Paper reading guide

This is the central index for the site’s paper-reading routes. It is a curated starting sequence, not an exhaustive bibliography or a leaderboard. Read the papers in order within a lane, and use the linked guides for the concepts, checkpoints, and evaluation questions that make the sequence useful.

Guide coverage: detailed learning paths for ML/deep learning, computer vision, NLP, LLMs, and natural-image Vision & Language; a research-level computational-pathology curriculum covering WSI engineering, dense morphology, MIL, self-supervision, foundation encoders, pathology VLMs, spatial/molecular modeling, and clinical evidence; and a critical-reading protocol. This page connects those maintained routes.

Coverage boundary: this remains a curated curriculum rather than an exhaustive historical bibliography or a live leaderboard. Use the Paper Radar for provisional new arXiv discoveries; inclusion there is not an endorsement.

Review dates: each chapter distinguishes its route-wide audit date from individual paper metadata checks. Adding a new paper does not imply that every older paper, publication status, or artifact in that field was reverified.

The AI background registry and computational-pathology registry are searchable, structured views of the maintained routes. They expose year, publication state, semantic membership, official artifacts, and metadata-review date, and supply the same records shown in My Library.

Each detailed background chapter renders its complete maintained route from that catalog. Its shorter staged table explains the recommended conceptual spine; use the generated full list for all foundational papers, branches, benchmarks, and current endpoints in the maintained scope.

Looking specifically for self-supervised learning? Open the pathology SSL reading route. It integrates its maintained papers by objective lineage and pathology research question, with publication status, author artifacts, checkpoints, and a fair-comparison protocol.

How years and publication status are shown

The guide uses the peer-reviewed venue year when a paper has a version of record. When a widely read preprint appeared earlier, both dates are shown as “preprint year → venue year.” “arXiv” and “medRxiv” mean preprint, not a peer-reviewed venue. Model cards are labelled as artifacts rather than papers. Links go to the publisher, proceedings, anthology, OpenReview record, or author-posted preprint.

Find your reading route

1,117 unique papers · 53 detailed routes · 6 fields.

Choose a field, then a learning question. Each chapter has a short conceptual spine and a complete maintained reading list. Shared papers keep one catalog record; field and route counts overlap and should not be added together.

ML & deep learning13 routes · 390 unique papers

Mathematical foundations, classical learning, neural networks, generative models, and reliable evaluation. Start with the field overview →

• Math / statistics6 papers

  Linear algebra, calculus, probability, information theory, estimation, and statistical inference for ML.

• Classical ML14 papers

  Linear models, kernels, trees, ensembles, clustering, dimensionality reduction, and model selection.

• Neural networks54 papers

  Backpropagation, initialization, optimization, normalization, regularization, and generalization.

• Generative / representation65 papers

  Latent-variable, adversarial, autoregressive, flow, diffusion, and representation-learning objectives.

• Knowledge distillation140 papers

  Teacher–student transfer through output, feature, relation, sequence, self-distillation, multimodal, and domain-specific objectives.

• Loss / objectives61 papers

  Proper scoring, imbalance, metric geometry, overlap, noisy-label robustness, calibration, and multimodal alignment objectives.

• Transfer / adaptation / continual31 papers

  Representation transfer, domain adaptation, few-shot and meta-learning, semi-supervised learning, parameter-efficient adaptation, and continual plasticity.

• AI / reinforcement learning20 papers

  Search, logic, probabilistic reasoning, Markov decision processes, value learning, policy learning, planning, and offline RL.

• Evaluation / reproducibility37 papers

  Splits, metrics, calibration, uncertainty, ablations, statistical comparison, shift, and reproducible evidence.

• Causal inference9 papers

  Causal graphs, identification, interventions, treatment effects, orthogonal estimation, and causal evaluation with machine learning.

• Interpretability33 papers

  Feature attribution, local and global explanations, counterfactual reasoning, faithfulness, stability, and human interpretation.

• Privacy / federated12 papers

  Distributed optimization, federated averaging, differential privacy, privacy accounting, leakage, and deployment trade-offs.

• Graph / geometric ML10 papers

  Message passing, graph expressivity, invariant and equivariant architectures, relational structure, and geometric inductive bias.

Computer vision8 routes · 290 unique papers

Image formation, recognition, dense prediction, representation learning, generation, video, 3D, and evaluation. Start with the field overview →

• Image / recognition41 papers

  Image formation, features, convolutional networks, classification, transfer learning, and modern backbones.

• Detection / segmentation51 papers

  Localization, object detection, semantic and instance segmentation, open-world detection, and promptable masks.

• Visual representations86 papers

  Supervised transfer, contrastive learning, self-distillation, masked prediction, predictive learning, and scaling.

• Generative vision48 papers

  Autoregressive image models, VAEs, GANs, diffusion, flow matching, controllable synthesis, and editing.

• Restoration / inverse problems33 papers

  Denoising, super-resolution, paired and unpaired translation, transformer restoration, learned image priors, and measurement-consistent reconstruction.

• Video / 3D / embodied43 papers

  Temporal models, tracking, video representation, geometry, neural rendering, 3D reconstruction, and embodied perception.

• Vision evaluation39 papers

  Dataset design, metrics, corruption and distribution shift, shortcuts, calibration, fairness, and efficiency.

• 3D / spatial vision36 papers

  Depth, correspondence, camera geometry, neural rendering, point tracking, static and dynamic reconstruction, and spatial foundation models.

Natural-language processing6 routes · 100 unique papers

Text representations, sequence models, transformers, language tasks, retrieval, generation, and evaluation. Start with the field overview →

• Text / sequence models21 papers

  Tokenization, sparse and distributed representations, recurrent networks, encoder–decoders, and attention.

• Transformers / pretraining45 papers

  Transformer mechanics, encoder, decoder, and encoder–decoder objectives, adaptation, and efficient transfer.

• NLP tasks / evaluation20 papers

  Classification, extraction, retrieval, QA, translation, summarization, generation, factuality, and multilingual evaluation.

• Structure / extraction8 papers

  Dependency syntax, spans, coreference, entities, relations, events, structured prediction, and information extraction.

• Multilingual / low resource14 papers

  Cross-lingual transfer, multilingual pretraining, low-resource datasets, many-to-many translation, language coverage, and evaluation equity.

• Generation / MT evaluation26 papers

  Summarization-specific pretraining, machine translation, reference and learned metrics, factuality, and human evaluation.

Large language models8 routes · 163 unique papers

Language-model scaling, post-training, retrieval, tools, agents, efficiency, safety, and evaluation. Start with the field overview →

• LLM pretraining40 papers

  Decoder-only modelling, scaling laws, data and token design, mixture-of-experts, context, and systems efficiency.

• LLM post-training36 papers

  Instruction tuning, parameter-efficient adaptation, preference learning, reward modelling, and alignment trade-offs.

• RAG / tools / agents32 papers

  Retrieval-augmented generation, long context, tool use, planning, memory, agents, and system-level evaluation.

• LLM evaluation / safety41 papers

  Capabilities, contamination, factuality, calibration, bias, harmful behavior, privacy, interpretability, and inference cost.

• LLM systems31 papers

  Tensor, pipeline and data parallelism, state partitioning, attention kernels, quantization, scheduling, KV memory, and production serving.

• Reasoning / verification21 papers

  Rationale prompting, self-improvement, search over thoughts, process supervision, verifiable rewards, and test-time scaling.

• LLM interpretability11 papers

  Transformer circuits, superposition, sparse features, causal tracing, model editing, unlearning objectives, and evaluation.

• Code / software engineering12 papers

  Code representation, program synthesis, infilling, functional evaluation, repository repair, iterative reuse, and contamination-aware benchmarks.

Natural-image vision & language8 routes · 253 unique papers

Natural-image image–text representations, captioning, VQA, grounding, multimodal LLMs, generation, video, agents, and evaluation. Start with the field overview →

• Image–text representation55 papers

  Dual encoders, fusion encoders, contrastive and captioning objectives, retrieval, zero-shot transfer, and data curation.

• Captioning / VQA47 papers

  Image description, visual question answering, compositional reasoning, scene text, documents, charts, and spatial reasoning.

• Grounding / open vocabulary35 papers

  Referring expressions, phrase grounding, open-vocabulary detection and segmentation, promptable masks, and grounded generation.

• Multimodal LLMs66 papers

  Vision-to-language bridges, interleaved pretraining, visual instruction tuning, high-resolution inputs, reasoning, and adaptation.

• Generation / video / agents72 papers

  Text-to-image and video, unified understanding and generation, multimodal agents, GUI control, robotics, and navigation.

• V&L evaluation / safety82 papers

  Grounding, hallucination, compositionality, multimodal reasoning, data provenance, bias, adversarial robustness, and safety.

• Document intelligence19 papers

  OCR, layout-aware and OCR-free document parsing, screenshots, charts, structured generation, and document evaluation.

• Multimodal retrieval21 papers

  Universal multimodal embeddings, instruction-conditioned retrieval, hard-negative training, document and video retrieval, and reranking.

Computational pathology10 routes · 342 unique papers

All curated pathology routes, from slide learning to clinical evaluation. Start with the field overview →

• WSI engineering51 papers

  Slide formats, quality control, stain handling, tiling, and reproducible feature extraction.

• Cell / tissue models47 papers

  Dense tissue labels, detection, nuclei instances, cell phenotyping, and pathomics.

• MIL / WSI81 papers

  Weak supervision, multiple-instance aggregation, and whole-slide context.

• Pathology SSL75 papers

  A maintained objective-to-pathology route spanning foundations, domain adaptation, WSI hierarchy, and current research.

• Foundation encoders72 papers

  Tile, slide, and image–text foundation representations and their scale-up.

• Pathology VLM81 papers

  Image–text alignment, multimodal assistants, and whole-slide multimodal models.

• Spatial / molecular47 papers

  Molecular prediction, spatial transcriptomics, graphs, multimodal fusion, and survival.

• Evaluation94 papers

  Reporting guidance, controlled benchmarks, generalization, and clinical claims.

• Generative pathology19 papers

  Patch and gigapixel synthesis, morphology and topology control, virtual staining, multimodal generation, privacy, and utility evaluation.

• Robustness / shift / privacy28 papers

  Site, scanner, stain, and population shift; shift monitoring; federated collaboration; privacy accounting; and external validation.

Choose a depth

Level	Intended reader	Read until you can…
L0 · orientation	new undergraduate or a reader changing fields	define the problem, input, output, supervision, and primary metric
L1 · foundation	advanced undergraduate or early master’s student	reconstruct the main objective and data flow without the paper
L2 · research	graduate student or early-stage researcher	audit splits, baselines, ablations, uncertainty, and reproducibility
L3 · synthesis	independent researcher	compare incompatible assumptions, design a fair experiment, and state what would falsify a claim

Difficulty describes the prerequisite burden, not paper quality. A short L3 paper can require more judgment than a long L1 paper.

Use one reading protocol everywhere

Before beginning a lane, open How to read a paper. For each paper:

1. Triage: write the problem, central claim, unit of analysis, and one threat to generalization.
2. Reconstruct: trace one sample from raw input through preprocessing, model, output, and metric.
3. Audit: recover the split unit, selection procedure, baseline budgets, uncertainty, and possible data overlap.
4. Test: reproduce the smallest decisive mechanism or table row.
5. Recall: close the source and write the method, evidence, limitation, and three questions from memory.

For WSI papers, always retain the patient → case → specimen → block → slide → tile hierarchy. A tile-level split is not a patient-level split.

Background paper ladders

These lanes supply the vocabulary needed by the computational-pathology route. Readers who already know a topic can use the exit task in its detailed guide instead of reading every anchor again.

A. Machine learning and deep learning

Start with the ML and deep-learning foundations. It also points to mathematics, statistical learning, PyTorch, evaluation, and milestone projects that are more suitable than research papers for first exposure.

Order	Level	Primary reading · year · venue/status	Read it for
1	L1	Learning representations by back-propagating errors — 1986 · Nature (journal)	computational graphs, credit assignment, and gradient-based representation learning
2	L1	Dropout: A Simple Way to Prevent Neural Networks from Overfitting — 2014 · JMLR (journal)	regularization as stochastic model perturbation
3	L1–L2	Adam: A Method for Stochastic Optimization — 2014 arXiv preprint → 2015 · ICLR (conference)	adaptive optimization and the difference between an update rule and a convergence guarantee
4	L1–L2	Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift — 2015 · ICML (conference)	optimization, normalization, train/evaluation behavior, and batch dependence

Exit check: derive the gradient for one layer, identify every source of regularization in an experiment, and explain why a lower training loss does not prove better generalization. Continue through the dedicated mathematics/statistics, classical-ML, neural-network, generative/representation, AI/RL, and evaluation chapters for the complete route.

B. Computer vision

Use the computer-vision learning path for receptive fields, task definitions, metrics, checkpoints, and pathology-specific cautions.

Order	Level	Primary reading · year · venue/status	Read it for
1	L1	Deep Residual Learning for Image Recognition — 2016 · CVPR (conference)	residual paths and optimization at depth
2	L1–L2	Fully Convolutional Networks for Semantic Segmentation — 2015 · CVPR (conference); U-Net: Convolutional Networks for Biomedical Image Segmentation — 2015 · MICCAI (conference)	dense prediction, resolution recovery, and skip connections
3	L2	Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks — 2015 · NeurIPS (conference); Mask R-CNN — 2017 · ICCV (conference)	proposal-based detection and instance segmentation
4	L2	An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale (ViT) — 2020 preprint → 2021 · ICLR (conference); Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows — 2021 · ICCV (conference)	tokenized images, inductive bias, hierarchy, and attention cost
5	L2	A Simple Framework for Contrastive Learning of Visual Representations (SimCLR) — 2020 · ICML (conference); Emerging Properties in Self-Supervised Vision Transformers (DINO) — 2021 · ICCV (conference); Masked Autoencoders Are Scalable Vision Learners (MAE) — 2022 · CVPR (conference)	three influential self-supervised objective families

Exit check: calculate a receptive field, choose a task-appropriate metric, and compare a CNN and ViT by data regime, locality, resolution, and compute rather than name.

C. Natural-language processing

Use the NLP learning path for tokenization, attention shapes, clinical-text leakage, and generation evaluation.

Order	Level	Primary reading · year · venue/status	Read it for
1	L1	Efficient Estimation of Word Representations in Vector Space — 2013 · arXiv preprint	predictive embeddings and distributional representations
2	L1	Sequence to Sequence Learning with Neural Networks — 2014 · NeurIPS (conference)	the encoder–decoder abstraction and its bottleneck
3	L1–L2	Neural Machine Translation by Jointly Learning to Align and Translate — 2014 arXiv preprint → 2015 · ICLR (conference)	content-dependent attention over encoder states
4	L2	Attention Is All You Need — 2017 · NeurIPS (conference)	self-attention, positional information, masking, and parallel sequence modelling
5	L2	BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding — 2019 · NAACL-HLT (conference); Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5) — 2019 preprint → 2020 · JMLR (journal)	encoder pretraining versus a text-to-text encoder–decoder interface

Exit check: write the shapes in multi-head attention, explain the attention mask, and identify when a report benchmark can be solved through label or template leakage.

D. Large language models and the multimodal bridge

Use the LLM learning path for pretraining, post-training, retrieval, tools, agents, evaluation, and safety. Rows 4–5 are the bridge into the separate natural-image Vision & Language route below.

Order	Level	Primary reading · year · venue/status	Read it for
1	L2	Scaling Laws for Neural Language Models — 2020 · arXiv preprint; Training Compute-Optimal Large Language Models — 2022 · NeurIPS (conference)	empirical scaling regimes and compute allocation—not guarantees of usefulness
2	L2	Training Language Models to Follow Instructions with Human Feedback (InstructGPT) — 2022 · NeurIPS (conference)	supervised instruction tuning, preference data, reward modelling, and RLHF
3	L2	Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — 2020 · NeurIPS (conference)	separating retrieval quality from generation faithfulness
4	L2	Learning Transferable Visual Models From Natural Language Supervision (CLIP) — 2021 · ICML (conference); Sigmoid Loss for Language Image Pre-Training (SigLIP) — 2023 · ICCV (conference)	dual-encoder alignment, retrieval, and zero-shot scoring
5	L2–L3	BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models — 2023 · ICML (conference); Visual Instruction Tuning (LLaVA) — 2023 · NeurIPS (conference)	bridges to frozen models versus multimodal instruction tuning

Exit check: identify the trained and frozen components, source of image–text supervision, visual evidence level, and evaluation that actually tests grounding.

E. Natural-image Vision & Language

The dedicated field overview separates natural-image multimodality from pathology VLMs and from text-only LLMs. Its comprehensive reading map is organized by scientific output rather than model release date:

Route	Detailed guide	Shortest concept sequence
image–text representation	alignment and retrieval	DeViSE → CLIP/ALIGN → LiT/CoCa/SigLIP → data filtering and patch-level alignment
captioning and VQA	captioning, VQA, OCR, and reasoning	Show and Tell/Show-Attend-Tell → VQA v1/v2 → GQA/TextVQA/OK-VQA → modern generative interfaces
grounding and open vocabulary	boxes, masks, and referring	MDETR → RegionCLIP/GLIP/OWL-ViT → ODISE/Grounding DINO/SAM → reasoning segmentation and absent-referent rejection
multimodal LLMs	bridges and visual instruction tuning	Frozen/Flamingo → BLIP-2 → LLaVA/InstructBLIP → high-resolution and current multimodal reasoning systems
generation, video, and agents	multimodal generation and action	VQ-VAE/latent diffusion → unified understanding–generation → video–language → GUI, robotics, and observation-planning agents
evidence and safety	evaluation, data, and safety	Winoground/CHAIR → MMMU/MM-Vet → hallucination and grounding diagnostics → multimodal red-teaming

Start with CLIP (2021), MDETR (2021), BLIP-2 (2023), and LLaVA (2023). Then branch by the output your research actually needs: embedding, text, box, mask, image/video, or action.

For every multimodal system, record the image granularity, visual encoder, token budget, connector/fusion mechanism, language backbone, trained/frozen components, image–text/instruction/preference data, supported outputs, and evidence that the image was necessary. Always run text-only, image-only, shuffled-image, and absent-evidence controls where the task permits them.

Computational-pathology research route

Follow these stages in order unless your research question gives you a reason to branch. The field overview is the map of pathology context, WSI processing, model families, tasks, and evaluation. Every row below links to a dedicated maintained chapter.

Stage	Reading	Status
0. Tissue, labels, WSI systems, and data	Computational-pathology overview, WSI engineering and QC, and the WSI Data Atlas	concepts, primary systems/QC papers, reproducibility contract, and reviewed dataset-family records
1. Dense morphology	Cell, gland, and tissue modeling	tissue segmentation, cell detection, nuclei instances, phenotyping, benchmarks, and metrics
2. Slide-level learning	MIL and whole-slide learning	prerequisite-aware MIL taxonomy, dated primary readings, controlled comparisons, and WSI projects
3. Representation objectives	Pathology self-supervised learning	comprehensive objective-to-pathology ladder with 52 dated papers
4. Foundation representations	Foundation encoders	tile, slide, multimodal, and medical-vision releases with access and evaluation guidance
5. Language and multimodal systems	Pathology vision–language models	alignment, retrieval, generation, grounding, whole-slide multimodality, and safety
6. Spatial biology, molecular targets, and outcomes	Spatial, molecular, and outcome modeling	molecular screening, spatial transcriptomics, graphs, histology–omics fusion, and survival
7. Evidence and translation	Evaluation and clinical translation and the paper-reading audit	claim-matched metrics, external/reader/prospective evidence, risk of bias, and reporting standards

The compact lanes below contain only high-value anchors. They are arranged by method dependency so that representation objectives precede models that consume their features; use the stage table above and the dedicated chapters for the full core route.

A. WSI problem formulation and engineering

1. L1–L2: Read the field overview before an architecture paper. Recover specimen preparation, stain, scanner, microns per pixel (MPP), pyramid level, tissue detection, tile sampling, and label provenance.
2. L1–L2: Read OpenSlide: A Vendor-Neutral Software Foundation for Digital Pathology — 2013 · Journal of Pathology Informatics, then the WSI engineering and QC chapter. Trace how one physical specimen becomes a multiresolution image, QC decision, coordinate, tile, and embedding.
3. L2: Select one WSI family from the Data Atlas. Verify access, license, hierarchy, and split unit at the official source before designing an experiment.

Required output: one end-to-end diagram from patient and slide selection to tiles, features, aggregation, prediction, and patient-level evaluation. Annotate every point at which leakage or selection bias can enter.

B. Dense morphology

Use the cell, gland, and tissue modeling chapter for task contracts and the full benchmark lineage. The shortest anchor route is:

1. L1: U-Net: Convolutional Networks for Biomedical Image Segmentation — 2015 · MICCAI, for dense prediction and skip-connected localization.
2. L2: HoVer-Net: Simultaneous Segmentation and Classification of Nuclei in Multi-Tissue Histology Images — 2020 · Medical Image Analysis, for separating touching nuclei and assigning cell types.
3. L3: CellViT: Vision Transformers for Precise Cell Segmentation and Classification — 2024 · Medical Image Analysis, for transferring transformer representations into a pathology instance-prediction system.

C. Self-supervised representations

The pathology SSL guide contains the comprehensive, annotated SSL registry, design questions, and fair-comparison protocol. This central route intentionally keeps only the objective families and the transition into pathology:

1. General objectives, L1–L2: A Simple Framework for Contrastive Learning of Visual Representations (SimCLR) — 2020 · ICML (conference); then Momentum Contrast for Unsupervised Visual Representation Learning (MoCo) — 2020 · CVPR (conference).
2. Collapse avoidance and self-distillation, L2: Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning (BYOL) — 2020 · NeurIPS (conference); then Emerging Properties in Self-Supervised Vision Transformers (DINO) — 2021 · ICCV (conference).
3. Masked prediction, L2: Masked Autoencoders Are Scalable Vision Learners (MAE) — 2022 · CVPR (conference); then iBOT: Image BERT Pre-Training with Online Tokenizer — 2022 · ICLR (conference).
4. Pathology adaptation, L2–L3: RetCCL: Clustering-Guided Contrastive Learning for Whole-Slide Image Retrieval — 2022 online → 2023 · Medical Image Analysis (journal); Transformer-Based Unsupervised Contrastive Learning for Histopathological Image Classification (CTransPath) — 2022 · Medical Image Analysis (journal); then Scaling Vision Transformers to Gigapixel Images via Hierarchical Self-Supervised Learning (HIPT) — 2022 · CVPR (conference).
5. Scale and slide context, L3: Towards a General-Purpose Foundation Model for Computational Pathology (UNI) — 2024 · Nature Medicine (journal); then A Whole-Slide Foundation Model for Digital Pathology from Real-World Data (Prov-GigaPath) — 2024 · Nature (journal).

Do not ask only which loss “wins.” Record view construction, scale, false negatives, hierarchy, pretraining overlap, exact weights, and downstream adaptation.

D. Multiple-instance and weakly supervised slide learning

The full MIL and whole-slide learning chapter extends this short sequence with the classical MIL assumptions, instance-selection, attention, clustering, Transformer, graph, multi-scale, survival, and slide-foundation lineages. Use the table below as the shortest entry route.

Order	Level	Primary reading · year · venue/status	Question to carry forward
1	L1–L2	Attention-based Deep Multiple Instance Learning — 2018 · ICML (conference)	Why must bag aggregation be permutation invariant, and what does attention weight actually establish?
2	L2	Clinical-grade computational pathology using weakly supervised deep learning on whole slide images — 2019 · Nature Medicine (journal)	How do sampling and dataset scale interact with sparse positive evidence?
3	L2	Data-efficient and weakly supervised computational pathology on whole-slide images (CLAM) — 2021 · Nature Biomedical Engineering (journal)	What does instance-level clustering add to attention MIL, and how is localization evaluated?
4	L2–L3	Dual-Stream Multiple Instance Learning Network for Whole Slide Image Classification With Self-Supervised Contrastive Learning (DSMIL) — 2021 · CVPR (conference)	How does the critical instance interact with the remaining bag, and which gains come from the encoder?
5	L2–L3	TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification — 2021 · NeurIPS (conference)	Which relationships among instances are modelled, and what spatial assumptions or memory costs are introduced?

Fair comparison: freeze the same encoder; cache the same patient-split features; use the same folds, tuning budget, stopping rule, and metrics; report uncertainty and external-site performance. An attention map is not automatically a validated explanation.

E. Foundation encoders and slide models

First read the foundation-encoder guide. It distinguishes tile, slide, vision–language, and general medical encoders and records the exact scientific and access questions needed to compare releases.

Use this order to compare roles rather than to construct a global ranking:

1. Historical domain encoders, L2: CTransPath — 2022 · Medical Image Analysis (journal); then Scaling Self-Supervised Learning for Histopathology with Masked Image Modeling (Phikon) — 2023 · medRxiv preprint, together with the official Phikon model card — official model artifact, not a paper; checked 2026.
2. Pathology image encoders, L2–L3: Towards a General-Purpose Foundation Model for Computational Pathology (UNI) — 2024 · Nature Medicine (journal); then A Foundation Model for Clinical-Grade Computational Pathology and Rare Cancers Detection (Virchow) — 2024 · Nature Medicine (journal).
3. Pathology image–text alignment, L2–L3: A Visual-Language Foundation Model for Computational Pathology (CONCH) — 2024 · Nature Medicine (journal).
4. Tile plus slide context, L3: A Whole-Slide Foundation Model for Digital Pathology from Real-World Data (Prov-GigaPath) — 2024 · Nature (journal).
5. Medical—not pathology-specific—context, optional L2: the official MedSigLIP model card — 2025 · model release/card, not a paper; its linked technical reference is the MedGemma Technical Report — 2025 · arXiv technical report.
6. Comparative evidence, L3: Benchmarking foundation models as feature extractors for weakly supervised computational pathology — published online 2025 · Nature Biomedical Engineering (journal; volume year 2026).

Model names are not sufficient experimental identifiers. Record the checkpoint, transforms, input size and MPP, embedding level, pretraining data, access terms, adaptation, aggregator, evaluation population, and possible overlap. The current page is a maintained study route rather than a timeless global ranking.

F. Pathology vision–language and generative systems

Read the natural-image Vision & Language route first. The dedicated pathology VLM chapter provides the broader taxonomy and paper route; the sequence here is the shortest entry:

1. L2: Learning Transferable Visual Models From Natural Language Supervision (CLIP) — 2021 · ICML (conference); then Sigmoid Loss for Language Image Pre-Training (SigLIP) — 2023 · ICCV (conference), for the general dual-encoder objectives.
2. L2–L3: A Visual-Language Foundation Model for Computational Pathology (CONCH) — 2024 · Nature Medicine (journal), for pathology-specific image–text representation learning and zero-shot transfer.
3. L3: A Multimodal Generative AI Copilot for Human Pathology (PathChat) — 2024 · Nature (journal), for a pathology vision–language assistant and multimodal instruction tuning.
4. L3: A Multimodal Whole-Slide Foundation Model for Pathology (TITAN) — 2025 · Nature Medicine (journal), for multimodal whole-slide representation learning.

For every system, separate tile/region evidence from whole-slide evidence, retrieval from generation, visual grounding from memorized medical knowledge, and automated metrics from blinded expert review. Audit report-derived labels and captions for leakage. These anchors are a concept ladder; use the dedicated chapter for the broader, dated catalog and evaluation protocol.

G. Evaluation, generalization, and clinical claims

1. L1–L3: Read evaluation and clinical translation and apply the site’s computational-pathology audit to every paper, including foundation-model papers.
2. L2: Use TRIPOD+AI — 2024 · BMJ (reporting guideline), to check transparent reporting for prediction-model development and validation. A reporting checklist does not by itself establish low bias or clinical utility.
3. L3: Read the foundation-model benchmark — published online 2025 · Nature Biomedical Engineering (journal; volume year 2026), as a study in controlled comparison and external datasets—not as a permanent universal ranking.
4. L3: Compare it with A benchmark study of vision and pathology foundation models for computational pathology — 2026 · Nature Communications (journal; currently an unedited publisher early-access manuscript). Audit its 32-model comparison for task-dependent rankings, external/out-of-domain evaluation, and small differences among leading models.
5. L3: Rebuild the evidence table with exact population, organ, stain, target, encoder version, WSI aggregator, adaptation, split, metric, confidence interval, external site, pretraining overlap, and resource cost.

Clinical utility requires more than retrospective discrimination. Look for calibration, operating points, subgroup and site shift, abstention, workflow comparator, reader-study design, prospective or temporal validation, and consequences of errors.

Short routes by research question

Research question	Minimum sequence
Which encoder should I use?	CV self-supervision → pathology SSL → foundation-encoder guide → one frozen-feature comparison on an external cohort
How should I classify a WSI with slide labels?	WSI formulation → attention MIL (2018) → CLAM (2021) → DSMIL or TransMIL (both 2021) → patient-level external evaluation
How do I build or assess a pathology VLM?	NLP transformers → CLIP (2021)/SigLIP (2023) → CONCH (2024) → PathChat (2024) or TITAN (2025) → grounding and leakage audit
Can morphology predict a biomarker or outcome?	label provenance → WSI/MIL pipeline → encoder comparison → calibration, censoring or prevalence analysis → external validation
How do I review a new foundation model?	exact artifact record → pretraining-overlap audit → equal frozen probe/MIL head → adaptation-budget comparison → deployment fit

What to record after each paper

Keep one row per exact paper version and one row per exact model checkpoint:

paper:
  title: ""
  version_or_doi: ""
  publication_status: ""
question:
  task: ""
  unit_of_analysis: patient | case | slide | region | tile
method:
  input_and_mpp: ""
  encoder_checkpoint: ""
  aggregation: ""
  supervision_and_objective: ""
evidence:
  population_and_split: ""
  baselines_and_budgets: ""
  metrics_and_uncertainty: ""
  external_validation: ""
risks:
  pretraining_overlap: none | possible | known | unknown
  leakage_and_confounding: []
  missing_information: []
recall:
  one_sentence_claim: ""
  strongest_evidence: ""
  strongest_limitation: ""
  questions: []

Do not commit personal notes, reading history, review schedules, or tracker backups to this repository. My Library stores records in the current browser profile and explains the privacy and backup boundary.

Keeping the guide current

• The curated ladders change through reviewed repository updates, not automatically.
• The Paper Radar surfaces recent arXiv records by transparent reading-priority signals; it does not score scientific quality.
• A paper should move from Radar to a learning ladder only after a human verifies its identity, relevance, primary links, publication state, artifacts, and limitations.
• Newer does not mean better. Seminal papers remain when they explain a lineage; newer papers are added when they introduce a distinct idea or materially change the evidence.

When proposing an addition, state where it belongs, which paper it follows, what new idea it teaches, and which existing item it might replace. That keeps this page a reading path instead of an unreviewed link dump.