Skip to content
Computational Pathology Reading Roadmap

Captioning, VQA, OCR, and visual reasoning

Captioning and visual question answering (VQA) turn visual evidence into language. They differ in what the prompt determines: captioning leaves content selection largely to the model, while VQA requests a specific fact or inference. OCR and knowledge-intensive VQA add two further boundaries—reading visible symbols and retrieving information that is not in the pixels.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates.

ECCV 2026 conference additions and metadata checked through 2026-09-12; the broader literature audit boundary above is unchanged.

Prerequisite path

Section titled “Prerequisite path”
  1. Complete sequence-to-sequence learning, attention, tokenization, and generation in the NLP route.
  2. Know classification, detection, patch tokens, and calibration from computer vision.
  3. Read image–text representation and fusion so that you can separate a visual encoder, fusion module, and language decoder.

You are ready for research-level reading when you can state whether an answer is visible, readable, retrievable, inferable, or unanswerable.

Output, supervision, and evidence taxonomy

Section titled “Output, supervision, and evidence taxonomy”
Task Output Typical supervision Evidence contract
Image captioning one or more free-form sentences human captions, dense captions, synthetic captions every stated entity/attribute/relation should be visually supported
VQA short answer, choice, or explanation image–question–answer triples answer should change when relevant visual evidence changes
Scene-text VQA text string or answer using visible text OCR tokens, boxes, questions, answers transcription and downstream reasoning must be evaluated separately
Knowledge VQA answer plus ideally sources images, questions, external corpus/knowledge graph distinguish visual premise, retrieved fact, and model prior
Dense description regions/points plus phrases or detailed narrative region captions, localized narratives, pointing language should be attributable to a region or object
Visual reasoning answer, program, rationale, or structured trace compositional questions, programs, rationales intermediate claims need visual or symbolic verification

Multiple-choice accuracy can reward language priors. Free-form generation can hide a wrong concept behind paraphrase. Always add image-null, shuffled-image, and counterfactual controls.

Staged paper route

Section titled “Staged paper route”

Encoder–decoder captioning and visual attention

Section titled “Encoder–decoder captioning and visual attention”
Order Primary paper and year Venue / status Read it for
1 Show and Tell (2015) CVPR, peer reviewed CNN-to-LSTM conditional language modelling
2 Show, Attend and Tell (2015) ICML, peer reviewed soft/hard visual attention and region-conditioned word generation
3 Bottom-Up and Top-Down Attention (2018) CVPR, peer reviewed detector regions as visual tokens for captioning and VQA
4 NoCaps (2019) ICCV, peer reviewed novel-object captioning beyond the paired-caption vocabulary
5 OSCAR (2020) ECCV, peer reviewed object tags as anchors for region–word fusion across captioning and VQA
6 VinVL (2021) CVPR, peer reviewed stronger object/attribute features and their effect on language tasks
7 CoCa (2022) TMLR, peer reviewed one model for contrastive alignment and autoregressive captioning
8 BEiT-3 (2023) CVPR, peer reviewed a shared multiway encoder transferred across captioning, retrieval, and VQA
9 PaLI (2023) ICLR, peer reviewed jointly scaling language, vision, and multilingual task mixtures

Caption metrics: consensus, propositions, and their limits

Section titled “Caption metrics: consensus, propositions, and their limits”
Metric Primary paper Core signal What it still cannot prove
CIDEr / CIDEr-D Vedantam et al. (2015) TF–IDF-weighted one- through four-gram cosine agreement with multiple independent references; CIDEr-D clips repetition and penalizes length mismatch whether a matched phrase is visually grounded, correctly attached, negated, or clinically important
SPICE Anderson et al. (2016) dependency parse → textual scene graph → object/attribute/relation tuples → proposition F1, with diagnostic sub-scores parser correctness, fluency, discourse, pixel evidence, and facts omitted from references
BLEU / ROUGE / METEOR task-dependent origin papers and implementations precision-, recall-, or alignment-oriented lexical agreement complete semantic truth or visual necessity

CIDEr asks whether wording matches reference consensus; SPICE asks whether parsed semantic propositions match. Neither metric reads the image. Report the exact metric version, tokenizer/parser, reference count, corpus, and uncertainty, then add object/relation hallucination, region grounding, counterfactual images, and blinded human review.

VQA and compositional structure

Section titled “VQA and compositional structure”
Order Primary paper and year Venue / status Read it for
1 VQA (2015) ICCV, peer reviewed the task, human answer variability, and early benchmark design
2 CLEVR (2017) CVPR, peer reviewed controlled attributes, relations, counts, and executable question programs
3 VQA v2 (2017) CVPR, peer reviewed complementary images designed to reduce question-only shortcuts
4 VizWiz (2018) CVPR, peer reviewed authentic user questions, imperfect images, and explicitly unanswerable inputs
5 GQA (2019) CVPR, peer reviewed scene-graph-generated compositional questions and functional programs
6 OFA (2022) ICML, peer reviewed expressing vision–language tasks as sequence-to-sequence generation
7 GIT (2022) TMLR, peer reviewed a simple image encoder and language decoder scaled across tasks
8 Molmo and PixMo (2025; preprint 2024) CVPR, peer reviewed fully disclosed human captions, QA, pointing, and an open training recipe

Balanced answer frequencies help, but do not prove image use. Report a text-only baseline and construct minimal image pairs in which one visual fact changes while the question stays fixed.

Reading text in images

Section titled “Reading text in images”
Order Primary paper and year Venue / status Read it for
1 TextVQA (2019) CVPR, peer reviewed questions whose answers require scene-text recognition
2 OCR-VQA (2019) ICDAR, peer reviewed book-cover text, OCR, and large-scale automatically generated questions
3 M4C (2020) CVPR, peer reviewed iterative decoding and copying OCR tokens with a pointer mechanism
4 DocVQA (2021) WACV, peer reviewed OCR, layout, and answer extraction across document images
5 ChartQA (2022) Findings of ACL, peer reviewed visual attributes plus logical and arithmetic reasoning over charts
6 Qwen2-VL (2024) arXiv preprint native-resolution visual tokens and a modern OCR/document-capable MLLM
7 Qwen2.5-VL (2025) arXiv preprint document, chart, spatial, and long-video extensions
8 ChartQAPro (2025) Findings of ACL, peer reviewed real-world dashboards, diverse chart sources, conversational and unanswerable questions
9 POLYCHARTQA (2026) ACL, peer reviewed chart QA across ten languages and the English/non-English performance gap

For OCR-facing tasks, log image resolution, crop/tiling policy, OCR engine, normalized and raw strings, script/language, and whether the answer can be copied directly.

External knowledge and evidence

Section titled “External knowledge and evidence”
Order Primary paper and year Venue / status Read it for
1 OK-VQA (2019) CVPR, peer reviewed open-ended questions requiring knowledge beyond the image
2 KRISP (2021) CVPR, peer reviewed combining neural implicit knowledge with a symbolic graph
3 A-OKVQA (2022) ECCV, peer reviewed world-knowledge questions with direct answers and human rationales
4 KAT (2022) NAACL, peer reviewed retrieving explicit knowledge and fusing it with implicit language-model knowledge
5 BLIP-2 (2023) ICML, peer reviewed accessing frozen LLM knowledge through a learned visual querying bridge

An answer can be factually true but unsupported by the retrieved source, or faithful to a retrieved source that is itself wrong. Score visual premise, retrieval relevance, source authority, answer correctness, citation entailment, and abstention separately.

Method detail: decoding and reasoning

Section titled “Method detail: decoding and reasoning”

Caption/VQA systems may use:

  • a fixed answer classifier, which cannot generate unseen answers;
  • autoregressive decoding, which supports open vocabulary but adds exposure and hallucination errors;
  • copying/pointer mechanisms for OCR or retrieved entities;
  • constrained decoding into labels, boxes, or programs;
  • rationales before or after the answer;
  • self-consistency or verifier reranking.

A generated chain of thought is an output sequence, not direct access to the model’s causal computation. Test whether perturbing the cited region, text, or intermediate program changes the answer as predicted.

Evaluation contract

Section titled “Evaluation contract”
Output Primary metric Required complement
Short VQA answer VQA soft accuracy or exact match per-answer-type results, calibration, text-only/image-null controls
Caption CIDEr/SPICE and human pairwise quality object/attribute/relation hallucination and omission
OCR string exact/normalized edit distance recognition versus question-reasoning decomposition
Knowledge answer answer accuracy retrieval recall, source quality, citation support, abstention
Rationale answer correctness claim-by-claim visual support and intervention consistency
Dense narrative human coverage and correctness pointing/region linkage, unsupported-detail rate

Controlled experiment

Section titled “Controlled experiment”

Select a VQA dataset with answer annotations and construct four inputs per question:

  1. correct image;
  2. paired counterfactual image where the requested fact differs;
  3. random image from the same answer class;
  4. blank image.

Compare a text-only model, frozen dual-encoder features, a cross-modal fusion model, and a connector-based MLLM. Hold the language backbone, prompt, decoding, split, and adaptation budget fixed where possible. Report:

  • accuracy and calibration in all four conditions;
  • the fraction of answers unchanged after counterfactual replacement;
  • visual-token attention or pointing only as a diagnostic, not proof;
  • human-reviewed unsupported claims for free-form answers;
  • latency and visual-token count by resolution.

For knowledge VQA, add retrieval-off, irrelevant-retrieval, contradictory-source, and unanswerable conditions.

Failure modes

Section titled “Failure modes”
  • Question words or answer priors dominate weak visual evidence.
  • Reference captions omit valid details and reward generic safe sentences.
  • Lexical metrics reward overlap while missing contradictions or hallucinated objects.
  • OCR normalization hides character, unit, or sign errors.
  • A model copies a visible answer without understanding the requested relation.
  • Retrieved passages leak the answer or introduce unsupported facts.
  • Multiple-choice distractors expose the benchmark-generation template.
  • A rationale mentions a plausible region but is invariant to removing it.
  • Model-as-judge scoring inherits the same visual blind spots as the candidate model.
  • Synthetic questions, captions, and judgments share one proprietary teacher.

Self-check and exit task

Section titled “Self-check and exit task”
  1. Why did VQA v2 pair complementary images?
  2. When does a fixed answer classifier outperform generation for the wrong reason?
  3. How would you separate OCR failure from reasoning failure?
  4. What makes a knowledge answer correct but ungrounded?
  5. Which caption metrics can detect an unsupported object, and what do they miss?
  6. What intervention would test whether a rationale uses its claimed evidence?

Exit task: build a 100-example adversarial audit with correct, counterfactual, blank, and unanswerable cases. Produce an error matrix for perception, OCR, retrieval, reasoning, generation, and calibration rather than one accuracy number.

Bridge to computational pathology

Section titled “Bridge to computational pathology”

A pathology report is not an ordinary caption. It is a structured, selective document about a specimen or case, and many statements depend on slide-wide sampling, clinical context, ancillary tests, or conventions not visible in one tile. Report-generation evaluation must distinguish observations, measurements, diagnoses, uncertainty, omissions, and unsupported recommendations. Patient-level splitting and report template leakage are mandatory audits. Continue to the pathology VLM report-generation route for WSI-specific evidence and clinical evaluation.

Anatomical token selection and staged CT report generation

A coarse abnormality language model guides structure-specific evidence selection for a second report model. Training-only anatomical masks and a downstream report reward align the stages. Inspect which component GRPO updates and which signals come from segmentation supervision.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Mimicking Radiologists: A Coarse-to-Fine Framework with Structural Sparse Tokens for Dual-LLM Computed Tomography Report GenerationHong Liu, Dong Wei, Yefeng Zheng, Xian Wu, and Liansheng Wang2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #9339. Publisher proceedings/version of record was not verified on 2026-09-12. The paper-linked CTRG repository README still describes the earlier Structure Observation Driven Image-Text Contrastive Learning work and LLaMA-2-7B. A matching dual-LLM/GRPO release was not verified; do not treat this link as reproduction evidence.This dual-LLM CT reporting system first proposes abnormalities from coarse structural tokens, then selects local evidence to generate the full report. AP-LTF combines text–token similarity, learned scoring and attention weights, while training-only anatomical masks discourage attention outside the intended structure. GRPO refines the proposal policy using the downstream report reward.ECCV Spotlight program · Conference manuscript · Paper-linked repository (earlier method; matching release unverified)

Reading question: Can an early abnormality proposal remove evidence needed by the final report, and how is that failure measured?

Evidence boundary: Labeler-based clinical efficacy scores are proxy metrics, not clinical validation. The linked repository documents an earlier method; a matching release is not yet verified.

Spatial-frequency and response-token distillation targets

SFKD selectively aligns wavelet-localized Fourier amplitudes across backbone families. TRAG aligns response-to-vision attention using token-specific entropy-weighted KL. Both teach structure beyond final logits, but use different correspondence assumptions.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Token-level Response-visual Attention Guidance for Multimodal LLMs Knowledge DistillationJaehyun Jang, Eunseop Yoon, Hee Suk Yoon, SooHwan Eom, Mark A. Hasegawa-Johnson, and Chang D. Yoo2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. Teacher and student share a frozen SigLIP vision encoder and resolution; this is not cross-resolution KD. The official code repository lists training/evaluation scripts as planned. Attention similarity correlations are not causal guarantees of faithful reasoning.TRAG distills how a multimodal teacher attends to visual tokens while producing a response. It aggregates middle-layer response-to-vision attention, renormalizes over visual keys and assigns each response token its own forward/reverse KL mixture using teacher attention entropy. This supplements language and unimodal distillation losses rather than replacing the full training recipe.Official ECCV poster record · Official code

Reading question: Are the grids, semantic units and probability supports genuinely aligned before applying the loss?

Evidence boundary: SFKD has architecture-specific regressions; TRAG shares a visual encoder and resolution. Neither establishes cross-magnification performance.

Perceptual experts, geometry and explicit grounding tokens

Continuous expert-supervised visual tokens, geometry-aware fusion and coarse-to-fine grounding alter how a language model accesses visual evidence. Compare the representation interface, auxiliary supervision and deployed computation before comparing answer scores.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual TokensYiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, and XuDong Wang2026 · ECCV · acceptedECCV 2026 Oral verified in the official program (event 6085); scientific claims use the conference-hosted manuscript #5952. Publisher proceedings/version of record was not verified on 2026-09-12. First preprint appeared in November 2025; stable ID preserves that origin.CoVT teaches a VLM to generate continuous visual thoughts alongside ordinary language. These latent tokens acquire perceptual meaning through supervision from segmentation, depth, edge, and self-supervised feature experts. The answer conditions on this compact visual reasoning sequence; dense predictions are optional explanatory outputs rather than mandatory inference-time tool calls.ECCV program · Conference manuscript · Author project · Code
Make Geometry Matter for Spatial ReasoningShihua Zhang, Qiuhong Shen, Shizun Wang, Tianbo Pan, and Xinchao Wang2026 · ECCV · acceptedECCV 2026 Oral verified in the official program (event 6086); scientific claims use the conference-hosted manuscript #482. Publisher proceedings/version of record was not verified on 2026-09-12.GeoSR asks whether a spatial-reasoning VLM actually uses its geometry tokens. It combines training-time masking of appearance tokens with learned token-and-channel-wise fusion, so geometric evidence becomes useful rather than an ignored auxiliary stream. The contribution is controlled utilization of pretrained geometry, not simply adding another encoder.ECCV program · Conference manuscript · Author project · Code

Reading question: Does a claimed reasoning gain arise from more visual evidence, better alignment, a changed output space or a larger inference budget?

Evidence boundary: Expert distillation is not proof of faithful reasoning; stronger pointing or spatial QA need not improve every tracking metric.

Long-video memory, scripts and entity continuity

Retrieval-based episodic memory, audio-visual script generation and entity-centric evaluation answer different long-video questions. Track what is stored, how evidence is retrieved, and whether identities and events remain consistent across time.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
OmniScript: Towards Audio-Visual Script Generation for Long-Form Cinematic VideoJunfu Pu, Yuxin Chen, Teng Wang, and Ying Shan2026 · ECCV · acceptedECCV 2026 Oral verified in the official program (event 6045); scientific claims use the conference-hosted manuscript #9919. Publisher proceedings/version of record was not verified on 2026-09-12.OmniScript converts cinematic audio and video into a hierarchical scene-event script. The model first reasons about plot and character relationships, then emits time-grounded fields for actions, dialogue, expressions, and audio cues. Its research contribution couples an audio-visual architecture, structured training data, and an evaluation protocol that separately checks what happened and when.ECCV program · Conference manuscript · Author project · Code

Reading question: Can the system recover the right episode and preserve who did what, when, and in which scene?

Evidence boundary: Compressed memory can omit decisive events; fluent scripts and pooled benchmark scores can conceal identity or temporal failures.

Distilling visual context and sensitivity to evidence

SCOPD teaches a sparse-context student from a full-context EMA teacher on student rollouts. CW-OPD instead tests how predictions change when answer-critical visual evidence is edited between two worlds.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language ModelsAhmadreza Jeddi, Enming Zhang, Jasper Gerigk, Hakki Karaimer, Mozhgan Nasr Azadani, Jiayun Luo, Minh Ngoc Le, Gholamali Aminian, Hugo Buurmeijer, Yongchao Chen, Leonid Sigal, Igor Gilitschenski, Konstantinos G. Derpanis, Marco Pavone, and Babak Taati2026 · arXiv · preprintPreprint checked 2026-10-02; project page still labels code and models as forthcoming. Visual-token retention is not an optical-magnification intervention.SCOPD treats visual-token pruning partly as a representation-utilization problem. A full-context teacher supervises responses generated by the sparse student; SCOPD+ concentrates the supervision on visually sensitive positions.Versioned full text · Author project page
Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language ModelsYuanhao Sun, Huawei Ji, Jiaxin Ding, Luoyi Fu, and Xinbing Wang2026 · arXiv · preprintPreprint checked 2026-10-02. The paper calls the method CW-OPD; its linked code repository uses the name CWAD. No conference acceptance inferred.CW-OPD supervises how a VLM prediction changes when answer-critical visual evidence changes. It combines ordinary per-world distillation with a cross-world belief-transition target, rather than assuming a correct answer proves the student used the right evidence.Versioned full text · Author implementation (repository name CWAD)

Reading question: Does the student learn visual evidence, or merely the teacher’s answer distribution?

Evidence boundary: Token pruning and counterfactual image editing are not WSI acquisition changes. Track EMA updates, teacher-only hints, editor quality and training cost before comparing distillation methods.

Distill visual grounding by withholding textual shortcuts

Hide to See masks parts of the student’s reasoning prefix while the teacher sees the complete prefix. The asymmetric distillation setup asks the student to rely more on the image rather than copy answers from preceding text.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM DistillationSeonghoon Yu, Dongjun Nam, Byung-Kwan Lee, and Jeany Son2026 · NeurIPS · acceptedNeurIPS 2026 acceptance is author-reported by NVIDIA Research and the official implementation, checked 8 October 2026. No exact title or author match was found in the dated official oral/poster snapshot; presentation tier and final proceedings are unverified. Review uses arXiv v4. Training-time textual-prefix masking is not WSI magnification transfer.Masking-KD distills a reasoning VLM while hiding selected textual shortcuts from the student. The teacher sees the complete image, question, and earlier response; the student loses access to its most attended earlier reasoning tokens and must predict the same next-token distribution with the image still available. The intervention is reasoning-prefix masking, not image masking or an inference-time pruning algorithm.Versioned full text · Official implementation · Author project · Author acceptance record

Reading question: Does masking improve image-dependent reasoning, or only alter the difficulty of next-token prediction?

Evidence boundary: Acceptance is author-reported; no exact match was located in the inspected official program snapshot. Prefix masking changes training context, not image resolution, and image-dependence proxies are not clinical grounding tests.

Complete maintained paper route

Section titled “Complete maintained paper route”

47 canonical readings · 2015–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.

Show all 47 readings in maintained order

  37. #372026ACLPeer reviewedL2–L3

    Afri-MCQA: Multimodal Cultural Question Answering for African Languages

    Atnafu Lambebo Tonja, Srija Anand, Emilio Villa-Cueva, Israel Abebe Azime, Jesujoba Oluwadara Alabi, Muhidin A. Mohamed, Debela Desalegn Yadeta, Negasi Haile Abadi, Abigail Oppong, Nnaemeka Casmir Obiefuna, Idris Abdulmumin, Naome A Etori, Eric Peter Wairagala, Kanda Patrick Tshinu, Imanigirimbabazi Emmanuel, Gabofetswe Malema, Alham Fikri Aji, David Ifeoluwa Adelani, and Thamar Solorio

    Peer-reviewed ACL 2026 long paper and Best Social Impact Paper award recipient.

    Official artifacts: Author manuscript · Official dataset

    Metadata checked

  39. #392026arXivPreprintL2–L3

    VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

    Junxiang Xu, Ruisi Wang, Fanyi Pu, Maijunxian Wang, Ran Ji, Tongxi Zhou, Chenyang Gu, Jing Zuo, Hongcan Xiao, Yimeng Geng, Wanqi Yin, Wei Chen, Oscar Qian, Zhengan Yan, Ziqi Huang, Haiwen Diao, Liang Pan, Bo Li, Xiangyu Fan, Dezhi Luo, Fengyuan Yu, Zehong Zhao, Qingying Gao, Tinghui Zhu, Yilan Zhang, Jingqi Tong, Pinyuan Feng, Zhengze Jiang, Letian Wang, Ziyu Guo, Renrui Zhang, Jieneng Chen, Sonia Joseph, Constantin Venhoff, Saman Motamed, Mengyue Yang, Chandra Sripada, Alan Yuille, Philip Torr, Lvmin Zhang, Vikash Kumar, Daniel Khashabi, Nikolaus Kriegeskorte, Raphaël Millière, Vincent C. Müller, Anyi Rao, Quan Wang, Ziwei Liu, Dahua Lin, Lei Yang, Hokin Deng, and Zhongang Cai

    arXiv preprint first submitted 26 August 2026; not peer reviewed as of the audit date.

    Official artifacts: Official project · Official code and benchmark

    Metadata checked

  40. #402026ECCVAccepted · proceedings pendingL3

    Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens

    Yiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, and XuDong Wang

    ECCV 2026 Oral verified in the official program (event 6085); scientific claims use the conference-hosted manuscript #5952. Publisher proceedings/version of record was not verified on 2026-09-12. First preprint appeared in November 2025; stable ID preserves that origin.

    Official artifacts: ECCV program · Conference manuscript · Author project · Code

    Metadata checked

  43. #432026ECCVAccepted · proceedings pendingL3

    Mimicking Radiologists: A Coarse-to-Fine Framework with Structural Sparse Tokens for Dual-LLM Computed Tomography Report Generation

    Hong Liu, Dong Wei, Yefeng Zheng, Xian Wu, and Liansheng Wang

    ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #9339. Publisher proceedings/version of record was not verified on 2026-09-12. The paper-linked CTRG repository README still describes the earlier Structure Observation Driven Image-Text Contrastive Learning work and LLaMA-2-7B. A matching dual-LLM/GRPO release was not verified; do not treat this link as reproduction evidence.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Paper-linked repository (earlier method; matching release unverified)

    Metadata checked

  44. #442026ECCVAccepted · proceedings pendingL3

    Token-level Response-visual Attention Guidance for Multimodal LLMs Knowledge Distillation

    Jaehyun Jang, Eunseop Yoon, Hee Suk Yoon, SooHwan Eom, Mark A. Hasegawa-Johnson, and Chang D. Yoo

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. Teacher and student share a frozen SigLIP vision encoder and resolution; this is not cross-resolution KD. The official code repository lists training/evaluation scripts as planned. Attention similarity correlations are not causal guarantees of faithful reasoning.

    Official artifacts: Official ECCV poster record · Official code

    Metadata checked

  45. #452026arXivPreprintL3

    SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models

    Ahmadreza Jeddi, Enming Zhang, Jasper Gerigk, Hakki Karaimer, Mozhgan Nasr Azadani, Jiayun Luo, Minh Ngoc Le, Gholamali Aminian, Hugo Buurmeijer, Yongchao Chen, Leonid Sigal, Igor Gilitschenski, Konstantinos G. Derpanis, Marco Pavone, and Babak Taati

    Preprint checked 2026-10-02; project page still labels code and models as forthcoming. Visual-token retention is not an optical-magnification intervention.

    Official artifacts: Versioned full text · Author project page

    Metadata checked

  47. #472026NeurIPSAccepted · proceedings pendingL3

    Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM Distillation

    Seonghoon Yu, Dongjun Nam, Byung-Kwan Lee, and Jeany Son

    NeurIPS 2026 acceptance is author-reported by NVIDIA Research and the official implementation, checked 8 October 2026. No exact title or author match was found in the dated official oral/poster snapshot; presentation tier and final proceedings are unverified. Review uses arXiv v4. Training-time textual-prefix masking is not WSI magnification transfer.

    Official artifacts: Versioned full text · Official implementation · Author project · Author acceptance record

    Metadata checked