanov, Jianfeng Zhang, Themis Palpanas, Ievgen Redko
Writing lesson: Pair each motivating data property with a pipeline component, then organize experiments around explicit questions and display overlap, data volume and evaluation protocol beside the results.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A useful synthetic-data companion for SSL curriculum design, although its experiments are time-series classification rather than CV or whole-slide pathology.
Relevant reading routes: Generative / representation · Transfer / adaptation / continual · Causal inference
016 · ICLR 2026 oral · Full text + selected visuals inspected
Aideen Fay, Inés García-Redondo, Qiquan Wang, Haim Dubossarsky, Anthea Monod
Writing lesson: Explain a technical diagnostic with a small worked construction, then keep its sampling unit, control conditions and exception-bearing summary table explicit.
Scientific selection: Included for writing analysis only. Useful representation-analysis writing and safety diagnostics, but not a priority KD, CV/VLM, SSL or pathology method in this addition pass.
Relevant reading routes: LLM interpretability · LLM evaluation / safety · Interpretability
017 · ICLR 2026 oral · Full text + selected visuals inspected
Carl Edwards, Chi Han, Gawon Lee, Thao Nguyen, Sara Szymkuć, Chetan Prasad, Bowen Jin, Jiawei Han, Ying Diao, Ge Liu, Hao Peng, Bartosz Grzybowski, Martin Burke, Heng Ji
Writing lesson: Let the domain constraint determine both representation and evaluation, and diagram exceptions or fallback paths alongside the idealized method.
Scientific selection: Included for writing analysis only. Molecular generation offers transferable multimodal-tokenization writing, but chemical graphs are not natural-image VLM or whole-slide pathology data.
Relevant reading routes: Graph / geometric ML · Text / sequence models · Generative / representation
018 · ICLR 2026 oral · Full text + selected visuals inspected
Jing Yao, Shitong Duan, Xiaoyuan Yi, Dongkuan Xu, Peng Zhang, Tun Lu, Ning Gu, Zhicheng Dou, Xing Xie
Writing lesson: For a new evaluation instrument, organize evidence around what it measures, whether it responds to controlled interventions, and how stable its results are—not only whether it separates models.
Scientific selection: Included for writing analysis only. Relevant dynamic-benchmark methodology, but not a priority KD, CV/VLM, SSL or pathology addition.
Relevant reading routes: LLM evaluation / safety · Evaluation / reproducibility
019 · ICLR 2026 oral · Full text + selected visuals inspected
Bryon Tjanaka, Henry Chen, Matthew Fontaine, Stefanos Nikolaidis
Writing lesson: Explain the exact failure mechanism with a matched before/after schematic, then pair effectiveness metrics with resource costs and an explicit comparability note.
Scientific selection: Candidate for a later scientific review; not added by this writing study. An optional CV-generation search companion; less direct than the priority distillation and SSL candidates.
Relevant reading routes: Generative vision · Generative / representation
020 · ICLR 2026 oral · Full text + selected visuals inspected
Wei Zhuo, Siqiang Luo
Writing lesson: Use explicit criteria and train/freeze diagrams to define an adaptation method, then make every inference-time operation appear in the executable pseudocode.
Scientific selection: Included for writing analysis only. Useful graph-transfer presentation but not a direct priority-domain method; material mathematical and implementation ambiguities warrant caution before scientific promotion.
Relevant reading routes: Graph / geometric ML · Transfer / adaptation / continual
021 · ICLR 2026 oral · Full text + selected visuals inspected
Fengyu Cai, Tong Chen, Xinran Zhao, Sihao Chen, Hongming Zhang, Sherry Wu, Iryna Gurevych, Heinz Koeppl
Writing lesson: Pair an intuitive reformulation with an explicit information-flow diagram, then isolate its contribution using matched backbone and data controls.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Self-supervised retrieval is relevant to representation learning, but less directly aligned with the current CV/VLM and KD priorities.
Relevant reading routes: NLP tasks / evaluation · Generative / representation
022 · ICLR 2026 oral · Full text + selected visuals inspected
Zongrui Li, Xinhua Ma, Minghui HU, Yunqing Zhao, Yingchen Yu, Qian Zheng, Chang Liu, Jiang Xudong, Song Bai
Writing lesson: Show the failure in the representation the application actually uses, then connect a visually explicit intermediate target to an analytic final estimator.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong CV candidate: a supervised target reformulation and video-prior transfer for inverse geometry, not knowledge distillation or pathology.
Relevant reading routes: 3D / spatial vision · Restoration / inverse problems · Generative vision
023 · ICLR 2026 oral · Full text + selected visuals inspected
Laiguo Yin, Yixin Zhang, YUQING SUN, Lizhen Cui
Writing lesson: Separate a reusable representation's construction, standalone evaluation and downstream re-ranking, and include negative cases that reveal when semantic transfer fails.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant VLM representation candidate with a prototype contrastive objective, but scientific inclusion should preserve external supervision and theoretical caveats.
Relevant reading routes: Multimodal retrieval · Image–text representation · Loss / objectives
024 · ICLR 2026 oral · Full text + selected visuals inspected
Seul Lee, Karsten Kreis, Srimukh Veccham, Meng Liu, Danny Reidenbach, Saee Paliwal, Weili Nie, Arash Vahdat
Writing lesson: Explain a structured generative method at three levels—tree operation, sequence representation and executable search—and measure the quality–compute trade-off.
Scientific selection: Included for writing analysis only. A useful discrete-flow and structured-search writing example, but chemical synthesis is outside the current prioritized CV/VLM, SSL and KD additions.
Relevant reading routes: Generative / representation
025 · ICLR 2026 oral · Full text + selected visuals inspected
Alexi Gladstone, Ganesh Nanduru, Md Mofijul Islam, Peixuan Han, Hyeonjeong Ha, Aman Chadha, Yilun Du, Heng Ji, Jundong Li, Tariq Iqbal
Writing lesson: Use one visual mechanism across modalities, separate training and inference algorithms, and report absolute metrics plus actual computation alongside scaling slopes.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong self-supervised/CV candidate for energy-based iterative prediction, provided preliminary scale, compute overhead and modality limitations remain explicit.
Relevant reading routes: Generative / representation · Visual representations · Restoration / inverse problems
026 · ICLR 2026 oral · Full text + selected visuals inspected
Johnny Wei, Ameya Godbole, Mohammad Aflah Khan, Ryan Wang, Xiaoyuan Zhu, James Flemings, Nitya Kashyap, Krishna Gummadi, Willie Neiswanger, Robin Jia
Writing lesson: For a benchmark or model-suite paper, connect each artifact variation to a causal question, make interventions inspectable, and publish invalidated analyses as well as favorable findings.
Scientific selection: Included for writing analysis only. Excellent controlled-evaluation writing study, but memorization auditing is outside the current prioritized KD/CV/VLM/SSL scientific additions.
Relevant reading routes: LLM evaluation / safety · Evaluation / reproducibility · Privacy / federated
027 · ICLR 2026 oral · Full text + selected visuals inspected
Hongjian Jiang, Michael Hahn, Georg Zetzsche, Anthony W. Lin
Writing lesson: For a theoretical paper, organize intermediate cases around the obstacle each removes, reuse one concrete running example, and distinguish theorem assumptions from experimental implementations.
Scientific selection: Included for writing analysis only. Useful theory-writing example, but not a priority KD, CV/VLM, SSL or pathology-method addition.
Relevant reading routes: Transformers / pretraining · Reasoning / verification · Math / statistics
028 · ICLR 2026 oral · Full text + selected visuals inspected
Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens
Writing lesson: For a software-method paper, explain one unifying semantic rule, demonstrate equivalent code side by side, and publish representative outputs and failure messages as evidence of usability.
Scientific selection: Included for writing analysis only. Excellent software-writing example, but a tensor notation tool rather than a priority KD/CV/VLM/SSL method.
Relevant reading routes: Neural networks · LLM systems · Evaluation / reproducibility
029 · ICLR 2026 oral · Full text + selected visuals inspected
Yingjie Ma, Bijal Bharadva, Xin Zhang, Joann Qiongna Chen
Writing lesson: Use challenge-to-component signposting and separate privacy threat models visibly; qualify the narrative against every metric rather than converting mixed evidence into a uniform success story.
Scientific selection: Included for writing analysis only. Mobility-trajectory privacy is outside the selected scientific scope; substantive accounting and claim-evidence caveats also argue against unqualified scientific promotion.
Relevant reading routes: Privacy / federated · AI / reinforcement learning · Generative / representation
030 · ICLR 2026 oral · Full text + selected visuals inspected
Ang Lv, Jin Ma, Yiyuan Ma, Siyuan Qiao
Writing lesson: Pair each design principle with an observable mechanism, then use ablations to distinguish that mechanism from plausible cheaper explanations; report both analytic complexity and actual distributed throughput.
Scientific selection: Included for writing analysis only. Relevant auxiliary-loss/MoE writing example, but evaluated on language pretraining rather than the current KD, CV/VLM, SSL or pathology priorities.
Relevant reading routes: LLM systems · LLM pretraining · Loss / objectives · Interpretability
031 · ICLR 2026 oral · Full text + selected visuals inspected
Tommaso Mencattini, Riccardo Cadei, Francesco Locatello
Writing lesson: Build a method narrative around an explicit failure mechanism, then reuse the same visual comparison grammar to test whether the proposed remedy actually addresses it.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong causal-discovery and representation-interpretability fit; check existing coverage before scientific promotion.
Relevant reading routes: Causal inference · Interpretability
032 · ICLR 2026 oral · Full text + selected visuals inspected
Lei Wang, Yu Cheng, Yining Shi, Zhiwen Mo, Zhengju Tang, Wenhao Xie, Tong Wu, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang
Writing lesson: Choose an executable running example and preserve its terminology across API explanation, intermediate representation, optimization mechanism, and empirical evaluation.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Useful systems-level companion to efficient attention/inference; confirm the closest existing systems category before promotion.
Relevant reading routes: LLM systems
033 · ICLR 2026 oral · Full text + selected visuals inspected
Blake Gaines, Jinbo Bi
Writing lesson: For theoretical papers, introduce one concrete object through multiple linked views and visibly distinguish theorems from empirical observations.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A useful advanced foundation on representation geometry; promotion should preserve architectural assumptions.
Relevant reading routes: Neural networks · Interpretability
034 · ICLR 2026 oral · Full text + selected visuals inspected
Aaradhya Pandey, Arnab Auddy, Haolin Zou, Arian Maleki, Sanjeev Kulkarni
Writing lesson: When changing an evaluation or guarantee, first demonstrate the specific failure of the old formulation and explicitly trace the new definition through theorem, experiment, and practical calibration.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A substantial theoretical unlearning contribution suited to an advanced privacy route, not an off-the-shelf LLM deletion guarantee.
Relevant reading routes: Privacy / federated · Math / statistics
035 · ICLR 2026 oral · Full text + selected visuals inspected
Roy Betser, Eyal Gofer, Meir Yossef Levi, Guy Gilboa
Writing lesson: Separate alternative explanatory routes and their assumptions, and choose empirical diagnostics that map one-to-one onto the theoretical ingredients.
Scientific selection: Added to the scientific catalog with a full review. Selected and integrated as infonce-gaussian-2026 with bilingual review, argument graph, Table 1, and quiz; links contrastive-learning theory with loss and representation routes.
Relevant reading routes: Loss / objectives · Visual representations
036 · ICLR 2026 oral · Full text + selected visuals inspected
Yuanchen Ju, Yongyuan Liang, Yen-Jen Wang, Gireesh Nandiraju, Yuanliang Ju, Seungjae (Jay) Lee, Qiao Gu, Elvis Hsieh, Furong Huang, Koushil Sreenath
Writing lesson: Use motivating controlled comparisons to justify each representation choice, then keep the representation, dataset, model, benchmark, and execution layer terminologically distinct.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Direct fit for structured visual reasoning and embodied planning; preserve the separation between graph planning and low-level control.
Relevant reading routes: Generation / video / agents · Video / 3D / embodied
037 · ICLR 2026 oral · Full text + selected visuals inspected
Hao Tan, jun lan, Zichang Tan, Senyuan Shi, Ajian Liu, Chuanbiao Song, Huijia Zhu, Weiqiang Wang, Jun Wan, Zhen Lei
Writing lesson: Build the evaluation taxonomy around the claimed deployment gap, then align each training stage with a distinct ablation and preserve the exact averaging protocol in table captions.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Useful case study of preference-aligned visual forensics and OOD protocol design, with clearly bounded image-level evidence.
Relevant reading routes: Vision evaluation · V&L evaluation / safety · LLM post-training
038 · ICLR 2026 oral · Full text + selected visuals inspected
Shibo Hong, jiahao ying, Haiyuan Liang, Mengdi Zhang, Jun Kuang, Jiazheng Zhang, Yixin Cao
Writing lesson: Define each generalization axis and label source before presenting results, and publish the operational rubric—not just the names of evaluation dimensions.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority VLM evaluation resource spanning understanding, generation and interleaved output, with useful human-versus-teacher supervision caveats.
Relevant reading routes: V&L evaluation / safety · LLM evaluation / safety
039 · ICLR 2026 oral · Full text + selected visuals inspected
Zhuoran Jin, Hongbang Yuan, Kejian Zhu, Jiachun Li, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao
Writing lesson: Use the same small set of conceptual deliverables in the introduction, headings, diagrams and experimental questions; expose missing capabilities alongside scores.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority VLM reward-modeling example connecting free-form criteria, multi-task training, model architecture and evaluation limitations.
Relevant reading routes: V&L evaluation / safety · LLM post-training
040 · ICLR 2026 oral · Full text + selected visuals inspected
Tao Ren, Zishi Zhang, Jinyang Jiang, Zehao Li, Shentao Qin, Yi Zheng, Guanghao Li, Qianyou Sun, Yan Li, Jiafeng Liang, Xinping Li, Yijie Peng
Writing lesson: Make an optimizer's tradeoff concrete with one computation-graph legend and a table separating peak device memory, host memory, step time and time to a common target.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A relevant visual-generation optimization method with a concrete hybrid-gradient design and unusually useful resource ablations; not a knowledge-distillation method.
Relevant reading routes: Generative vision · Loss / objectives
041 · ICLR 2026 oral · Full text + selected visuals inspected
Rohit Jena, Vedant Zope, Pratik A Chaudhari, James Gee
Writing lesson: Build a systems paper around a measured bottleneck, show the memory movement visually, and couple each scale claim to accuracy, hardware, approximation flags and communication cost.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority adjacent infrastructure for very large biomedical images; include in pathology engineering only as a transferable registration technique, clearly not a validated WSI benchmark.
Relevant reading routes: Restoration / inverse problems · WSI engineering
042 · ICLR 2026 oral · Full text + selected visuals inspected
Zheng Huang, Enpei Zhang, Weikang Qiu, Yinghao Cai, Carl Yang, Elynn Chen, Xiang Zhang, Rex Ying, Dawei Zhou, Yujun Yan
Writing lesson: Use a motivating representation comparison to justify architecture, then make training-only supervision visibly distinct from information available at inference.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A distinctive CV/VLM cross-modal reconstruction method with object-compositional text supervision, useful as adjacent representation-learning guidance rather than a pathology method.
Relevant reading routes: Generative vision · Image–text representation
043 · ICLR 2026 oral · Full text + selected visuals inspected
Isaac Han, Sangyeon Park, Seungwon Oh, Donghu Kim, Hojoon Lee, KyungJoong Kim
Writing lesson: Turn an intuitive tradeoff into explicit proxies, state the regime where those proxies are useful, and show intervention-time behavior rather than relying only on final scores.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority visual representation/adaptation method with a cheap geometry-aware reset; relevant to continually updated encoders, but not itself distillation or a pathology evaluation.
Relevant reading routes: Transfer / adaptation / continual · Visual representations
044 · ICLR 2026 oral · Full text + selected visuals inspected
Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, Alexandre Alahi
Writing lesson: For a feedback-based method, use one closed-loop figure and include an adversarially simple baseline that can game the headline metric.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority video/visual generation method with model-error replay, residual error estimation and flow-matching adaptation; adjacent to residual-learning interests but not residual KD.
Relevant reading routes: Generative vision · Video / 3D / embodied · Loss / objectives
045 · ICLR 2026 oral · Full text + selected visuals inspected
Pratyaksh Sharma, Alexandra M Proca, Lucas Prieto, Pedro Mediano
Writing lesson: A theory-driven paper can be taught through a controlled ladder of models and a shared visual vocabulary, with each theorem term linked to a measurable geometric behavior.
Scientific selection: Included for writing analysis only. Strong explanatory-writing and representation-geometry case study, but not a direct KD, CV/VLM, SSL or pathology method under this update's scientific priorities.
Relevant reading routes: Interpretability · Text / sequence models
046 · ICLR 2026 oral · Full text + selected visuals inspected
Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, Ivan Vulić
Writing lesson: When proposing a new reasoning representation, visualize the actual information flow and test the proposed explanation through an initialization ablation, not only a leaderboard comparison.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority visual reasoning/planning paper with a genuinely image-based autoregressive policy and concrete RL supervision; useful alongside multimodal reasoning routes.
Relevant reading routes: Generation / video / agents · Video / 3D / embodied
047 · ICLR 2026 oral · Full text + selected visuals inspected
Wayne Chi, Valerie Chen, Ryan Shar, Aditya Mittal, Jenny Liang, Wei-Lin Chiang, Anastasios Angelopoulos, Ion Stoica, Graham Neubig, Ameet Talwalkar, Chris Donahue
Writing lesson: For an in-the-wild benchmark, show the actual collection interaction, rejected examples and a controlled context ablation; these establish what the benchmark measures beyond its headline score.
Scientific selection: Included for writing analysis only. Useful benchmark-writing case, but code editing is outside the selected KD/CV/VLM/SSL/pathology promotion priorities.
Relevant reading routes: Code / software engineering · LLM evaluation / safety
048 · ICLR 2026 oral · Full text + selected visuals inspected
Jinchuan Tian, Sang-gil Lee, Zhifeng Kong, Sreyan Ghosh, Arushi Goel, Chao-Han Huck Yang, Wenliang Dai, Zihan Liu, Hanrong Ye, Shinji Watanabe, Mohammad Shoeybi, Bryan Catanzaro, Rafael Valle, Wei Ping
Writing lesson: For a multi-stage system, align the challenge list, model names, training diagram and experimental subsections; show an intermediate representation before discussing higher-level behavior.
Scientific selection: Included for writing analysis only. An informative multimodal systems-writing example, but audio-only generation is outside the current scientific promotion priorities.
Relevant reading routes: LLMs · LLM post-training · Generative / representation
049 · ICLR 2026 oral · Full text + selected visuals inspected
Yuxuan Zhou, Fei Huang, Heng Li, Fengyi Wu, Tianyu Wang, jianwei zhang, Junyang Lin, Zhi-Qi Cheng
Writing lesson: Present a theoretically simple but impractical algorithm first, then show the exact modification that removes its overhead; report acceptance and measured latency separately.
Scientific selection: Included for writing analysis only. Valuable theory-to-systems writing case; speculative verification itself is not a selected KD/CV/VLM/SSL/pathology addition.
Relevant reading routes: LLM systems
050 · ICLR 2026 oral · Full text + selected visuals inspected
Chinmay Datar, Taniya Kapoor, Abhishek Chandra, Qing Sun, Erik L. Bolager, Iryna Burak, Anna Veselovska, Massimo Fornasier, Felix Dietrich
Writing lesson: Organize numerical experiments around the failure modes a method is intended to solve, and compare runtime only alongside explicit accuracy targets and conventional numerical baselines.
Scientific selection: Included for writing analysis only. Strong scientific-computing writing example, but this is a forward PDE solver rather than a selected KD/CV/VLM/SSL/pathology method.
Relevant reading routes: Neural networks · Evaluation / reproducibility
051 · ICLR 2026 oral · Full text + selected visuals inspected
Madhav Kanda, Shubham Dipak Ugare, Sasa Misailovic
Writing lesson: For generated scientific code, report executable validity and scientific diagnostics separately, and show an example that moves through rejection, inference and refinement.
Scientific selection: Included for writing analysis only. Relevant to reliable scientific-program synthesis but outside the selected scientific promotion priorities.
Relevant reading routes: Code / software engineering · Reasoning / verification · Math / statistics
052 · ICLR 2026 oral · Full text + selected visuals inspected
Kairong Luo, Zhenbo Sun, Haodong Wen, Xinyu Shi, Jiarui Cui, Chenyi Dang, Kaifeng Lyu, Wenguang Chen
Writing lesson: When a familiar technique appears ineffective, organize the paper around a confounding interaction and vary its components independently before presenting a combined recipe.
Scientific selection: Included for writing analysis only. Useful controlled-optimization writing study, but text-only pretraining is not among the currently prioritized scientific additions.
Relevant reading routes: LLM pretraining · Transfer / adaptation / continual · Evaluation / reproducibility
053 · ICLR 2026 oral · Full text + selected visuals inspected
Kaiwen Zheng, Huayu Chen, Haotian Ye, Haoxiang Wang, Qinsheng Zhang, Kai Jiang, Hang Su, Stefano Ermon, Jun Zhu, Ming-Yu Liu
Writing lesson: Make the training/evaluation boundary visible in the table itself, and distinguish idealized objective derivations, practical stabilizers and measured wall-clock gains.
Scientific selection: Candidate for a later scientific review; not added by this writing study. High-priority candidate for visual generation and flow-matching objective design; selection should account for existing DiffusionNFT coverage.
Relevant reading routes: Generative vision · Generative / representation · Loss / objectives
054 · ICLR 2026 oral · Full text + selected visuals inspected
Parham Rezaei, Filip Kovačević, Francesco Locatello, Marco Mondelli
Writing lesson: Connect each theoretical insight to a controlled empirical intervention, and organize extra experiments by the alternative explanation they test.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Directly evaluates synthetic subset selection for ResNet and transformer distillation; microscopy evidence is a small RxRx1 perturbation-classification experiment, not WSI.
Relevant reading routes: Image / recognition · Knowledge distillation · Math / statistics
055 · ICLR 2026 oral · Full text + selected visuals inspected
Marco Fumero, Luca Moschella, Emanuele Rodolà, Francesco Locatello
Writing lesson: Pair a low-dimensional visual explanation with application-level evidence, and repeat a compact setup–analysis–takeaway structure.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Interesting empirical autoencoder/foundation-model probe, but theoretical issues should be resolved before a strong scientific recommendation; pathology appears only as patch reconstruction, not WSI validation.
Relevant reading routes: Interpretability · Visual representations · Vision evaluation
056 · ICLR 2026 oral · Full text + selected visuals inspected
Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville
Writing lesson: Design a control that isolates the proposed failure mechanism, then teach any new metric visually before presenting its cross-model distribution.
Scientific selection: Included for writing analysis only. Strong evaluation-writing example, but text-only conversation reliability is outside the current KD/CV/VLM/SSL/pathology promotion priority.
Relevant reading routes: LLM evaluation / safety · Generation / MT evaluation
057 · ICLR 2026 oral · Full text + selected visuals inspected
Noah Golowich, Allen Liu, Abhishek Shetty
Writing lesson: Make the key mathematical object answer an intuitive question, then demonstrate one concrete consequence with a control that would fail if the proposed distinction did not matter.
Scientific selection: Included for writing analysis only. Valuable theory–experiment narrative, but output-logit structure is not a KD method and is outside the prioritized scientific promotion scope.
Relevant reading routes: LLM interpretability · Text / sequence models · Math / statistics
058 · ICLR 2026 oral · Full text + selected visuals inspected
Ran Xu, Yuchen Zhuang, Yishan Zhong, Yue Yu, Zifeng Wang, Xiangru Tang, Hang Wu, May Dongmei Wang, Peifeng Ruan, Donghan Yang, Tao Wang, Guanghua Xiao, Xin Liu, Carl Yang, Yang Xie, Wenqi Shi
Writing lesson: For an environment paper, align each claimed capability with its own evidence: execution validity, training utility, generalization, cost and failure analysis.
Scientific selection: Included for writing analysis only. Biomedical code-agent infrastructure, mainly text/EHR/tabular tasks; no WSI, visual-model or priority KD/SSL evaluation. Retain as a writing and benchmark-design example rather than categorizing it as computational pathology.
Relevant reading routes: Code / software engineering · RAG / tools / agents · LLM post-training
059 · ICLR 2026 oral · Full text + selected visuals inspected
Nan Huo, Xiaohan Xu, Jinyang Li, Per Jacobsson, Shipei Lin, Bowen Qin, Binyuan Hui, Xiaolong Li, Ge Qu, Shuzheng Si, Linheng Han, Edward Alexander, Xintong Zhu, Rui Qin, Ruihan Yu, Yiyao Jin, Feige Zhou, Weihao Zhong, Yun Chen, Hongyu Liu, Chenhao Ma, Fatma Ozcan, Yannis Papakonstantinou, Reynold Cheng
Writing lesson: Define interactive protocols with a control-flow figure and a caption that explains scoring dependencies; then test the simulator itself rather than treating it as a transparent evaluator.
Scientific selection: Included for writing analysis only. Useful benchmark-writing example, but text-to-SQL interaction is outside the selected KD, CV/VLM, SSL and pathology scientific scope.
Relevant reading routes: RAG / tools / agents · Code / software engineering · LLM evaluation / safety
060 · ICLR 2026 oral · Full text + selected visuals inspected
Jatan Shrestha, Santeri Heiskanen, Kari Hepola, Severi Rissanen, Pekka Jääskeläinen, Joni Pajarinen
Writing lesson: Use a small geometric figure for a sampling procedure, then pair aggregate rankings with explicit failure-case diagnostics and complete per-task tables.
Scientific selection: Included for writing analysis only. General offline multi-objective design optimization, with no KD/SSL or direct visual representation method; retain as a useful writing and generative-optimization reference outside this promotion round.
Relevant reading routes: Generative / representation · Loss / objectives
061 · ICLR 2026 oral · Full text + selected visuals inspected
Zhiyu Mou, Yiqin Lv, Miao Xu, Qi Wang, Yixiu Mao, Jinghao Chen, Qichen Ye, Chao Li, Rongquan Bai, Chuan Yu, Jian Xu, Bo Zheng
Writing lesson: Connect each proposed mechanism to a named failure, a mathematical or operational justification, and a dedicated falsification experiment; give readers an appendix map.
Scientific selection: Included for writing analysis only. Useful theory-to-industrial-evidence writing example; specialized advertising policy optimization rather than a direct vision, language or WSI contribution.
Relevant reading routes: AI / reinforcement learning
062 · ICLR 2026 oral · Full text + selected visuals inspected
Brandon Annesi, Dario Bocchi, Chiara Cammarota
Writing lesson: Teach a difficult proof technique on a minimal example, distinguish asymptotic mathematics from finite-system evidence, and keep counterexamples adjacent to the main trend.
Scientific selection: Included for writing analysis only. Valuable theory-writing example on overparameterization and Hessian geometry, but outside the current KD, CV/VLM, SSL and pathology promotion priorities.
Relevant reading routes: Neural networks · Math / statistics
063 · ICLR 2026 oral · Full text + selected visuals inspected
Junlin Han, Shengbang Tong, David Fan, Yufan Ren, Koustuv Sinha, Philip Torr, Filippos Kokkinos
Writing lesson: Use finding-led sections for a multi-experiment study, make controls explicit, and validate a derived recipe at a new scale while retaining negative qualitative examples.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-fit candidate with direct evidence on language data mixtures, multimodal transfer, evaluation shortcuts and scale-sensitive visual perception.
Relevant reading routes: LLM pretraining · Multimodal LLMs · V&L evaluation / safety
064 · ICLR 2026 oral · Full text + selected visuals inspected
Yixian Xu, Yusong Wang, Shengjie Luo, Kaiyuan Gao, Tianyu He, Di He, Chang Liu
Writing lesson: Pair a small conceptual figure with a mechanistic comparison table, then present empirical results with explicit baseline provenance and evaluation-tool versions.
Scientific selection: Included for writing analysis only. Geometric generative modeling for molecules rather than current priority domains; retain the SDE sign discrepancy for author/code clarification before any later implementation-level promotion. This is not WSI pathology.
Relevant reading routes: Generative / representation · Graph / geometric ML
065 · ICLR 2026 oral · Full text + selected visuals inspected
Shuyang Jiang, Yusheng Liao, Ya Zhang, Yanfeng Wang, Yu Wang
Writing lesson: Show the training credit-assignment mechanism with a running example, then report both quality and resource costs plus a dedicated proxy-quality check.
Scientific selection: Included for writing analysis only. Useful token-level reward and reasoning-efficiency writing example, outside current priority domains; the base model's Distill name does not make this a new knowledge-distillation algorithm.
Relevant reading routes: Reasoning / verification · LLM post-training · Loss / objectives
066 · ICLR 2026 oral · Full text + selected visuals inspected
Zheng Zhao, Yeskendir Koishekenov, Xianjun Yang, Naila Murray, Nicola Cancedda
Writing lesson: For an analysis instrument, state the access and compute assumptions early, disclose failed transfer, and distinguish a successful intervention example from a generally validated repair method.
Scientific selection: Included for writing analysis only. Substantive LLM interpretability example but no tested KD, CV/VLM, SSL or pathology method; keep in the complete oral writing corpus without priority catalog promotion.
Relevant reading routes: LLM interpretability · Reasoning / verification
067 · ICLR 2026 oral · Full text + selected visuals inspected
Vinod Raman, Travis Dick, Matthew Joseph
Writing lesson: Explain why the chosen objective matters, state impossibility and assumptions before guarantees, and label benchmark baselines that operate under stronger information access.
Scientific selection: Included for writing analysis only. Privacy-theory and metric-design writing example, not a tested KD, CV/VLM, SSL or pathology method within the current promotion priorities.
Relevant reading routes: Privacy / federated · Math / statistics
068 · ICLR 2026 oral · Full text + selected visuals inspected
Jianwen Jiang, Weihong Zeng, Zerong Zheng, Jiaqi Yang, Chao Liang, Wang Liao, Han Liang, Weifeng Chen, XING WANG, Yuan Zhang, Mingyuan Gao
Writing lesson: Explain the conditioning failure before the architecture, visualize train/test differences, and triangulate proxy scores, human judgments and explicit failure cases.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-fit candidate connecting multimodal semantic planning to controllable avatar-video synthesis; distinct conditioning and warm-up mechanisms support a substantive review.
Relevant reading routes: Generative vision · Generation / video / agents
069 · ICLR 2026 oral · Full text + selected visuals inspected
Yu Guoqi, Juncheng Wang, Chen Yang, Jing Qin, Angelica Aviles-Rivero, Shujun Wang
Writing lesson: Present the reusable module and complete network at separate scales, attach dimensions to diagrams, and identify the independent evaluation unit before reporting sample totals.
Scientific selection: Included for writing analysis only. Useful architecture and subject-level evaluation example, but EEG/ECG and wearable signals are not WSI computational pathology, nor a tested priority CV/VLM/KD/SSL contribution.
Relevant reading routes: Neural networks · Evaluation / reproducibility
070 · ICLR 2026 oral · Full text + selected visuals inspected
Konwoo Kim, Suhas Kotha, Percy Liang, Tatsunori Hashimoto
Writing lesson: Distinguish observed points, fitted curves and asymptotic predictions visually; connect each methodological addition to a demonstrated failure; use a matched-budget ablation for the key mechanism.
Scientific selection: Added to the scientific catalog with a full review. Direct priority KD evidence: compress independent data-constrained ensembles and improve same-size students through sequence distillation plus retained real training data.
Relevant reading routes: Knowledge distillation · LLM pretraining
071 · ICLR 2026 oral · Full text + selected visuals inspected
Federico Danieli, Pau Rodriguez, Miguel Sarabia, Xavier Suau, Luca Zappella
Writing lesson: For systems papers, connect mathematical reformulation to hardware implementation, then evaluate correctness, runtime and task quality as distinct claims with explicit comparison units.
Scientific selection: Included for writing analysis only. Substantive sequence-model systems contribution; its demonstrated application is language-model recurrence parallelization rather than KD, CV/VLM, SSL or pathology.
Relevant reading routes: Text / sequence models · LLM systems
072 · ICLR 2026 oral · Full text + selected visuals inspected
Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Jiahao Wang, Vishal Patel, Paul Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, Jieneng Chen
Writing lesson: Define the downstream decision a benchmark should improve, visualize the evaluation loop, and expose control/input differences beside scores rather than hiding them in implementation details.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-fit CV/VLM candidate connecting generated visual futures to closed-loop recognition, navigation, question answering and manipulation outcomes.
Relevant reading routes: Video / 3D / embodied · Generation / video / agents · V&L evaluation / safety
073 · ICLR 2026 oral · Full text + selected visuals inspected
Guojian Zhan, Letian Tao, Pengcheng Wang, Yixiao Wang, Yuxin Chen, Yiheng Li, Hongyang Li, Masayoshi Tomizuka, Shengbo Li
Writing lesson: Separate target construction from target fitting; align each claimed mechanism with its own ablation, and report policy quality and deployment cost as independent outcomes.
Scientific selection: Included for writing analysis only. Relevant mean-flow RL and boundary-loss example, but the demonstrated contribution is generative control rather than a priority visual representation, KD or pathology method; real-robot validation remains future work.
Relevant reading routes: AI / reinforcement learning · Loss / objectives
074 · ICLR 2026 oral · Full text + selected visuals inspected
Wenyi Wang, Piotr Piękos, Li Nanbo, Firas Laakom, Yimeng Chen, Mateusz Ostaszewski, Mingchen Zhuge, Jürgen Schmidhuber
Writing lesson: Define the surrogate-target mismatch, test the surrogate without direct target leakage, and separate theorem assumptions, heuristic implementation and transfer evidence into distinct claims.
Scientific selection: Included for writing analysis only. Useful non-priority example of agentic search, idealization boundaries and benchmark claims; it edits coding-agent scaffolds rather than visual models, KD objectives, SSL representations or pathology systems.
Relevant reading routes: Code / software engineering · RAG / tools / agents
075 · ICLR 2026 oral · Full text + selected visuals inspected
Jing Gu, Xian Liu, Yu Zeng, Ashwin Nagarajan, Fangrui Zhu, Daniel Hong, Yue Fan, Qianqi Yan, Kaiwen Zhou, Ming-Yu Liu, Xin Wang
Writing lesson: Define the operational rubric, illustrate a scored example, and validate the evaluator separately from the system being evaluated.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong priority-fit candidate for video-generation and VLM-judge evaluation; scientifically useful only with explicit operational-physics and evaluator-validity caveats.
Relevant reading routes: Generative vision · V&L evaluation / safety
076 · ICLR 2026 oral · Full text + selected visuals inspected
Tongtian Zhu, Tianyu Zhang, Mingze Wang, Zhanpeng Zhou, Can Wang
Writing lesson: Turn a surprising curve into a sequence of controlled questions, define the evaluated population, and explain how every diagnostic visualization is constructed.
Scientific selection: Included for writing analysis only. Useful writing example for empirical discovery plus conditional optimization theory; priority scientific additions focus on KD, visual representations and pathology, whereas this is decentralized communication scheduling and parameter averaging, not distillation.
Relevant reading routes: Privacy / federated · Neural networks
077 · ICLR 2026 oral · Full text + selected visuals inspected
Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, Bingyi Kang
Writing lesson: Let a compact representation explain the design, then separate benchmark results, sufficiency tests and implementation trade-offs into different evidence blocks.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong priority-fit candidate joining self-supervised visual initialization, teacher pseudo-label transfer and any-view geometry; no pathology validation should be inferred.
Relevant reading routes: 3D / spatial vision · Visual representations · Knowledge distillation
078 · ICLR 2026 oral · Full text + selected visuals inspected
Thibaud Gloaguen, Mark Vero, Robin Staab, Martin Vechev
Writing lesson: State who controls each stage, measure the harmful behavior against an adapted benign baseline, and audit the evaluator before interpreting robustness.
Scientific selection: Included for writing analysis only. A security-writing example with clear threat boundaries and evaluator caveats; distillation is one tested downstream trigger, not a priority KD improvement or visual/pathology method.
Relevant reading routes: LLM evaluation / safety · LLM post-training
079 · ICLR 2026 oral · Full text + selected visuals inspected
Zilin Xiao, Qi Ma, Mengting Gu, Chun-cheng Chen, Xintao Chen, Vicente Ordonez, Vijai Mohan
Writing lesson: Present a selectable compute budget with its quality and storage costs, then distinguish controlled baselines from inherited literature numbers.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong VLM-priority candidate for compact multi-vector retrieval, nested contrastive objectives and quality–cost evaluation; not itself knowledge distillation.
Relevant reading routes: Multimodal retrieval · Document intelligence · Image–text representation
080 · ICLR 2026 oral · Full text + selected visuals inspected
Zhao Yang, Yi Duan, Jiwei Zhu, Ying Ba, Chuan Cao, Bing Su
Writing lesson: Distinguish the motivating empirical failure, the hypothesized mechanism and the actual implemented operator; validate each with the corresponding control.
Scientific selection: Included for writing analysis only. Useful hypothesis-versus-identification writing study in regulatory genomics; no visual, distillation or WSI experiment supports treating this as a priority pathology paper.
Relevant reading routes: Causal inference · Loss / objectives
081 · ICLR 2026 oral · Full text + selected visuals inspected
Feng Hong, Jiangchao Yao, Yifei Shen, Dongsheng Li, Ya Zhang, Yanfeng Wang
Writing lesson: Use an explanatory diagnostic, a mechanism-specific control and a disaggregated outcome table to make a minority-performance argument testable.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant CV objective-design candidate with an explicit minority-capacity mechanism; no WSI evaluation or KD transfer is demonstrated.
Relevant reading routes: Generative vision · Loss / objectives
082 · ICLR 2026 oral · Full text + selected visuals inspected
Jingzhe Shi, Qinwei (Martin) Ma, Hongyi Liu, Hang Zhao, Jenq-Neng Hwang, Lei Li
Writing lesson: Present an explanatory schematic, raw and normalized evidence, and a transparent assumption map; use synthetic data to test quantities otherwise hidden in real data.
Scientific selection: Included for writing analysis only. Relevant theory-writing example, but outside the current KD/CV/VLM/SSL/pathology scientific-addition priority and with assumptions needing careful mathematical review.
Relevant reading routes: LLM pretraining · Math / statistics
083 · ICLR 2026 oral · Full text + selected visuals inspected
Riccardo Poiani, Martino Bernasconi, Andrea Celli
Writing lesson: For theory papers, pair each result with its assumptions, interpretation and one decisive proof idea; explain precisely why a familiar argument fails in the harder setting.
Scientific selection: Included for writing analysis only. A useful theory exposition example, but not a priority KD, CV/VLM, SSL or pathology method.
Relevant reading routes: AI / reinforcement learning · Math / statistics
084 · ICLR 2026 oral · Full text + selected visuals inspected
Peter Holderrieth, Uriel Singer, Tommi Jaakkola, Ricky T. Q. Chen, Yaron Lipman, Brian Karrer
Writing lesson: Validate a reusable computational primitive separately from its downstream combination, and label compute and resolution wherever comparisons change.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong training-free generative-vision sampling candidate; not KD and not a demonstrated pathology application.
Relevant reading routes: Generative vision · Generative / representation · Generation / video / agents
085 · ICLR 2026 oral · Full text + selected visuals inspected
Joonghyuk Shin, Zhengqi Li, Richard Zhang, Jun-Yan Zhu, Jaesik Park, Eli Shechtman, Xun Huang
Writing lesson: For interactive systems, report latency and throughput separately, isolate infrastructure optimizations, and show how training reproduces the specific inference mechanism being claimed.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Direct video KD candidate using a bidirectional teacher, causal student and distribution-matching scores; useful for temporal distillation, not a cross-magnification pathology experiment.
Relevant reading routes: Knowledge distillation · Generative vision · Video / 3D / embodied
086 · ICLR 2026 oral · Full text + selected visuals inspected
Yueqi Song, Ketan Ramaneti, Zaid Sheikh, Ziru Chen, Boyu Gou, Tianbao Xie, Yiheng Xu, Danyang Zhang, Apurva Gandhi, Fan Yang, Joseph Liu, Tianyue Ou, Zhihao Yuan, Frank F Xu, Shuyan Zhou, Xingyao Wang, Xiang Yue, Tao Yu, Huan Sun, Yu Su, Graham Neubig
Writing lesson: For data-infrastructure papers, pair the interface abstraction with a worked conversion, provenance, actual sampling policy and downstream controls; state what the effort proxy does and does not measure.
Scientific selection: Included for writing analysis only. Useful systems/data-paper writing example, but not a selected KD, CV/VLM, SSL or pathology method.
Relevant reading routes: RAG / tools / agents · Evaluation / reproducibility
087 · ICLR 2026 oral · Full text + selected visuals inspected
Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev
Writing lesson: Organize security evidence around distinct desired properties, disclose the adversary and query budget, and put the benefit–distortion trade-off beside the headline success rate.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A distinct use of logit distillation to imprint a conditional statistical behavior; relevant to the KD map but not selected for immediate scientific promotion.
Relevant reading routes: Knowledge distillation · LLM evaluation / safety
088 · ICLR 2026 oral · Full text + selected visuals inspected
Zijian Zhao, Sen Li
Writing lesson: For an application paper, define the operating constraints first, show the complete decision path, disclose simulator assumptions, and report operational trade-offs alongside the headline reward.
Scientific selection: Included for writing analysis only. Useful application and systems-writing example outside the selected KD/CV/VLM/SSL/pathology additions.
Relevant reading routes: AI / reinforcement learning
089 · ICLR 2026 oral · Full text + selected visuals inspected
Xiaofeng Zhang, Yuanchao Zhu, Chaochen Gu, Xiaosong Yuan, Qiyan Zhao, Jiawei Cao, Barrett Tang, Sinan Fan, Yaomin Shen, Chen Shen, Hao Tang
Writing lesson: A diagnostic paper benefits from observation–intervention–failure-case structure, but causal language, implementation semantics and every reported improvement must survive a consistency audit.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant VLM reliability candidate, but unresolved source inconsistencies require code-level clarification before a definitive scientific review.
Relevant reading routes: V&L evaluation / safety · Multimodal LLMs · Interpretability
090 · ICLR 2026 oral · Full text + selected visuals inspected
Eran Malach, Omid Saremi, Sinead Williamson, Arwen Bradley, Aryo Lotfi, Emmanuel Abbe, Joshua Susskind, Etai Littwin
Writing lesson: When a system overcomes a model limitation, state exactly which resource boundary changed and keep the theorem, training construction and empirical extrapolation ranges separate.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A useful teacher-trajectory distillation and architecture-generalization candidate, but not selected ahead of direct KD/CV/SSL additions.
Relevant reading routes: RAG / tools / agents · Text / sequence models · Knowledge distillation
091 · ICLR 2026 oral · Full text + selected visuals inspected
Deyu Zou, Yongqiang Chen, Jianxiang Wang, Garry YANG, Mufei Li, Qing Da, James Cheng, Pan Li, Yu Gong
Writing lesson: Make the bridge from an idealized latent mechanism to the implementable observable rule explicit, and test mistaken intervention as well as successful intervention.
Scientific selection: Included for writing analysis only. Active reasoning/RL writing study; using a distilled baseline is not itself a new KD method, and this is outside the current scientific-priority selection.
Relevant reading routes: Reasoning / verification · LLM post-training · AI / reinforcement learning
092 · ICLR 2026 oral · Full text + selected visuals inspected
Kieran Didi, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner, Christian Dallago, Jian Tang, Michael Bronstein, Martin Steinegger, Emine Kucukbenli, Arash Vahdat, Karsten Kreis
Writing lesson: Separate the quality of a learned prior from the benefit of search, plot improvement against a clearly defined compute budget, and expose the evaluation pipeline as carefully as the architecture.
Scientific selection: Included for writing analysis only. Useful generative-model and compute-scaling writing study, but protein binder design is outside the selected KD/CV/VLM/SSL/pathology scientific scope.
Relevant reading routes: Generative / representation · Graph / geometric ML · Evaluation / reproducibility
093 · ICLR 2026 oral · Full text + selected visuals inspected
Rafał Karczewski, Markus Heinonen, Alison Pouplin, Søren Hauberg, Vikas Garg
Writing lesson: Build a theory-driven method as obstruction → revised mathematical object → tractable estimator → task-specific evidence, and place cost accounting beside quality.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A useful diffusion-geometry and image-interpolation candidate; not a distillation method or pathology experiment.
Relevant reading routes: Generative / representation · Generative vision
094 · ICLR 2026 oral · Full text + selected visuals inspected
Wenbin Ouyang, Sirui Li, Yining Ma, Cathy Wu
Writing lesson: Separate the correctness of a transformation from the quality of its learned controller, and pair endpoint scores with time-to-quality curves when claiming acceleration.
Scientific selection: Included for writing analysis only. A useful graph-learning and optimization writing example, but vehicle-routing segmentation is not visual segmentation, KD, SSL or pathology.
Relevant reading routes: Graph / geometric ML · ML / deep learning
095 · ICLR 2026 oral · Full text + selected visuals inspected
Houyi Li, Ka Man Lo, Shijie Xuyang, Ziqi Wang, Wenzhen Zheng, Haocheng Zhang, Zhao Li, Shuigeng Zhou, Xiangyu Zhang, Daxin Jiang
Writing lesson: For efficiency papers, organize the narrative around controlled resources and expose the accounting in per-run tables rather than relying on a single performance-versus-size chart.
Scientific selection: Included for writing analysis only. An LLM resource-accounting study; relevant background for fair experiments but not a priority KD, CV/VLM, SSL or pathology addition.
Relevant reading routes: LLM pretraining · LLM systems · Evaluation / reproducibility
096 · ICLR 2026 oral · Full text + selected visuals inspected
Siyuan Wang, Gaokai Zhang, Li Lyna Zhang, Ning Shang, Fan Yang, Dongyao Chen, Mao Yang
Writing lesson: For a data-centric method, show a concrete transformation of one example, then separately test the data recipe, training objective and transfer conditions.
Scientific selection: Included for writing analysis only. A long-context RL/data-construction study; distilled models are baselines, not a new KD method, and the experiments are text-only.
Relevant reading routes: Reasoning / verification · LLM post-training · NLP tasks / evaluation
097 · ICLR 2026 oral · Full text + selected visuals inspected
Junsong Chen, Yuyang Zhao, Jincheng YU, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Zhou Daquan, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie
Writing lesson: For a systems/model paper, give readers a hierarchy of pipeline → module → state update, and attach every efficiency claim to an explicit evaluation regime.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong priority CV/video candidate with linear-attention architecture, recurrent context and actual few-step DMD training in its LongSANA stage.
Relevant reading routes: Generative vision · Video / 3D / embodied · Generative / representation · Knowledge distillation
098 · ICLR 2026 oral · Full text + selected visuals inspected
Haobin Li, Yijie Lin, Peng Hu, Mouxing Yang, Xi Peng
Writing lesson: When proposing robustness at multiple stages, name the corruption types explicitly, label where each component acts, and pair ablations with failure cases and deployment costs.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-adjacent multimodal robustness and retrieval candidate using visual/textual attributes and VLM reranking; no pathology validation or KD claim.
Relevant reading routes: Multimodal retrieval · Image–text representation · Graph / geometric ML · Loss / objectives
099 · ICLR 2026 oral · Full text + selected visuals inspected
Thomas Rupf, Marco Bagatella, Marin Vlastelica, Andreas Krause
Writing lesson: Organize evaluation around the information assumptions your method changes, and separate theorem conditions, practical approximations and measured behavior.
Scientific selection: Included for writing analysis only. Strong task-inference and exploration exposition, but primarily state-based reinforcement learning rather than the selected KD, CV/VLM or pathology priority.
Relevant reading routes: AI / reinforcement learning · Generative / representation
100 · ICLR 2026 oral · Full text + selected visuals inspected
Mingyang Liu, Yongshan Chen, Zhiyuan Fan, Gabriele Farina, Asuman Ozdaglar, Kaiqing Zhang
Writing lesson: For a theory paper with several regimes, provide an assumption-by-result matrix and reusable algorithm diagrams before asking readers to navigate proofs.
Scientific selection: Included for writing analysis only. The main contribution is online-learning/game theory with an illustrative LLM router, outside the selected scientific-addition priorities.
Relevant reading routes: AI / reinforcement learning · Math / statistics · LLM systems
101 · ICLR 2026 oral · Full text + selected visuals inspected
Kyu Beom Han, Jaeyoon Kim, Woo Jae Kim, Jinhwan Seo, Sung-eui Yoon
Writing lesson: When proposing a physical consistency loss, visualize the missing supervision and create a direct measurement of the hidden quantity rather than relying only on final image quality.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong CV inverse-rendering and objective-design candidate; physically constrained residual matching is not teacher–student knowledge distillation.
Relevant reading routes: Restoration / inverse problems · 3D / spatial vision · Loss / objectives
102 · ICLR 2026 oral · Full text + selected visuals inspected
Marco Bondaschi, Nived Rajaraman, Xiuying Wei, Razvan Pascanu, Caglar Gulcehre, Michael Gastpar, Ashok Makkuva
Writing lesson: Build a mechanistic paper around a tractable task with a known optimal estimator, then distinguish observed learning, representational construction and capacity limits.
Scientific selection: Included for writing analysis only. Useful sequence-model mechanism and theory exposition, but not a direct KD, CV/VLM, SSL or pathology method selected for scientific promotion.
Relevant reading routes: Text / sequence models · Neural networks · Interpretability
103 · ICLR 2026 oral · Full text + selected visuals inspected
Zhenxing Xu, Yizhe Zhang, Weidong Bao, Hao Wang, Ming Chen, Haoran Ye, Wenzheng Jiang, Hui Yan, Ji Wang
Writing lesson: For an agentic controller, show the measurable state, role-specific prompts and feedback loop, then report task quality beside end-to-end overhead.
Scientific selection: Included for writing analysis only. Metaheuristic hyperparameter control is outside the selected KD/CV/VLM/SSL/pathology priorities; reporting inconsistencies further argue against immediate scientific promotion.
Relevant reading routes: RAG / tools / agents · AI / reinforcement learning
104 · ICLR 2026 oral · Full text + selected visuals inspected
Pinyi Zhang, Ting-En Lin, Yuchuan Wu, Jingyang Chen, Zongqi Wang, Hua Yang, Bing Zhao, Fei Huang, Yongbin Li, Kai Zhang
Writing lesson: Separate training improvements from test-time sampling gains and show whose preferences, which judges and what compute define each evaluation.
Scientific selection: Included for writing analysis only. Primarily personalized language reward modeling, outside the selected KD, CV/VLM, SSL and pathology scientific priorities.
Relevant reading routes: LLM post-training · LLM evaluation / safety · Generation / MT evaluation
105 · ICLR 2026 oral · Full text + selected visuals inspected
Zhuoyang Zhang, Luke Huang, Chengyue Wu, Shang Yang, Kelly Peng, Yao Lu, Song Han
Writing lesson: Explain a parallel generator with a visibility-and-cache diagram, then report matched sequential and parallel variants with both quality and explicitly conditioned latency.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority CV-generation candidate: locality-aware image-token parallelization with a matched sequential control; not knowledge distillation.
Relevant reading routes: Generative vision · Generation / video / agents · Generative / representation
106 · ICLR 2026 oral · Full text + selected visuals inspected
Rahul Thomas, Arka Pal
Writing lesson: For a numerical algorithm, report objective quality, computational cost, failure rate, fallback behavior, and approximation error together rather than letting one performance number stand for all five.
Scientific selection: Included for writing analysis only. Valuable optimization-and-serving writing example, but not a priority KD, CV/VLM, SSL, or pathology method; draft verification is not teacher–student training.
Relevant reading routes: LLM systems · Math / statistics
107 · ICLR 2026 oral · Full text + selected visuals inspected
Seonggyun Lee, Sungjun Lim, Seojin Park, Soeun Cheon, Kyungwoo Song
Writing lesson: When a method exploits extra data, explain its format mismatch visually and include matched-extra-data baselines alongside label-budget comparisons.
Scientific selection: Included for writing analysis only. Semi-supervised language preference optimization rather than self-supervised representation learning or KD; useful writing study, outside the current scientific-addition priorities.
Relevant reading routes: LLM post-training · Loss / objectives
108 · ICLR 2026 oral · Full text + selected visuals inspected
Hideaki Kim, Tomoharu Iwata
Writing lesson: Organize a theory-to-algorithm paper as a dependency chain, and distinguish approximation cost, statistical error, and different scaling dimensions in both prose and experiments.
Scientific selection: Included for writing analysis only. Strong mathematical-method writing example for temporal point processes, without a direct priority KD, CV/VLM, SSL, or pathology contribution.
Relevant reading routes: Math / statistics · Classical ML · Loss / objectives
109 · ICLR 2026 oral · Full text + selected visuals inspected
Pengxiang Li, Yefan Zhou, Dilxat Muhtar, Lu Yin, Shilin Yan, Li Shen, Yi Liang, Soroush Vosoughi, Shiwei Liu
Writing lesson: Pair an internal-dynamics visualization with an executable decision rule, then state exactly which samples and output regions the motivating statistic describes.
Scientific selection: Included for writing analysis only. Language diffusion inference is relevant background, but this training-free rule is not itself distillation, CV/VLM, SSL representation learning, or pathology.
Relevant reading routes: LLM systems · Reasoning / verification · Generative / representation
110 · ICLR 2026 oral · Full text + selected visuals inspected
Jonathn Chang, Leonhard Piff, Suvadip Sana, Jasmine Li, Lionel Levine
Writing lesson: For a subjective benchmark, define the measurement population and criteria first, publish the elicitation protocol, and validate several distinct failure modes instead of treating one correlation as universal validity.
Scientific selection: Included for writing analysis only. Relevant evaluation-writing case study, but not a priority KD, CV/VLM, SSL, or pathology method.
Relevant reading routes: LLM evaluation / safety · Evaluation / reproducibility · LLM post-training
111 · ICLR 2026 oral · Full text + selected visuals inspected
Guhao Feng, Shengjie Luo, Kai Hua, Ge Zhang, Wenhao Huang, Di He, Tianle Cai
Writing lesson: Use an explicit barrier-to-design-to-experiment mapping, while keeping theoretical assumptions and practical stabilization rules visible in the same narrative.
Scientific selection: Included for writing analysis only. Useful fast-weight and objective-design background, but evaluated as language long-context adaptation rather than the priority CV/VLM, representation-SSL, KD, or pathology methods.
Relevant reading routes: LLM pretraining · LLM systems · Transfer / adaptation / continual · Loss / objectives
112 · ICLR 2026 oral · Full text + selected visuals inspected
Karen Ullrich, Jingtong Su, Claudia Shi, Arjun Subramonian, Amir Bar, Ivan Evtimov, Nikolaos Tsilivis, Randall Balestriero, Julia Kempe, Mark Ibrahim
Writing lesson: A benchmark paper can be organized around a controlled intervention axis, with an explicit observation/action/reward contract and captions that carry filtering and aggregation details.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority VLM/UI-agent robustness resource with configurable appearance/content and source-qualified state-based evaluation.
Relevant reading routes: Generation / video / agents · V&L evaluation / safety · Vision evaluation
113 · ICLR 2026 oral · Full text + selected visuals inspected
Ritik Raina, Abe Leite, Alexandros Graikos, Seoyoung Ahn, Dimitris Samaras, Gregory Zelinsky
Writing lesson: Tie each proposed conditioning stream to an experimental hypothesis, then display mechanism, behavioral protocol, and failure examples as distinct evidence layers.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A relevant DINOv2-conditioned generation and human-perception evaluation candidate, but it does not introduce a new SSL or KD training method.
Relevant reading routes: Generative vision · Visual representations · Vision evaluation
114 · ICLR 2026 oral · Full text + selected visuals inspected
Satchel Grant, Simon Jerome Han, Alexa Tartaglini, Christopher Potts
Writing lesson: Organize a methodological critique around separate questions of existence, relevance, and mitigation, and state exactly which scientific claim each counterexample challenges.
Scientific selection: Included for writing analysis only. Useful causal-interpretability writing and methodological caution, but not a priority KD, CV/VLM, SSL, or pathology method in this selective scientific pass.
Relevant reading routes: Interpretability · Causal inference · Loss / objectives · LLM interpretability
115 · ICLR 2026 oral · Full text + selected visuals inspected
Ning Yang, Hai Lin, Yibo Liu, Baoliang Tian, Guoqing Liu, Haijun Zhang
Writing lesson: Map each methodological motivation to a named component and a corresponding ablation, while keeping benchmark labels and training protocol visible in every result table.
Scientific selection: Included for writing analysis only. A text preference-optimization study rather than a selected KD/CV/VLM/SSL method; its derivation and reporting ambiguities further argue against immediate scientific promotion.
Relevant reading routes: LLM post-training · Loss / objectives
116 · ICLR 2026 oral · Full text + selected visuals inspected
Usha Bhalla, Alex Oesterling, Claudio Mayrink Verdun, Hima Lakkaraju, Flavio Calmon
Writing lesson: Make a representation-learning prior visible in an example first, formalize it as assumptions, and test the desired property alongside reconstruction and downstream utility.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A genuine temporal contrastive/self-supervised representation-learning method with residual low-level reconstruction, relevant to the SSL concept route beyond images.
Relevant reading routes: Generative / representation · Loss / objectives · Interpretability · LLM interpretability
117 · ICLR 2026 oral · Full text + selected visuals inspected
Xuan Ju, Tianyu Wang, Yuqian Zhou, HE Zhang, Qing Liu, Cherry Zhao, Zhifei Zhang, Yijun Li, Yuanhao Cai, Shaoteng Liu, Daniil Pakhomov, Zhe Lin, Soo Ye Kim, Qiang Xu
Writing lesson: When unification is the contribution, use a shared visual grammar for all input/output modes and independently ablate architecture choices and training mixtures.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong CV/VLM priority fit: unified instruction-driven image/video editing and generation with interleaved conditioning and flow matching.
Relevant reading routes: Generative vision · Video / 3D / embodied · Generation / video / agents · Generative / representation
118 · ICLR 2026 oral · Full text + selected visuals inspected
Utkarsh Mishra, David He, Yongxin Chen, Danfei Xu
Writing lesson: A transferable algorithm is easier to explain when a small running failure case, a two-component diagram, and domain-specific evaluations all use the same local/global vocabulary.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority CV generative-inference candidate with concrete panorama and long-video results, alongside a substantial robotics application.
Relevant reading routes: Generative vision · Video / 3D / embodied · Generative / representation · AI / reinforcement learning
119 · ICLR 2026 oral · Full text + selected visuals inspected
Haoyue Dai, Immanuel Albrecht, Peter Spirtes, Kun Zhang
Writing lesson: Use a recurring small graph to connect algebra, graph operations, and identifiability; report oracle guarantees and finite-sample evidence as distinct claims.
Scientific selection: Included for writing analysis only. Valuable theory exposition, but not a direct KD, CV/VLM, SSL, or computational-pathology method in the selected scientific-priority pass.
Relevant reading routes: Causal inference · Math / statistics · Graph / geometric ML
120 · ICLR 2026 oral · Full text + selected visuals inspected
Haochen Zhou, Guofeng Ding, Mouxing Yang, Peng Hu, Yijie Lin, Xi Peng
Writing lesson: Build a failure taxonomy first, give each proposed component a visible place in the mechanism diagram, and assign each experimental question an explicit evidence destination.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-relevant unsupervised cross-view representation and noisy image–text correspondence method, pending technical promotion rather than automatically added.
Relevant reading routes: Visual representations · Image–text representation · Loss / objectives · Classical ML
121 · ICLR 2026 oral · Full text + selected visuals inspected
Zhiheng Xi, Jixuan Huang, Chenyang Liao, Baodai Huang, Jiaqi Liu, Honglin Guo, yajie yang, Rui Zheng, Junjie Ye, Jiazheng Zhang, Wenxiang Chen, Wei He, Yiwen Ding, Guanyu Li, Zehui Chen, Zhengyin Du, Xuesong Yao, Yufei Xu, Jiecao Chen, Tao Gui, Zuxuan Wu, Qi Zhang, Xuanjing Huang, Yu-Gang Jiang
Writing lesson: For a systems-and-algorithm paper, explain the reusable interface first, diagnose the training failure with paired traces, then connect the proposed schedule to both effectiveness and resource cost.
Scientific selection: Included for writing analysis only. Useful agent-training and systems-writing example, but this pass prioritizes KD, CV/VLM, SSL, and pathology methods; this text-based multi-turn RL framework is not a direct fit.
Relevant reading routes: RAG / tools / agents · AI / reinforcement learning · LLM systems
122 · ICLR 2026 oral · Full text + selected visuals inspected
Xiangjun Tang, Biao Zhang, Peter Wonka
Writing lesson: When introducing a representation, show both its internal data flow and the downstream artifact it improves. Align qualitative crops with the quantitative metric, and explicitly separate raw outputs from rendering enhancements.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A relevant CV/flow-based representation candidate, but less immediate to KD and pathology than the selected first additions.
Relevant reading routes: Generative vision · Video / 3D / embodied
123 · ICLR 2026 oral · Full text + selected visuals inspected
Xiangjian Jiang, Nikola Simidjievski, Mateja Jamnik
Writing lesson: For a new evaluation metric, start with a concrete counterexample, separate validity from usefulness and cost, and make the appendix follow the same research-question structure as the main paper.
Scientific selection: Included for writing analysis only. A strong benchmark-writing example for tabular generation, but not a direct KD, CV/VLM, SSL, or pathology-method priority in this update.
Relevant reading routes: Evaluation / reproducibility · Classical ML · Generative / representation
124 · ICLR 2026 oral · Full text + selected visuals inspected
Neil Kale, Chen Bo Calvin Zhang, Kevin Zhu, Ankit Aich, Paula Rodriguez, Christina Knight, Zifan Wang
Writing lesson: Define the threat model before the architecture, then show paired metrics and counterexamples that reveal where an apparently stronger monitor still fails.
Scientific selection: Included for writing analysis only. Strong threat-model and safety-evaluation writing study. Weak-to-strong monitoring is not teacher–student training, and no priority KD/CV/VLM/SSL/pathology method is introduced.
Relevant reading routes: LLM evaluation / safety · RAG / tools / agents
125 · ICLR 2026 oral · Full text + selected visuals inspected
Aayush Karan, Yilun Du
Writing lesson: Explain the nearest plausible alternative with a minimal worked example, retain that alternative as an experimental control, and attach inference cost to a training-free claim.
Scientific selection: Included for writing analysis only. Useful inference-time reasoning exposition, but outside the selected KD/CV/VLM/SSL/pathology scientific-addition priorities.
Relevant reading routes: Reasoning / verification · LLM systems · Math / statistics
126 · ICLR 2026 oral · Full text + selected visuals inspected
Zhaomin Wu, Haodong Zhao, Ziyang Wang, Jizhou Guo, Qian Wang, Bingsheng He
Writing lesson: Use an analogy to organize a definition, not replace it; explicitly identify each approximation between the formal object and the implemented measurement.
Scientific selection: Included for writing analysis only. Behavioral fingerprinting includes distilled models but proposes no distillation method; retained as a writing/evaluation example rather than a priority KD addition.
Relevant reading routes: LLM evaluation / safety · LLM interpretability · Evaluation / reproducibility
127 · ICLR 2026 oral · Full text + selected visuals inspected
Xiaoyuan Cheng, Wenxuan Yuan, Yiming Yang, Yuanzhao Zhang, Sibo Cheng, Yi He, Zhuo Sun
Writing lesson: Make each theoretical motivation correspond to a labeled objective term and a distinct empirical diagnostic, then provide implementation-specific algorithms in the appendix.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant self-supervised dynamical representation and anti-collapse objective, but mathematical assumptions need a specialist check before promotion; not KD or pathology.
Relevant reading routes: Visual representations · Video / 3D / embodied · Generative / representation · Loss / objectives
128 · ICLR 2026 oral · Full text + selected visuals inspected
Marco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, Andrea Tirinzoni
Writing lesson: Build a bridge between neighboring research communities by defining their objects before combining them. Present an intuition figure, an executable algorithm, qualified theory and question-led experiments as complementary views of one contribution.
Scientific selection: Added to the scientific catalog with a full review. Added to SSL, reinforcement learning and visual-representation routes with a source-checked bilingual review. Reward-free state–action learning and reward-conditioned task inference are not evidence of pathology or cross-magnification transfer.
Relevant reading routes: Generative / representation · AI / reinforcement learning · Visual representations · Video / 3D / embodied
129 · ICLR 2026 oral · Full text + selected visuals inspected
Raghu Arghal, Kevin He, Shirin Saeedi Bidokhti, Saswati Sarkar
Writing lesson: When connecting a formal model to LLM simulations, visualize the information available to each role and report systematic departures from the theoretical agent assumptions.
Scientific selection: Included for writing analysis only. A theory-to-simulation writing case about information mediation, outside prioritized KD, CV/VLM, SSL and pathology methods.
Relevant reading routes: LLM evaluation / safety · RAG / tools / agents · Math / statistics
130 · ICLR 2026 oral · Full text + selected visuals inspected
Zeeshan Hayder, Ali Cheraghian, Lars Petersson, Mehrtash Harandi, Richard Hartley
Writing lesson: Name each optimization failure, show the corresponding mechanism, and give it a separate diagnostic or ablation. Keep an exact derivation, a practical approximation and their respective guarantees visibly distinct.
Scientific selection: Added to the scientific catalog with a full review. Direct fit for gradient-aware distillation and objective balancing; include explicit analytical and implementation caveats. This is not cross-magnification KD or residual-error distillation.
Relevant reading routes: Knowledge distillation · Loss / objectives · Image / recognition · Detection / segmentation
131 · ICLR 2026 oral · Full text + selected visuals inspected
Zhehao Huang, Yuhang Liu, Baijiong Lin, Yixin Lou, Zhengbao He, Hanling Tian, Tao Li, Xiaolin Huang
Writing lesson: When preserving one capability while adding another, report both task success and the specific failure mechanism the preservation component is intended to prevent.
Scientific selection: Included for writing analysis only. Relevant parameter-merging study but not teacher–student KD; it uses previously distilled reasoning models rather than proposing a distillation algorithm, and is outside current scientific priorities.
Relevant reading routes: Transfer / adaptation / continual · LLM post-training · Reasoning / verification
132 · ICLR 2026 oral · Full text + selected visuals inspected
Geon-Hyeong Kim, Yu Jin Kim, Byoungjip Kim, Honglak Lee, Kyunghoon Bae, Youngsoo Jang, Moontae Lee
Writing lesson: For a simple method, spend the saved architectural complexity on explicit data-transformation rules and evaluation that reveals both benefits and behavioral costs.
Scientific selection: Included for writing analysis only. Strong objective/evaluation-writing case, but not a priority KD, CV/VLM, SSL or pathology addition.
Relevant reading routes: LLM post-training · LLM evaluation / safety · Loss / objectives
133 · ICLR 2026 oral · Full text + selected visuals inspected
Zhaomin Wu, Mingzhe Du, See-Kiong Ng, Bingsheng He
Writing lesson: Pair every abstract behavioral construct with an inspectable controlled task, complete prompts and a sensitivity analysis that shows exactly what the metric measures.
Scientific selection: Included for writing analysis only. Useful benchmark-writing and operationalization case, but outside the prioritized KD, CV/VLM, SSL and pathology scientific additions.
Relevant reading routes: LLM evaluation / safety · Reasoning / verification
134 · ICLR 2026 oral · Full text + selected visuals inspected
Naimeng Ye, Arnav Ahuja, Georgios Liargkovas, Yunan Lu, Kostis Kaffes, Tianyi Peng
Writing lesson: Draw the actual time dependencies, report hit probability separately from latency reduction, and place semantic safety conditions beside any lossless-acceleration claim.
Scientific selection: Included for writing analysis only. A systems-writing and agent-latency example outside the selected CV/VLM, SSL and KD promotion priorities.
Relevant reading routes: RAG / tools / agents · LLM systems
135 · ICLR 2026 oral · Full text + selected visuals inspected
Guanhua Zhang, Ricardo Dominguez-Olmedo, Moritz Hardt
Writing lesson: Define what a comparison estimates, map each claim to a repeated visual grammar, and report stratified checks that reveal where a pooled trend weakens.
Scientific selection: Included for writing analysis only. Useful evaluation methodology and qualification practice, but outside the present prioritized CV/VLM, SSL and KD scientific additions.
Relevant reading routes: LLM evaluation / safety · Evaluation / reproducibility · Transfer / adaptation / continual
136 · ICLR 2026 oral · Full text + selected visuals inspected
Guanjun Wang, Binwu Wang, Jiaming Ma, Zhengyang Zhou, Pengkun Wang, Xu Wang, Yang Wang
Writing lesson: Organize a unification paper around independent failure axes, then test each axis and their intersection with clearly paired controls.
Scientific selection: Included for writing analysis only. Relevant general graph-imbalance writing example, but not a direct KD, visual SSL or pathology evaluation; deployment transfers remain hypotheses.
Relevant reading routes: Graph / geometric ML · Loss / objectives
137 · ICLR 2026 oral · Full text + selected visuals inspected
Changli Tang, Qinfan Xiao, Ke Mei, Tianyi Wang, Fengyun Rao, Chao Zhang
Writing lesson: For a unified model, explicitly enumerate supported input combinations, training pair types and evaluation query/target directions. Pair each capability claim with a controlled ablation and a small, interpretable example.
Scientific selection: Added to the scientific catalog with a full review. Added to multimodal representation, retrieval and video routes with a source-checked bilingual review. Embedding-based candidate selection is not free-form generation; audio-video results are not WSI evidence.
Relevant reading routes: Image–text representation · Multimodal retrieval · Multimodal LLMs · Video / 3D / embodied
138 · ICLR 2026 oral · Full text + selected visuals inspected
Ben Finkelshtein, Silviu Cucerzan, Sujay Kumar Jauhar, Ryen White
Writing lesson: Use matched interface diagrams and controlled perturbation grids to turn a broad benchmark into actionable conditional guidance.
Scientific selection: Included for writing analysis only. Useful interface-comparison and benchmark-writing study, outside this release's prioritized KD, visual SSL and pathology additions.
Relevant reading routes: Graph / geometric ML · RAG / tools / agents · LLM evaluation / safety
139 · ICLR 2026 oral · Full text + selected visuals inspected
Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Y Zou, Pan Lu
Writing lesson: Make training ownership, runtime collaboration and evaluation access separately visible, and pair behavioral claims with trace-level evidence that survives checking.
Scientific selection: Included for writing analysis only. Strong modular-agent writing example; not a priority visual/KD/pathology promotion. The offline SFT comparator alone is insufficient to frame it as a general distillation method.
Relevant reading routes: RAG / tools / agents · LLM post-training · Reasoning / verification
140 · ICLR 2026 oral · Full text + selected visuals inspected
Artyom Sorokin, Nazar Buzun, Aleksandr Anokhin, Egor VEDERNIKOV, Petr Anokhin, Mikhail Burtsev, Evgeny Burnaev
Writing lesson: Use explicit provenance symbols and benchmark-specific model/metric tables to make a resource-efficiency claim auditable.
Scientific selection: Included for writing analysis only. Useful retrieval-method and protocol-writing study, outside the current priority set of visual/KD/SSL/pathology catalog additions.
Relevant reading routes: NLP tasks / evaluation · RAG / tools / agents · AI / reinforcement learning
141 · ICLR 2026 oral · Full text + selected visuals inspected
Akshat Ramachandran, Marina Neseem, Charbel Sakr, Rangharajan Venkatesan, Brucek Khailany, Tushar Krishna
Writing lesson: Pair each algorithmic compression gain with an implementation mechanism and an iso-resource control, then distinguish memory capacity from user latency.
Scientific selection: Included for writing analysis only. Relevant systems-writing and inference-compression study, but not a direct KD, visual SSL or pathology promotion.
Relevant reading routes: LLM systems · Reasoning / verification
142 · ICLR 2026 oral · Full text + selected visuals inspected
Zhe Zhang, Yuanning Feng, Yuxuan Song, Keyue Qiu, Hao Zhou, Wei-Ying Ma
Writing lesson: Connect each compressed component to a measured bottleneck, show stage-specific losses beside the architecture, and retain best/worst-case evidence instead of one headline score.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Dual consistency and variable-length diffusion acceleration are relevant distillation ideas, but the evidence is biomolecular rather than WSI and mathematical presentation needs resolution before promotion.
Relevant reading routes: Knowledge distillation · Generative / representation
143 · ICLR 2026 oral · Full text + selected visuals inspected
Yuhao Wu, Yushi Bai, Zhiqiang Hu, Roy Ka-Wei Lee, Juanzi Li
Writing lesson: Use research-question headings and matched learning curves to connect design choices with evidence, and publish evaluation prompts plus specific observed failure modes.
Scientific selection: Included for writing analysis only. Useful long-output RL and evaluation writing study, but outside the current prioritized KD/CV/VLM/SSL/pathology scientific additions.
Relevant reading routes: Generation / MT evaluation · LLM post-training · Reasoning / verification
144 · ICLR 2026 oral · Full text + selected visuals inspected
Xuying Ning, Dongqi Fu, Tianxin Wei, Mengting Ai, Jiaru Zou, Ting-Wei Li, Hanghang Tong, Yada Zhu, Hendrik Hamann, Jingrui He
Writing lesson: For a benchmark paper, let the annotation units determine the metrics and organize the experiments around diagnostic questions; expose prompts and concrete success/failure traces alongside aggregate scores.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-relevant multimodal agent evaluation and process-supervised adaptation, but hold scientific promotion until split documentation and inconsistent table entries are checked.
Relevant reading routes: Generation / video / agents · V&L evaluation / safety · RAG / tools / agents
145 · ICLR 2026 oral · Full text + selected visuals inspected
Aakash Sunil Lahoti, Kevin Li, Berlin Chen, Caitlin Wang, Aviv Bick, Zico Kolter, Tri Dao, Albert Gu
Writing lesson: Organize a multi-component architecture around separate failure modes and preserve that mapping through derivation, ablation and measured outcomes. Compare old/new block diagrams in the same visual coordinate system, and report exactly how model size and runtime protocols are matched.
Scientific selection: Added to the scientific catalog with a full review. Added to sequence-modeling and efficient-serving routes with a source-checked bilingual review. Language and synthetic-sequence results motivate transfer questions but do not establish natural-image or pathology performance.
Relevant reading routes: Text / sequence models · Transformers / pretraining · LLM systems
146 · ICLR 2026 oral · Full text + selected visuals inspected
Changxin Tian, Jiapeng Wang, Qian Zhao, Kunlong Chen, Jia Liu, Ziqi Liu, Jiaxin Mao, Xin Zhao, Zhiqiang Zhang, JUN ZHOU
Writing lesson: Make an implementation change tangible with a before/after workflow, derive its central mapping in a few traceable steps, and separate the main comparison from sensitivity and compatibility controls.
Scientific selection: Included for writing analysis only. Relevant optimization/model-averaging background, but the studied setting is language-only pretraining rather than this update's priority KD, visual, SSL or pathology methods.
Relevant reading routes: LLM systems · Loss / objectives
147 · ICLR 2026 oral · Full text + selected visuals inspected
Ziyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, Zhuotao Tian
Writing lesson: Turn an efficiency paper into three distinct tests: quality at a fixed compression level, quality when spending saved capacity, and actual measured runtime including the compression overhead. State exactly which resource is held fixed.
Scientific selection: Added to the scientific catalog with a full review. Added to efficient serving, video and multimodal-LLM routes with a source-checked bilingual review. Transfer to large WSI token sets remains untested; residual tokens do not mean residual-error KD.
Relevant reading routes: LLM systems · Multimodal LLMs · Video / 3D / embodied
148 · ICLR 2026 oral · Full text + selected visuals inspected
Xinpeng Wang, Nitish Joshi, Barbara Plank, Rico Angell, He He
Writing lesson: Define labels and counterfactual controls before presenting a detector; explain its intervention with a timeline, show competing diagnostic signals, and keep discovery demonstrations separate from validated detection claims.
Scientific selection: Included for writing analysis only. A useful oversight and evaluation-writing case, but its experiments concern text math/code reward hacking rather than priority KD, visual, SSL or pathology methods.
Relevant reading routes: LLM evaluation / safety · Reasoning / verification
149 · ICLR 2026 oral · Full text + selected visuals inspected
Li Sun, Zhenhao Huang, Silei Chen, Lanxu Yang, Junda Ye, Sen Su, Philip Yu
Writing lesson: Reuse the staged explanation from geometric intuition to an explicit loss and then to an adaptation protocol; pair the overview figure with a self-contained legend and claim-linked evaluation.
Scientific selection: Included for writing analysis only. Graph self-supervised transfer is relevant background, but the paper is not a priority visual/pathology/KD method and its internal proof/table inconsistencies should be resolved before scientific promotion.
Relevant reading routes: Transfer / adaptation / continual · Math / statistics
150 · ICLR 2026 oral · Full text + selected visuals inspected
Danyal Rehman, Tara Akhound-Sadegh, Artem Gazizov, Yoshua Bengio, Alexander Tong
Writing lesson: Build a method paper around an explicit validity contract: which quantity must be accurate, why it matters downstream, and what the new computation changes. Use separate visuals for the mechanism, the cost–quality frontier and failure-sensitive diagnostics; connect idealized guarantees to their assumptions and distinguish them from finite-model measurements.
Scientific selection: Included for writing analysis only. Included for flow-method, likelihood-validity and cost–quality writing lessons. Molecular Boltzmann sampling is outside the selected KD/CV/VLM/SSL/pathology scientific scope.
151 · ICLR 2026 oral · Full text + selected visuals inspected
Panagiotis D. Grontas, Antonio Terpin, Efe Balta, Raffaello D'Andrea, John Lygeros
Writing lesson: For optimization layers, visually separate feasibility, objective quality and runtime, and expose numerical stabilization choices as part of the method.
Scientific selection: Included for writing analysis only. Valuable constrained-layer and numerical-method writing example; the self-supervised optimization loss is not visual SSL, and current evidence is not a direct KD/pathology method.
Relevant reading routes: Neural networks · Math / statistics
152 · ICLR 2026 oral · Full text + selected visuals inspected
Mingyu Kim, Young-Heon Kim, Mijung Park
Writing lesson: For inference guidance, make timing a controlled experimental variable and report safety together with fidelity and diversity, not one composite victory claim.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant training-free visual-generation guidance, but substantive theoretical and reporting caveats make this lower priority than direct KD/SSL additions.
Relevant reading routes: Generative vision · Generation / video / agents · V&L evaluation / safety
153 · ICLR 2026 oral · Full text + selected visuals inspected
Shijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li zhang, Tianfan Xue, Jian Zhang
Writing lesson: A convincing efficiency paper can turn a mechanistic hypothesis into a design whose training and inference paths are visibly different, then evaluate quality and deployment cost separately.
Scientific selection: Added to the scientific catalog with a full review. Strong priority candidate: teacher-generated reasoning descriptions supervise a compact visual encoder without retaining the LLM, directly relevant to cross-architecture knowledge transfer and efficient visual evaluation. Selected and independently checked for a bilingual scientific review.
Relevant reading routes: Knowledge distillation · Image–text representation · Vision evaluation
154 · ICLR 2026 oral · Full text + selected visuals inspected
Gongjie Zhang, Wenhao Li, Quanhao Qian, Jiuniu Wang, Deli Zhao, Shijian Lu, Ran Xu
Writing lesson: Start from a measurable physical ambiguity, show a controlled failure, and make each proposed component correspond to that failure in the architecture and ablation.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant to geometry-aware VLM representations and spatial robustness; a potential CV/VLM addition, not direct pathology evidence or a clean teacher-removal KD example.
Relevant reading routes: 3D / spatial vision · Grounding / open vocabulary · Multimodal LLMs
155 · ICLR 2026 oral · Full text + selected visuals inspected
Chao Wang, Licheng Jiao, Lingling Li, Jiaxuan Zhao, Guanchun Wang, Fang Liu, Shuyuan Yang
Writing lesson: Give each empirical question a named subsection, and show both the outer learning loop and its internal transformations in one hierarchical figure.
Scientific selection: Included for writing analysis only. Useful optimization-paper organization, but outside the current KD/CV/VLM/SSL/pathology promotion priorities; the convergence argument requires additional scrutiny.
Relevant reading routes: Neural networks · Evaluation / reproducibility
156 · ICLR 2026 oral · Full text + selected visuals inspected
Shawn Im, Changdae Oh, Zhen Fang, Sharon Li
Writing lesson: After a formal result, explain its basis objects, compose them into individual components, and finally show how the components cooperate in the full computation.
Scientific selection: Included for writing analysis only. An instructive theory-to-interpretation presentation, but outside the current priority for KD, CV/VLM, SSL and pathology scientific additions.
Relevant reading routes: LLM interpretability · Transformers / pretraining · Interpretability
157 · ICLR 2026 oral · Full text + selected visuals inspected
Wei Ju, Siyu Yi, Kangjie Zheng, Yifan Wang, Ziyue Qiao, Li Shen, Yongdao Zhou, Xiaochun Cao, Jiancheng Lv
Writing lesson: Organize supplementary evidence around distinct failure modes, and pair an efficiency claim with an accuracy–cost plot rather than reporting speed in isolation.
Scientific selection: Included for writing analysis only. Adjacent self-supervised graph methodology with reusable experimental organization, but not a direct visual/pathology/KD priority; the strongest optimality statements need independent correction before scientific promotion.
Relevant reading routes: Graph / geometric ML · Loss / objectives
158 · ICLR 2026 oral · Full text + selected visuals inspected
Zhengbo Wang, Jian Liang, Ran He, Zilei Wang, Tieniu Tan
Writing lesson: Explain a compression method by first rewriting the familiar operation, then show the precise algorithmic differences and the accuracy–resource tradeoff.
Scientific selection: Included for writing analysis only. Useful low-rank optimization background, but it is not a KD or cross-magnification method and is outside the present scientific-addition priorities.
Relevant reading routes: LLM systems · Neural networks
159 · ICLR 2026 oral · Full text + selected visuals inspected
Yahan Li, Jifan Yao, John Bunyi, Adam Frank, Angel Hwang, Ruishan Liu
Writing lesson: Build a benchmark narrative from an explicit rubric to annotated failures and then targeted stress tests, publishing the actual annotation interface and selection rules.
Scientific selection: Included for writing analysis only. Valuable expert-evaluation design, but mental-health QA is not computational pathology and falls outside the current visual/KD/SSL scientific-promotion priorities.
Relevant reading routes: LLM evaluation / safety · Generation / MT evaluation
160 · ICLR 2026 oral · Full text + selected visuals inspected
Yi-Ge Zhang, Jingyi Cui, Qiran Li, Yisen Wang
Writing lesson: Build a theory-driven paper around one observable phenomenon, then use parallel interventions and matched controls to connect the abstraction to implementation.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Directly relevant to SSL objective design, but its assumptions and removal terminology need explicit qualification before promotion.
Relevant reading routes: Visual representations · Loss / objectives
161 · ICLR 2026 oral · Full text + selected visuals inspected
Xiaoxiao Hu, Jiaqi Jin, Sheng Li, Wanli Peng, Xinpeng Zhang, Zhenxing Qian
Writing lesson: Give each advertised property its own evaluation axis and disclose what is excluded from speed measurements directly beside the figure.
Scientific selection: Included for writing analysis only. Useful provenance-writing case, but outside the immediate KD/SSL priorities and its strongest theoretical security wording requires qualification.
Relevant reading routes: Generative vision · Vision evaluation
162 · ICLR 2026 oral · Full text + selected visuals inspected
Rim Assouel, Declan Campbell, Yoshua Bengio, Taylor Webb
Writing lesson: For a mechanistic claim, align one circuit diagram with several distinct evidence types and organize follow-up experiments as explicit questions about the proposed mechanism.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong priority fit for VLM binding and mechanistic interpretation; useful conceptual complement to visual grounding and representation routes.
Relevant reading routes: Multimodal LLMs · V&L evaluation / safety · Interpretability
163 · ICLR 2026 oral · Full text + selected visuals inspected
ZHUN WANG, Tianneng Shi, Jingxuan He, Matthew Cai, Jialin Zhang, Dawn Song
Writing lesson: A benchmark paper should make its scoring oracle inspectable, name the evaluation subset beside each comparison, and preserve negative agent traces with versioned configuration details.
Scientific selection: Included for writing analysis only. Valuable benchmark and reproducibility example but not an immediate KD, CV/VLM, SSL, or pathology-method addition.
Relevant reading routes: Code / software engineering · LLM evaluation / safety · RAG / tools / agents
164 · ICLR 2026 oral · Full text + selected visuals inspected
Xinyang Li, Tengfei Wang, Zixiao Gu, Shengchuan Zhang, Chunchao Guo, Liujuan Cao
Writing lesson: Frame a hybrid method through complementary failure modes, then pair quantitative ablations with visual evidence when the chosen metrics miss the artifact being addressed.
Scientific selection: Added to the scientific catalog with a full review. High-priority cross-mode generative distillation candidate that directly connects KD with 3D consistency and few-step generation. Selected and independently checked for a bilingual scientific review.
Relevant reading routes: Knowledge distillation · Generative vision · 3D / spatial vision
165 · ICLR 2026 oral · Full text + selected visuals inspected
Weiduo Liao, Fei Han, Hisao Ishibuchi, Qingfu Zhang, Ying Wei
Writing lesson: For an add-on method, present matched with/without pairs across host algorithms, isolate capacity and representation-use controls, and disclose both reusable preparation cost and per-host overhead.
Scientific selection: Candidate for a later scientific review; not added by this writing study. High-priority relational/concept distillation for continual vision learning; directly useful beyond classic larger-teacher/smaller-student KD.
Relevant reading routes: Knowledge distillation · Transfer / adaptation / continual · Visual representations
166 · ICLR 2026 oral · Full text + selected visuals inspected
Qingyuan Tian, Wenhong Zhu, Xiaoran Liu, Xiaofeng Wang, Rui Wang
Writing lesson: Use a shared mathematical parameterization and a matching visual vocabulary to explain a small method change; evaluate distinct consequences with distinct diagnostics instead of treating one metric as a universal proxy.
Scientific selection: Included for writing analysis only. Useful position-encoding and context-extension writing example, but not directly a KD, SSL, CV/VLM or pathology priority in this paper.
Relevant reading routes: LLM systems · Transformers / pretraining
167 · ICLR 2026 oral · Full text + selected visuals inspected
Chenqi Li, Yu Liu, Timothy Denison, Tingting Zhu
Writing lesson: Use distinct train/inference panels and a direction-aware result table whenever transfer changes both the model path and the information available to it.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A useful cross-modal representation-transfer alternative to KD, with explicit teacher-tail reuse and resource tradeoffs; biosignal rather than pathology evidence.
Relevant reading routes: Knowledge distillation · Transfer / adaptation / continual
168 · ICLR 2026 oral · Full text + selected visuals inspected
Le-Trung Nguyen, Enzo Tartaglione, Van-Tam Nguyen
Writing lesson: Build an efficiency argument in layers: identify the stored tensors, test the structural assumption, show accuracy-resource frontiers, then measure real hardware and disclose where compression fails.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant adjacent vision-transformer compression method, though not teacher-student KD; useful for separating trainable-parameter, activation and inference costs.
Relevant reading routes: Visual representations · Transfer / adaptation / continual · LLM systems
169 · ICLR 2026 oral · Full text + selected visuals inspected
Devvrit Khatri, Lovish Madaan, Rishabh Tiwari, Rachit Bansal, Venkata Sai Surya Subramanyam Duvvuri, Manzil Zaheer, Inderjit Dhillon, David Brandfonbrener, Rishabh Agarwal
Writing lesson: Show observed versus predicted points unmistakably, explain fit parameters visually, and validate the final combined recipe rather than assuming independent component improvements add together.
Scientific selection: Included for writing analysis only. Strong experimental-design and scaling-writing example, outside the current KD/CV/VLM/SSL scientific-promotion priorities.
Relevant reading routes: LLM post-training · AI / reinforcement learning · Evaluation / reproducibility
170 · ICLR 2026 oral · Full text + selected visuals inspected
Noah Amsel, David Persson, Christopher Musco, Robert M. Gower
Writing lesson: For a numerical subroutine, connect the exact theorem's admissible algorithm class to a concrete implementation, then test whether better numerical accuracy actually improves the downstream task.
Scientific selection: Included for writing analysis only. Strong theory-to-implementation writing reference; optimizer infrastructure rather than a priority KD, SSL or pathology method, and its CV extension does not establish consistent superiority.
Relevant reading routes: ML / deep learning · LLM systems · Math / statistics
171 · ICLR 2026 oral · Full text + selected visuals inspected
zhipeng cai, Ching-Feng Yeh, Hu Xu, Zhuang Liu, Gregory P. Meyer, Xinjie Lei, Changsheng Zhao, Shang-Wen Li, Vikas Chandra, Yangyang Shi
Writing lesson: Organize an ablation-driven method as explicit questions and numbered findings, then show exactly how the final pipeline follows from those findings.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Direct priority VLM/CV contribution on geometric grounding without architecture changes; no pathology evidence.
Relevant reading routes: Multimodal LLMs · 3D / spatial vision · Grounding / open vocabulary
172 · ICLR 2026 oral · Full text + selected visuals inspected
Kartik Nair, Indradyumna Roy, Soumen Chakrabarti, Anirban Dasgupta, Abir De
Writing lesson: State the exact random object and assumptions, then give separate evidence for each logical link; place deployment algorithms side by side when they share persisted state.
Scientific selection: Included for writing analysis only. A substantial graph-retrieval theory and presentation example, but it does not directly establish KD, CV/VLM, SSL, or pathology performance.
Relevant reading routes: Graph / geometric ML
173 · ICLR 2026 oral · Full text + selected visuals inspected
Bartłomiej Marek, Lorenzo Rossi, Vincent Hanke, Xun Wang, Michael Backes, Franziska Boenisch, Adam Dziedzic
Writing lesson: Use research questions as navigation, but preserve the exact privacy unit, attacker information, and target dataset whenever summarizing a finding.
Scientific selection: Included for writing analysis only. Useful threat-model and benchmark-writing study, but outside the current priority additions and not independently validated as a privacy deployment recommendation.
Relevant reading routes: Privacy / federated · LLM evaluation / safety
174 · ICLR 2026 oral · Full text + selected visuals inspected
Marco P Abrate, Laurenz Muessig, Joshua Bassett, Hui Tan, Francesca Cacucci, Thomas Wills, Caswell Barry
Writing lesson: Build a mechanism narrative from measured experience to controlled simulation, then test competing explanations using several complementary diagnostics.
Scientific selection: Included for writing analysis only. A strong mechanistic scientific-writing example with self-supervised prediction, but a specialized neuroscience simulation rather than a priority general-purpose CV/SSL or pathology method.
Relevant reading routes: Video / 3D / embodied · Generative / representation
175 · ICLR 2026 oral · Full text + selected visuals inspected
Romain Froger, Pierre Andrews, Matteo Bettini, Amar Budhiraja, Ricardo Cabral, Virginie Do, Emilien Garreau, Jean-Baptiste Gaya, Hugo Laurençon, Maxime Lecanu, Kunal Malkan, Dheeraj Mekala, Pierre Ménard, Gerard Bertran, Ulyana Piterbarg, Mikhail Plekhanov, Mathieu Rita, Andrey Rusakov, Vladislav Vorotilov, Mengjuew Wang, Ian Yu, Amine Benhalloum, Grégoire Mialon, Thomas Scialom
Writing lesson: For systems benchmarks, define the environment clock, observability, verifier, and resource accounting as parts of the experiment—not implementation trivia.
Scientific selection: Included for writing analysis only. Useful agent-benchmark reporting study; current scientific additions prioritize KD, CV/VLM, SSL, and pathology rather than broad agent evaluation.
Relevant reading routes: RAG / tools / agents · LLM evaluation / safety
176 · ICLR 2026 oral · Full text + selected visuals inspected
Gyuhyeon Seo, Jungwoo Yang, Junseong Pyo, Nalim Kim, Jonggeun Lee, Yohan Jo
Writing lesson: Show how data generation, task prerequisites, and scoring connect, then analyze successful trajectories as well as failures to identify hidden dependence on feedback.
Scientific selection: Included for writing analysis only. A useful benchmark design and error-analysis study, outside the current prioritized scientific additions.
Relevant reading routes: RAG / tools / agents · LLM evaluation / safety
177 · ICLR 2026 oral · Full text + selected visuals inspected
Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Michael Bronstein, Anru Zhang, Joey Bose, Alexander Tong
Writing lesson: Expose every approximation between a likelihood theorem and the implemented loss, then use matched-protocol tables and concrete failures to delimit the practical claim.
Scientific selection: Included for writing analysis only. A useful objective-design and theory-to-implementation writing example, but the present experiments are protein sequences, text and code rather than the selected CV, VLM, SSL or knowledge-distillation additions.
Relevant reading routes: Loss / objectives · Generation / MT evaluation · Generative / representation
178 · ICLR 2026 oral · Full text + selected visuals inspected
Mason Wang, Anna Huang
Writing lesson: Introduce a representation through a precise familiar analogy, then pair a training/inference diagram with task-separated metrics and explicit perceptual-study limitations.
Scientific selection: Included for writing analysis only. Valuable representation-control and visualization writing, but audio experiments are outside the selected CV/VLM, SSL and KD promotion priorities.
Relevant reading routes: Generative / representation · Interpretability
179 · ICLR 2026 oral · Full text + selected visuals inspected
Zhaoyang Liu, JingJing Xie, Zichen Ding, Zehao Li, Bowen Yang, Zhenyu Wu, Xuehui Wang, Qiushi Sun, Shi Liu, Weiyun Wang, Shenglong Ye, Qingyun Li, Zeyue Tian, Gen Luo, Xiangyu Yue, Biqing Qi, Kai Chen, Bowen Zhou, Yu Qiao, Qifeng Chen, Wenhai Wang
Writing lesson: For a data-engineering contribution, connect acquisition decisions to task-aligned ablations, and publish the inconvenient platform failures and benchmark modifications needed to reproduce the claim.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A relevant VLM-agent data and evaluation resource, but lower priority than the selected KD/SSL and natural-image mechanism papers; requires catalog deduplication and artifact checks before any promotion.
Relevant reading routes: Generation / video / agents · Grounding / open vocabulary · V&L evaluation / safety
180 · ICLR 2026 oral · Full text + selected visuals inspected
Pascal Bergsträßer, Ryan Cotterell, Anthony W. Lin
Writing lesson: For a theory paper, use one small worked construction, state quantifier direction explicitly, and organize each theorem around a visible chain of reductions and reused witnesses.
Scientific selection: Included for writing analysis only. A strong theoretical-writing example, but not a selected empirical CV/VLM, SSL or KD addition; theorem-scope and printed-formula questions require expert checking before a technical tutorial.
Relevant reading routes: Transformers / pretraining · Math / statistics · Evaluation / reproducibility
181 · ICLR 2026 oral · Full text + selected visuals inspected
Shikang Zheng, Guantao Chen, Qinming Zhou, Yuqi Lin, Lixuan He, Chang Zou, Peiliang Cai, Jiacheng Liu, Linfeng Zhang
Writing lesson: For acceleration work, distinguish calibration, arithmetic savings, measured latency and quality reference; validate every predictor equation against its implementation before turning an analogy into a guarantee.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant efficient-generation method compatible with distilled models, but predictor and speed-claim inconsistencies require clarification before scientific promotion.
Relevant reading routes: Generative vision · Generative / representation
182 · ICLR 2026 oral · Full text + selected visuals inspected
Taishi Nakamura, Satoki Ishikawa, Masaki Kawamura, Okamoto, Daisuke Nohara, Jun Suzuki, Rio Yokota
Writing lesson: Build a scaling paper around a falsifiable mismatch, hold comparison axes explicit, and use supplementary stress tests to narrow rather than inflate the headline claim.
Scientific selection: Included for writing analysis only. Useful scaling-study writing example but no new KD, SSL, vision or pathology method in the current scientific priority scope.
Relevant reading routes: LLM pretraining · Reasoning / verification
183 · ICLR 2026 oral · Full text + selected visuals inspected
Mohamed Malhou, Ludovic Perret, Kristin Lauter
Writing lesson: Introduce structured input with one worked diagram, then pair complexity with measured padding and resource costs. Carry the data-generation restrictions into every solver-performance claim.
Scientific selection: Included for writing analysis only. Strong structured-method writing example, but its demonstrated task is finite-field symbolic algebra rather than the current KD, SSL, vision or pathology priorities.
Relevant reading routes: Math / statistics · Graph / geometric ML
184 · ICLR 2026 oral · Full text + selected visuals inspected
Anej Svete, Ashish Sabharwal
Writing lesson: Give theory readers a theorem map and worked toy example, then attach assumptions and resource blow-ups to every informal equivalence claim.
Scientific selection: Included for writing analysis only. Valuable theory-writing example outside the prioritized scientific additions in KD, SSL, vision and pathology.
Relevant reading routes: Reasoning / verification · Generation / MT evaluation · Math / statistics
185 · ICLR 2026 oral · Full text + selected visuals inspected
Jialiang Zhang, Haoran Geng, Yang You, Congyue Deng, Pieter Abbeel, Jitendra Malik, Leonidas Guibas
Writing lesson: Derive a structured operator step by step, then test both its motivating synthetic function and a downstream application; show exactly which feature types every layer changes.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A substantive structured CV method with hand reconstruction evidence, but lower priority than direct KD/SSL additions.
Relevant reading routes: 3D / spatial vision · Video / 3D / embodied · Graph / geometric ML
186 · ICLR 2026 oral · Full text + selected visuals inspected
Sanghyun Jo, Ziseok Lee, Wooyeol Lee, Jonghyun Choi, Jaesik Park, Kyungsu Kim
Writing lesson: Show the intermediate representation, explain what each pipeline stage repairs, and include numerical negative-transfer cases alongside attractive qualitative successes.
Scientific selection: Added to the scientific catalog with a full review. Selected as trace-2026 for boundary pseudo-label distillation and explicit negative pathology-transfer evidence, not a pathology performance recommendation. The official repository contains README/assets but no released implementation at commit d9d94b9c02bb6ef590d7530e26fac0edc07faa24.
Relevant reading routes: Knowledge distillation · Detection / segmentation · Visual representations · Cell / tissue models
187 · ICLR 2026 oral · Full text + selected visuals inspected
Thomas Wimmer, Prune Truong, Marie-Julie Rakotosaona, Michael Oechsle, Federico Tombari, Bernt Schiele, Jan Eric Lenssen
Writing lesson: Use separate evaluation protocols for task adaptation and representation preservation, and expand one decisive architectural operation in a second figure.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong cross-resolution feature candidate, but not itself a demonstrated cross-magnification pathology or teacher-student compression result.
Relevant reading routes: Visual representations · Detection / segmentation · 3D / spatial vision
188 · ICLR 2026 oral · Full text + selected visuals inspected
Yihan Wang, Jia Deng
Writing lesson: Illustrate the expensive operation and its replacement with matched diagrams, then distinguish mechanism ablations from benchmark and transfer evidence.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong CV correspondence and pretrained-feature reuse candidate; optical-flow residual refinement is not KD.
Relevant reading routes: Video / 3D / embodied · 3D / spatial vision · Visual representations
189 · ICLR 2026 oral · Full text + selected visuals inspected
Guillaume Braun, Bruno Loureiro, Minh Ha Quang, Masaaki Imaizumi
Writing lesson: Introduce observables and empirical phases before theorem statements, then keep the proof roadmap and appendix headings aligned with that narrative.
Scientific selection: Included for writing analysis only. Useful theory-writing structure, but outside the selected KD/CV/VLM/SSL priority and requiring specialist proof review before scientific promotion.
Relevant reading routes: Math / statistics · Neural networks · Loss / objectives
190 · ICLR 2026 oral · Full text + selected visuals inspected
Chunyang Li, Yilun Zheng, Xinting Huang, Tianqing Fang, Jiahao Xu, Lihui Chen, Yangqiu Song, Winston Hu
Writing lesson: For evaluation papers, make reference-label construction auditable and pair aggregate scores with concrete failures that distinguish system errors from evaluator errors.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant VLM/GUI evaluation candidate with useful grounding and operational-error distinctions, not a KD method.
Relevant reading routes: V&L evaluation / safety · Generation / video / agents · Code / software engineering
191 · ICLR 2026 oral · Full text + selected visuals inspected
Xinchen Zhang, Xiaoying Zhang, Youbin Wu, Yanbin Cao, Renrui Zhang, Ruihang Chu, Ling Yang, Yujiu Yang, Guang Shi
Writing lesson: Link diagnosis, intervention and application through explicit research questions, and disclose failures and metric definitions beside broad generalization language.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong VLM verification/refinement candidate, with useful synthetic-label construction and explicit transfer limitations; not automatically KD.
Relevant reading routes: V&L evaluation / safety · Generation / video / agents · Multimodal LLMs · Generative vision
192 · ICLR 2026 oral · Full text + selected visuals inspected
Xiaowen Jiang, Anton Rodomanov, Sebastian Stich
Writing lesson: For optimization writing, align assumptions, complexity tables and experimental axes, and explicitly price initialization and local work rather than reporting rounds alone.
Scientific selection: Included for writing analysis only. Useful federated optimization exposition, but not a direct KD, SSL, pathology or vision-method priority addition.
Relevant reading routes: Privacy / federated · Math / statistics
193 · ICLR 2026 oral · Full text + selected visuals inspected
Runyan Tan, Shuang Wu, Phillip Howard
Writing lesson: For a compact inference method, pair an implementable recipe with controlled operating-range plots, qualitative failures and explicit definitions of every aggregate metric.
Scientific selection: Included for writing analysis only. Useful general decoding and evaluation example, outside the selected KD/CV/VLM/SSL/pathology scientific-addition priorities.
Relevant reading routes: Generation / MT evaluation · LLM systems
194 · ICLR 2026 oral · Full text + selected visuals inspected
Zhiyu Pan, Yizheng Wu, Jiashen Hua, Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, Jieping Ye
Writing lesson: Separate data construction from deployment in both diagrams and prose, and audit every qualitative comparison cell as carefully as numeric tables.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong VLM self-improvement candidate with relevant visual-grounding and training-data lessons, retained for future scientific review rather than automatic promotion.
Relevant reading routes: Multimodal LLMs · Captioning / VQA · V&L evaluation / safety
195 · ICLR 2026 oral · Full text + selected visuals inspected
Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei-Ying Ma, Jingjing Liu, Mingxuan Wang, Hao Zhou
Writing lesson: For recurrent agents, document the complete state transition and show failure-state diffs, not only successful final answers or nominal context capacity.
Scientific selection: Included for writing analysis only. Relevant general memory-agent writing study, but not a direct priority KD, vision, SSL or pathology scientific addition.
Relevant reading routes: RAG / tools / agents · LLM systems · LLM post-training
196 · ICLR 2026 oral · Full text + selected visuals inspected
Hyojun Go, Dominik Narnhofer, Goutam Bhat, Prune Truong, Federico Tombari, Konrad Schindler
Writing lesson: For foundation-model composition, show module boundaries and trainability, evaluate retained abilities separately, and disclose every test-time refinement hidden behind a feedforward headline.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong CV and model-reuse candidate; teacher-output matching during stitched-decoder finetuning provides a genuine distillation connection, but it is not cross-magnification pathology evidence.
Relevant reading routes: Generative vision · 3D / spatial vision · Visual representations · Knowledge distillation
197 · ICLR 2026 oral · Full text + selected visuals inspected
Haijier Chen, Bo Xu, Shoujian zhang, Haoze Liu, Jiaxuan Lin, Jingrong Wang
Writing lesson: Use complementary deployment and training diagrams, distinguish inference inputs from training supervision, and evaluate each branch separately from the joint task.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong dual-teacher geometry/semantics distillation candidate; unresolved protocol and table inconsistencies should accompany any later scientific review.
Relevant reading routes: Knowledge distillation · Multimodal LLMs · Grounding / open vocabulary · 3D / spatial vision
198 · ICLR 2026 oral · Full text + selected visuals inspected
Dingdong WANG, Shujie LIU, Tianhua Zhang, Youjun Chen, Jinyu Li, Helen Meng
Writing lesson: Tie each proposed component to a named obstacle, expose annotation and judge prompts, and check qualitative examples as carefully as benchmark tables.
Scientific selection: Included for writing analysis only. Useful multimodal reward-design writing example, but speech-emotion specialization is outside the selected scientific scope.
Relevant reading routes: LLM post-training · LLM evaluation / safety
199 · ICLR 2026 oral · Full text + selected visuals inspected
Matthew Finlayson, Xiang Ren, Swabha Swayamdipta
Writing lesson: Define the observation interface and threat model early, distinguish measured experiments from extrapolations, and use analogies without importing stronger guarantees.
Scientific selection: Included for writing analysis only. A useful claims-calibration and theory-writing study, outside the selected scientific priority additions.
Relevant reading routes: LLM evaluation / safety · LLM interpretability
200 · ICLR 2026 oral · Full text + selected visuals inspected
Fan Chen, Audrey Huang, Noah Golowich, Sadhika Malladi, Adam Block, Jordan Ash, Akshay Krishnamurthy, Dylan Foster
Writing lesson: Introduce a counterexample and interpretable special case before the main theorem; label illustrative experiments, assumptions and proof dependencies separately.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Teacher-logit gradient truncation and coverage-sensitive selection are relevant KD theory candidates; retained as a candidate for a later scientific review.
Relevant reading routes: Knowledge distillation · LLM pretraining · Reasoning / verification · Math / statistics
201 · ICLR 2026 oral · Full text + selected visuals inspected
Zhiliang Peng, Jianwei Yu, Wenhui Wang, Yaoyao Chang, Yutao Sun, Li Dong, Yi Zhu, Weijiang Xu, Hangbo Bao, Zehua Wang, Shaohan Huang, Yan Xia, Furu Wei
Writing lesson: Use a token-colored rollout diagram to explain feedback, then separate capability demonstrations, controlled benchmarks, human ratings and measured runtime rather than combining them into one quality claim.
Scientific selection: Included for writing analysis only. Useful multimodal-generation writing example, but the demonstrated modality is speech rather than the current KD, SSL, CV/VLM and pathology scientific priorities.
Relevant reading routes: Generation / MT evaluation · Generative / representation
202 · ICLR 2026 oral · Full text + selected visuals inspected
Christoph Minixhofer, Ondrej Klejch, Peter Bell
Writing lesson: Validate an evaluation metric against humans under several domains, report raw ratings beside correlations, and explicitly distinguish direct validation from multilingual plausibility checks.
Scientific selection: Included for writing analysis only. Valuable benchmark-validation writing pattern, but its direct scientific evidence concerns speech rather than prioritized KD, SSL, CV/VLM or pathology methods.
Relevant reading routes: Generation / MT evaluation · Evaluation / reproducibility
203 · ICLR 2026 oral · Full text + selected visuals inspected
Rajiv Movva, Smitha Milli, Sewon Min, Emma Pierson
Writing lesson: Define an audit's observable target first, triangulate learned concepts with independent evidence, and attach every intervention claim to its actual model, metric and unit of uncertainty.
Scientific selection: Included for writing analysis only. Relevant interpretability and alignment writing example, but no direct KD, CV/VLM, SSL representation or pathology method is demonstrated in the current priority scope.
Relevant reading routes: LLM post-training · LLM interpretability · LLM evaluation / safety
204 · ICLR 2026 oral · Full text + selected visuals inspected
Jonathan Bragg, Mike D'Arcy, Nishant Balepur, Dan Bareket, Bhavana Dalvi Mishra, Sergey Feldman, Dany Haddad, Jena Hwang, Peter Jansen, Varsha Kishore, Bodhisattwa Prasad Majumder, Aakanksha Naik, Sigal Rahamimov, Kyle Richardson, Amanpreet Singh, Harshit Surana, Aryeh Tiktinsky, Rosni Vasu, Guy Wiener, Chloe Anastasiades, Stefanus Candra, Jason Dunkelberger, Daniel Emery, Rob Evans, Malachi Hamada, Regan Huff, Rodney Kinney, Matt Latzke, Jaron Lochner, Ruben Lozano-Aguilera, Ngoc-Uyen Nguyen, Smita Rao, Amber Tanaka, Brooke Vlahos, Peter Clark, Doug Downey, Yoav Goldberg, Ashish Sabharwal, Daniel Weld
Writing lesson: Make an evaluation paper auditable in layers: motivation-to-design mapping, cost–quality comparisons, exact aggregation rules, evaluator stress tests, then reproducible task and prompt specifications.
Scientific selection: Included for writing analysis only. Strong benchmark-writing and research-agent evaluation example, but outside this update's prioritized KD, CV/VLM, SSL and pathology method additions.
Relevant reading routes: RAG / tools / agents · LLM evaluation / safety · Evaluation / reproducibility
205 · ICLR 2026 oral · Full text + selected visuals inspected
Nam Nguyen, Thinh Nguyen, Bella Bose
Writing lesson: For theory-to-system papers, preserve the exact conditioning variables, feasible regimes and operational rate units all the way from definitions to figures; empirical agreement is not a substitute for a valid derivation.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Restoration/compression is a relevant CV direction, but substantive theoretical and protocol ambiguities prevent scientific promotion without clarification.
Relevant reading routes: Restoration / inverse problems · Generative / representation · Math / statistics
206 · ICLR 2026 oral · Full text + selected visuals inspected
Songyao Jin, Biwei Huang
Writing lesson: Teach a latent-structure theorem through a ladder of graphical examples, then test its assumptions directly and publish the graph-matching rules that determine reported recovery scores.
Scientific selection: Included for writing analysis only. Useful theorem-to-algorithm and assumption-stress-test writing example; its demonstrated applications are event-process graphs, not the current KD, SSL, CV/VLM or pathology priorities.
Relevant reading routes: Causal inference · Graph / geometric ML
207 · ICLR 2026 oral · Full text + selected visuals inspected
sina alemohammad, Zhangyang Wang, Richard Baraniuk
Writing lesson: Pair a simple intervention with trade-off plots, disclose the entire selection protocol, and audit equations, captions and model identifiers across the main paper and appendix before publication.
Scientific selection: Candidate for a later scientific review; not added by this writing study. A relevant natural-image generation and model-update candidate, with useful failure and trade-off analyses; proof, caption and model-identity inconsistencies should be clarified before a full scientific review. It is negative parameter extrapolation, not teacher–student knowledge distillation.
Relevant reading routes: Generative vision · Generative / representation · Evaluation / reproducibility
208 · ICLR 2026 oral · Full text + selected visuals inspected
Chengyang Ying, Huayu Chen, Xinning Zhou, Zhongkai Hao, Hang Su, Jun Zhu
Writing lesson: Use the same stage boundaries in the overview, pseudocode and experiments; show spatial coverage beside its numerical metric, and separate interaction efficiency from wall-clock cost and downstream adaptation.
Scientific selection: Included for writing analysis only. Useful unsupervised-control and generative-model writing example, but state-based URL and policy-guidance distillation are outside this pass's priority visual/pathology/compression-KD additions; the theoretical and implementation notation also warrant a separate technical review.
Relevant reading routes: AI / reinforcement learning · Generative / representation
209 · ICLR 2026 oral · Full text + selected visuals inspected
Gregory Yauney, Shahzaib Warraich, Swabha Swayamdipta
Writing lesson: When introducing a metric, visualize the decision it supports, compare its unit of analysis with existing metrics, and test the parameters that could change its conclusion.
Scientific selection: Included for writing analysis only. Useful evaluation-method and evidence-presentation study, outside the prioritized KD, CV/VLM, SSL and pathology scientific additions.
Relevant reading routes: LLM evaluation / safety · Evaluation / reproducibility · NLP tasks / evaluation
210 · ICLR 2026 oral · Full text + selected visuals inspected
Aditya Gopalan, Sayak Ray Chowdhury, Debangshu Banerjee
Writing lesson: Use a minimal counterexample and a recurring geometric diagram to connect a theoretical failure to an algorithm, while keeping local guarantees and practical implementation claims separate.
Scientific selection: Included for writing analysis only. A useful theory-to-algorithm writing case, but not a prioritized KD, CV/VLM, SSL or pathology addition in this release; implementation ambiguities need resolution before a scientific review.
Relevant reading routes: LLM post-training · Loss / objectives
211 · ICLR 2026 oral · Full text + selected visuals inspected
Yang Hu, Jieren Tan, Changliang Zou, Yajie Bao, Haojie Ren
Writing lesson: When replacing a conservative proxy constraint, display the proxy and actual target side by side, then label which algorithm enjoys which form of validity.
Scientific selection: Included for writing analysis only. Useful uncertainty/decision-theory exposition, but no direct KD, CV/VLM, SSL or pathology-specific method for this selected scientific pass.
Relevant reading routes: Math / statistics · Evaluation / reproducibility · Loss / objectives
212 · ICLR 2026 oral · Full text + selected visuals inspected
Emre Adabag, Marcus Greiff, John Subosits, Thomas Lew
Writing lesson: Connect a speed claim to an identifiable computational bottleneck, display what is reused, and show both controlled runtime and downstream solution quality.
Scientific selection: Included for writing analysis only. Strong optimization/control writing example without a direct KD/CV/VLM/SSL/pathology method in this priority pass.
Relevant reading routes: AI / reinforcement learning · ML / deep learning · Math / statistics
213 · ICLR 2026 oral · Full text + selected visuals inspected
Thomas MItchel, Hyunwoo Ryu, Vincent Sitzmann
Writing lesson: State the behavior that would falsify the central claim, design the metric around it, and show why superficially similar reconstruction or probe scores are insufficient.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Strong priority fit combining self-supervised representation learning, 3D vision and an explicit shortcut-resistant objective; pending scientific promotion.
Relevant reading routes: Visual representations · 3D / spatial vision · Video / 3D / embodied · Loss / objectives
214 · ICLR 2026 oral · Full text + selected visuals inspected
Wendong XU, Jing Xiong, Chenyang Zhao, Qiujiang Chen, Haoran Wang, Hui Shen, Zhongwei Wan, Jianbo Dai, Taiqiang Wu, He Xiao, Chaofan Tao, Zhuoqing Mao, Ying Sheng, Zhijiang Guo, Hongxia Yang, Bei Yu, Lingpeng Kong, Quanquan Gu, Ngai Wong
Writing lesson: For interactive benchmarks, diagram every actor and acceptance gate, distinguish opponent-dependent scores from standalone success, and publish complete protocol examples.
Scientific selection: Included for writing analysis only. A useful benchmark-protocol writing case, but not a prioritized KD/CV/VLM/SSL/pathology method.
Relevant reading routes: Code / software engineering · RAG / tools / agents · LLM evaluation / safety
215 · ICLR 2026 oral · Full text + selected visuals inspected
Zachary Izzo, Eshaan Nichani, Jason Lee
Writing lesson: State the precise comparison problem before the theorem, expose architectural assumptions, and pair each complexity factor with an experiment that actually varies it.
Scientific selection: Included for writing analysis only. A strong theorem-to-experiment writing example, outside the current prioritized scientific promotion topics.
Relevant reading routes: Transformers / pretraining · Math / statistics · LLM pretraining
216 · ICLR 2026 oral · Full text + selected visuals inspected
Harry Amad, Mihaela van der Schaar
Writing lesson: Define the new task precisely, use a controlled visual ablation to expose each inductive bias, and label hypothetical applications separately from demonstrated experiments.
Scientific selection: Included for writing analysis only. Useful conditional optimal-transport and surrogate-model writing, but current scientific priorities do not justify promoting its small-scale non-vision demonstrations.
Relevant reading routes: Generative / representation · Transfer / adaptation / continual · AI / reinforcement learning
217 · ICLR 2026 oral · Full text + selected visuals inspected
chunrui han, Guopeng Li, Jingwei Wu, Quan Sun, Yan Cai, Yuang Peng, Zheng Ge, Deyu Zhou, Haomiao Tang, Hongyu Zhou, Kenkun Liu, Shu-Tao Xia, Binxing Jiao, Daxin Jiang, Xiangyu Zhang, Yibo Zhu
Writing lesson: Use an inset to show local computation inside a simple global architecture, accompany it with a staged training table, and pair appealing galleries with concrete failure examples.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Priority-relevant continuous-token visual generation and tokenizer stabilization; suitable for later scientific curation after checkpoint, source and benchmark-protocol verification.
Relevant reading routes: Generative vision · Generation / video / agents · Generative / representation
218 · ICLR 2026 oral · Full text + selected visuals inspected
Yidan Xu, Yixin Wang, XuanLong Nguyen
Writing lesson: For a general framework, repeat a consistent chain: graphical-model assumption → mathematical identity → trainable approximation → separate tests of sample quality and representation utility. Put trade-offs and practical approximations next to the corresponding claim.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant flow-matching and structured-representation candidate. Retained for a later complete scientific review; prioritize checking the objective approximation, posterior interpretation and restricted image benchmarks before promotion.
Relevant reading routes: Generative / representation · Generative vision
219 · ICLR 2026 oral · Full text + selected visuals inspected
Haotian Ye, Qiyuan He, Jiaqi Han, Puheng Li, Jiaojiao Fan, Zekun Hao, Fitsum Reda, Yogesh Balaji, Huayu Chen, Sheng Liu, Angela Yao, James Y Zou, Stefano Ermon, Haoxiang Wang, Ming-Yu Liu
Writing lesson: Present a compression method as three contracts: what information is retained, how the rate is allocated, and what compute/side-information is needed to encode and decode. Pair the rate–quality plot with actual latency and clearly untested downstream tasks.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant adaptive video representation and VLM-efficiency candidate, not KD. Kept for a later scientific review to verify coding assumptions, mask overhead, evaluation normalization and repeated-table discrepancies.
Relevant reading routes: Video / 3D / embodied · Generative vision · Multimodal LLMs
220 · ICLR 2026 oral · Full text + selected visuals inspected
Guangnian Wan, Xinyin Ma, Gongfan Fang, Xinchao Wang
Writing lesson: For a safety paper, state who controls what, who observes what, and exactly what the evaluator scores. Pair the main risk measurement with utility, a benign control, failures and mitigation trade-offs; label any change in scoring protocol.
Scientific selection: Included for writing analysis only. Included for threat-model, evaluation and ethics-writing lessons. Not selected for the current KD/CV/VLM/SSL/pathology scientific additions.
221 · ICLR 2026 oral · Full text + selected visuals inspected
Haiquan Qiu, Quanming Yao
Writing lesson: Write a failure-analysis paper as a reproducible diagnostic chain: symptom → controlled substitution → error propagation → concrete arithmetic example → targeted repair. Distinguish evidence for one mechanism from claims that every related failure has the same cause.
Scientific selection: Candidate for a later scientific review; not added by this writing study. Relevant numerical-stability and training-systems candidate. Retained for later scientific integration after implementation-level verification of rounding, branch edge cases and the exact tested kernel; not a distillation method.
Relevant reading routes: LLM systems
222 · ICLR 2026 oral · Full text + selected visuals inspected
Pierre-Carl Langlais, Pavel Chizhov, Catherine Arnett, Carlos Hinostroza, Mattia Nee, Eliot Jones, Irène Girard, David Mach, Anastasia Stasenko, Ivan Yamshchikov
Writing lesson: For a dataset paper, build an audit trail from source and license metadata through transformations to measured utility. Distinguish document, word and tokenizer-dependent token counts; show subgroup results and imperfect-cleaning limits alongside aggregate scale.
Scientific selection: Included for writing analysis only. Included for dataset/provenance and multilingual-evaluation writing. This text-corpus resource is outside the selected scientific scope; writing coverage does not imply a dataset download or scientific catalog entry.
223 · ICLR 2026 oral · Full text + selected visuals inspected
Leonardo Defilippis, Yizhou Xu, Julius Girardin, Vittorio Erba, Emanuele Troiani, Lenka Zdeborova, Bruno Loureiro, Florent Krzakala
Writing lesson: Use a single phase diagram as the paper's organizing map, connect mathematical terms to interpretable phenomena, and place conjectural assumptions beside the numerical evidence that supports—but does not prove—them.
Scientific selection: Included for writing analysis only. A useful theory-writing example on regularization and spectra, but not a direct KD, CV/VLM, SSL or pathology method selected for this scientific update.
Relevant reading routes: Math / statistics · Loss / objectives
No papers match these filters. Try a shorter phrase or reset the filters.
