Gradients and computational graphs
Needed to separate sensitivity from contribution.
Interpretability and attribution
An attractive explanation is not necessarily a faithful one. This route makes the target, baseline, intervention, scope, and falsification test explicit. It also treats explanation validation as a scientific step: parameter and label randomization reveal whether a saliency map depends on the trained model and the learned input–label relationship, while signed and structural comparisons prevent a familiar-looking image from becoming the entire verdict. Passing a sanity check is useful evidence, not proof of completeness or clinical value.
Sparse autoencoders for vision and pathology
Section titled “Sparse autoencoders for vision and pathology”Learn a small active set of features, then test what those features mean and what the model does with them. A sparse autoencoder (SAE) usually takes a frozen model activation, maps it into a larger dictionary with few active coordinates, and reconstructs the activation from the selected dictionary vectors. It is not the same as a masked image autoencoder, sparse attention, or simply making a classifier’s weights sparse. A foundational language-model SAE feature-dictionary study provides the starting point; the maintained paper tables below extend it into natural-image features and pathology applications.
In classical sparse coding, coefficients are optimized for each example; an SAE learns an encoder that predicts those coefficients directly. Neural superposition—multiple features sharing representation dimensions—motivates this analysis, but is not itself evidence that recovered coordinates correspond to distinct biological mechanisms.
Read the methods in this order
Section titled “Read the methods in this order”| Learning question | Paper family | What to inspect |
|---|---|---|
| How is a sparse dictionary trained? | Foundational language-model SAE, TopK, JumpReLU | Reconstruction target, active-feature budget, threshold learning, decoder normalization, dead features, and actual sparsity |
| Can general and specific features coexist? | Matryoshka SAE (Bussmann), CLIP Matryoshka SAE (Zaigrajew), SAE++ | Distinct nested and cascaded designs, granularity, feature absorption, and coverage; the shared name does not identify one method |
| What do image features represent? | PatchSAE, Universal SAE, VLM monosemantic features | Patch-token versus whole-image activations, cross-model correspondence, localization, semantic labels, and selective interventions |
| Can concepts be deliberately aligned or controlled? | Concept Bottleneck SAE | Which concepts are supplied, which are learned, and whether interpretability and steerability agree |
| Does this extend to pathology? | Pathology SAE features, ProtoMIL, SAE-1N, PATH-DISSECT, EXPOSE | Tile/TMA/WSI unit, expert annotations, classifier reliance, slide outcomes, and scanner/site transfer |
| How should success be measured? | Noncanonical-units / BatchTopK study, SAEBench, and controlled applications | Completeness and atomicity tests; several independent measurements rather than one reconstruction or interpretability score |
This is a prerequisite-aware reading plan, not a ranking of papers or a claim that all variants have been tested on pathology. Language-model SAE methods provide useful tools, but their reported results remain language-model results.
Keep the training choices distinct
Section titled “Keep the training choices distinct”TopK limits active coordinates per example. BatchTopK instead allocates an activation budget across a batch, allowing different examples to use different numbers of features. JumpReLU learns feature-specific activation thresholds and uses gradient estimators to optimize a discontinuous sparse representation. Matryoshka SAE (Bussmann et al.) trains nested dictionary prefixes to reconstruct independently, encouraging preservation of coarser concepts alongside more specific ones. Its nested design should not be conflated with a similarly named vision SAE or a cascade that re-encodes learned dictionary atoms. TopK · BatchTopK original report · JumpReLU · Matryoshka SAE (Bussmann) · CLIP Matryoshka SAE (Zaigrajew)
Compare variants at measured sparsity and matched data/model settings, not only an equal dictionary width. Report inactive features and how examples are sampled. A feature index from one random seed or encoder is not automatically the same concept as that index from another.
Reconstruction, readability, and causal use are different tests
Section titled “Reconstruction, readability, and causal use are different tests”| Claim | Useful evidence | Insufficient by itself |
|---|---|---|
| The dictionary preserves activations | Held-out reconstruction, sparsity curves, model-output preservation | Attractive top-activating image examples |
| A feature has a coherent meaning | Independent concept labels, examples across activation strengths, precision and coverage | An automatically generated feature name |
| A downstream model relies on that feature | Matched-dose interventions, random controls, reconstruction-error controls, collateral-effect measurement | Correlation with a tissue label or a linear probe alone |
| Features improve WSI prediction | Patient-separated slide/case evaluation with matched encoder and aggregator | Tile classification or TMA results presented as whole-slide evidence |
| Removing a domain cue helps | Held-out site/scanner evaluation plus preservation of task-relevant morphology | Lower domain-classifier accuracy alone |
SAEBench shows why proxy metrics and practical usefulness must be evaluated separately. PATH-DISSECT provides a pathology example that separates represented tissue concepts from the features a frozen linear probe actually uses. Its interventions concern model behavior, not causal effects of tissue changes in patients.
A practical pathology experiment
Section titled “A practical pathology experiment”The following is an editorial study design, not a result demonstrated by every listed paper. Begin with frozen pathology embeddings and a conventional SAE; retain tile coordinates, calibrated microns per pixel, stain, scanner, specimen, and patient identifiers. Fit the dictionary on training patients only, balance sampling across slides, and keep validation/test patients separate. Compare against original embeddings and a simple matched downstream model.
Assess reconstruction, feature utilization, seed stability, expert-recognizable tissue concepts, rare-concept coverage, and controlled changes in model output. Then test a held-out site or scanner. For any domain-feature removal, measure whether useful morphology is also lost. For MIL, distinguish localization from slide-level AUC; improving one does not establish improvement in the other.
For a cross-magnification extension, align physical tissue fields of view, record the encoder input resize separately from WSI extraction dimensions, and test concept correspondence rather than matching latent indices blindly. A dictionary trained at one scale is not automatically an interpretable teacher at another. This is a proposed extension, not evidence that the SAE papers already solve cross-magnification distillation.
Continue with MIL and whole-slide learning, evaluation and clinical translation, and domain shift and robustness.
33 maintained papers · literature reviewed through 2 October 2026
Local surrogates, additive attributions, gradients, activation maps, circuits, and learned feature dictionaries answer different questions. An explanation needs a target, baseline, intervention, scope, stability test, and a decision about what evidence would falsify it.
Prerequisite contract
Controlled perturbation and human studies
Plausible-looking explanations can be unfaithful.
How the field developed
2016–2017
Local surrogate and additive attribution
LIME approximates a decision locally; SHAP assigns additive feature credit under a chosen value function and background distribution.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 01.01 “Why Should I Trust You?”: Explaining the Predictions of Any Classifier LIME explains one prediction by probing a black-box model near the instance, weighting perturbed samples by proximity, and fitting a sparse interpretable surrogate. The surrogate is optimized for local fidelity plus low complexity, not for reproducing the model globally. SP-LIME then selects a small, diverse set of local explanations using a submodular coverage objective to help people inspect a model. Official code 01.02 A Unified Approach to Interpreting Model Predictions SHAP defines a local explanation as an additive model over binary simplified features and proves that local accuracy, missingness, and consistency select one attribution: the Shapley value of a conditional-expectation game. The base value is the expected model output; feature attributions move that baseline to the prediction for the case being explained. The framework unifies several earlier additive explanation methods while making their hidden choices about missing features explicit. Official implementation Work: Apply both methods to one model while varying neighborhood and background distributions.
Exit evidence: Explain which output, coalition semantics, and locality each plot represents.
2017
Gradient paths and spatial localization
Integrated Gradients accumulates sensitivity from a baseline; Grad-CAM aggregates target gradients over convolutional feature maps.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 02.01 Axiomatic Attribution for Deep Networks Integrated Gradients attributes a model’s output difference between an input and a baseline by integrating ordinary input gradients along the straight path joining them. Multiplying each average path gradient by the input displacement yields feature credits that satisfy sensitivity, implementation invariance, and completeness. It needs no network modification and uses repeated gradient evaluations, but the answer remains relative to the selected baseline and output. — 02.02 Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization Grad-CAM localizes evidence for a chosen differentiable target by backpropagating its gradient to a convolutional feature tensor, globally averaging those gradients into one importance weight per channel, taking the weighted channel sum, and applying ReLU. The coarse class-discriminative map works without changing or retraining the network. Guided Grad-CAM multiplies it with guided backpropagation for high-resolution display. Official code Work: Run baseline, layer, label, randomization, and perturbation sanity checks.
Exit evidence: Separate visual sharpness, localization, sensitivity, and faithfulness.
2018
Randomization sanity checks
Model-parameter and label randomization test whether a visually plausible map actually depends on learned weights and the input–label relationship required by its intended explanation task.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 03.01 Sanity Checks for Saliency Maps An explanation intended to describe what a trained model learned should change when the learned parameters or the training input–label relationship is destroyed. Adebayo et al. turn that requirement into two randomization tests. If a saliency method remains invariant to a transformation that removes information required by the intended task, the method can be rejected for that task even when its picture still looks plausible. Official code and errata · Official supplement Work: Compare signed and absolute maps before and after cascading, independent-layer, and label randomization; reject invariances incompatible with the intended use.
Exit evidence: State which task-relevant dependency each explanation test establishes or fails to establish.
2021–2024
Mechanistic circuits and feature dictionaries
Circuit analysis decomposes transformer computation; sparse autoencoders seek a more interpretable feature basis for superposed activations.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 04.01 A Mathematical Framework for Transformer CircuitsAuthor-published technical article; not presented as a peer-reviewed conference paper. A Mathematical Framework for Transformer Circuits rewrites attention-only Transformers as sums of end-to-end computational paths through a shared residual stream. Each attention head separates into a QK circuit that decides where to read and an OV circuit that decides what output-direction change to write. Multiplying read/write matrices exposes virtual weights and head composition without treating intermediate query, key, and value vectors as the final explanatory objects. — 04.02 Sparse Autoencoders Find Highly Interpretable Features in Language Models Cunningham et al. train overcomplete sparse autoencoders on internal Pythia activations to recover feature directions that need not align with individual neurons. A ReLU encoder maps each activation to sparse nonnegative coefficients; tied decoder directions reconstruct the activation. An L1 penalty trades reconstruction fidelity for sparsity, aiming to separate features that were superposed in the model’s native basis. — Work: Trace one behavior through activations, interventions, and output change rather than relying on labels alone.
Exit evidence: Produce an intervention-backed mechanism claim with negative controls.
2018–2026
Concept directions, sparse codes, and causal tests
TCAV quantifies sensitivity along human-defined concept directions; sparse-autoencoder evaluation exposes an amortization gap; certified intervention testing separates a readable feature from a statistically supported causal claim.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 05.01 Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)Published in the ICML 2018 proceedings, PMLR 80. Core mechanism: human-defined concept directions. TCAV learns a linear direction separating activations for user-supplied concept examples from random counterexamples, then measures how moving an input representation along that direction changes a class score. This reading boundary matters: the score quantifies class sensitivity to an operationalized concept at a chosen layer, not causal necessity, pixel attribution, or proof that the model represents the human concept uniquely. Official implementation 05.02 Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional GeneralizationPublished in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337. Core mechanism: linear representation is not linear accessibility. The study distinguishes whether a latent factor is linearly present from whether a fixed linear probe or amortized sparse autoencoder can recover it after unseen factors are composed. Per-sample sparse coding closes failures that persist for learned encoders. This reading boundary matters: the experiments diagnose an accessibility failure in a controlled sparse-superposition setting, not a blanket claim that every linear probe or sparse autoencoder is invalid on real networks. — 05.03 Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic InterpretabilityPublished in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337. Core mechanism: a declared causal estimand. The framework turns activation patching or ablation claims into an expectation over an explicit input-and-intervention distribution, then estimates it with confidence sequences that remain valid under adaptive stopping and sampling. This reading boundary matters: the certificate is statistical for the declared bounded outcome and intervention policy, not a proof that a discovered circuit is uniquely causal or valid under every intervention. — Work: Compare concept directions, linear probes, sparse codes, and explicit interventions on one frozen model; pre-register the concept set, code budget, intervention distribution, behavioral score, stopping rule, and multiple-testing procedure.
Exit evidence: Distinguish human-readable sensitivity, compositional decodability, reconstruction, causal mediation, and anytime-valid statistical evidence; include negative concepts and randomization controls.
2025
Measure transferred and residual student features
UniCAM uses dependence and relevance measurements to separate features associated with distillation from features retained by a non-distilled counterpart.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 06.01 On Explaining Knowledge Distillation: Measuring and Visualising the Knowledge Transfer ProcessPublished in the WACV 2025 proceedings, pages 3467–3476. UniCAM explains what knowledge distillation changes by comparing a KD student with a same-architecture base model trained on raw labels. Partial distance correlation separates student-unique distilled features from base-unique residual features; gradient maps, feature similarity, and label-semantic relevance then characterize transfer layer by layer. Official implementation Work: Apply the diagnostic to matched teacher, distilled student, and non-distilled base models across seeds and layers.
Exit evidence: Report stability, relevance, similarity, compute, and counterfactual interventions before calling a measured feature explanatory.
2025
Pathologist-attention-inspired relation transfer
HVisKD transfers sample- and spatial-region relations between a pathology teacher and student, then compares reconstructed tissue attention with a small pathologist-selected audit set. The visual agreement is interpretive evidence, not proof of causal reasoning.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 07.01 Human Visual Attention-Inspired Knowledge Distillation Underlying Interpretable Computational PathologyPeer-reviewed open-access journal article 15:42124. Low- and high-magnification language motivates the design, but HVisKD transfers sample- and region-relation features between teacher and student given the same pathology patches; it is not cross-magnification distillation. HVisKD transfers relations among pathology features at two levels: sample relations span patches in a batch, and region relations span a spatial pyramid inside each feature map. A large CNN teaches a lightweight CNN on the same input patches; the deployed student predicts patch tissue subtype and those predictions can be assembled into a WSI heatmap. Open-access article · Official code Work: Separate the KD objective, attention-map construction, reader selection, agreement criterion, and external-slide sampling.
Exit evidence: Evaluate explanation stability across seeds, patients, sites, readers, and interventions rather than selected visual examples alone.
2025–2026
Readable prompt ensembles, local concepts and selective adaptation
TPB reweights mistakes across discrete prompt classifiers, CFM exposes spatially grounded sparse concepts, and Robotic Steering selects action-predictive attention heads. Compare what is selected or trained, how frozen features are reused, and whether the final intervention preserves unrelated capabilities.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 08.01 CFM: Language-aligned Concept Foundation Model for VisionECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #4902 checked on 2026-09-12. Publisher version of record was not independently verified. CFM converts dense language-aligned visual features into spatially grounded sparse concepts. A Matryoshka sparse autoencoder reconstructs patch features at nested dictionary capacities, separating coarse concepts from finer residual detail. Data-derived parent-child relationships then improve concept naming and support interpretable classification, captioning and open-vocabulary segmentation. ECCV Spotlight program · Conference manuscript · Official project or implementation 08.02 Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot DemonstrationsECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #13378 checked on 2026-09-12. Publisher version of record was not independently verified. Robotic Steering chooses which VLA attention heads to adapt using a few demonstrations. Head activations are scored by how well cross-trajectory nearest-neighbor retrieval predicts demonstrated actions. Low-rank updates then target selected query and output projections. The central contribution is data-driven adaptation location, not a new distillation loss or an inference-time steering vector. ECCV Spotlight program · Conference manuscript · Official project or implementation Work: Is the adaptation unit a prompt bank, a concept dictionary, an attention head, or another trainable branch?
Exit evidence: Frozen backbones do not imply label-free or zero-cost adaptation. Probe predictability is not causal proof, and Robotic Steering also trains FFN LoRA and its action head.
2025–2026
Counterexamples for spatial, semantic and action grounding
Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 09.01 Why Far Looks Up: Probing Spatial Representation in Vision-Language ModelsECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #10791 checked on 2026-09-12. Publisher version of record was not independently verified. The project page describes an oral; the catalog follows the official Spotlight record. Why Far Looks Up diagnoses vertical-distance entanglement: a VLM may treat higher image position as evidence of greater depth. Controlled behavioral tests are paired with hidden-state contrastive probes, separating a correct answer from the representation that supports it. This is a diagnostic study and benchmark, not a training algorithm that guarantees three-dimensional reasoning. ECCV Spotlight program · Conference manuscript · Official project or implementation Work: Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Exit evidence: A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations.
2025–2026
Recovering, steering and testing visual concepts
Intermediate features can expose information hidden by the final representation. Compare frozen-backbone negation adaptation, language-directed representations, circuit interventions and geometry-subspace probes; distinguish information accessibility from causal control.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 10.01 “World Knowledge” in the Weights: Reading Concept Circuits of Vision TransformersECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #8168. Publisher proceedings/version of record was not verified on 2026-09-12. VisionCLT adapts cross-layer transcoders to visualize how sparse concepts interact inside a ViT. It offers a global circuit assembled from learned encoder/decoder connections and an instance circuit based on active concepts for an image. These are complementary approximations to the model's computation, useful for testing shortcuts and interventions rather than merely highlighting salient pixels. ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript 10.02 Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace InterventionECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7101. Publisher proceedings/version of record was not verified on 2026-09-12. This study asks how geometry is encoded, not simply which backbone has the best depth score. It decomposes a trained linear probe by SVD, projects frozen ViT features into selected subspaces, and reuses the same frozen head. Matched aligned, random and residual projections test which directions carry the geometric information that this particular linear decoder can read. ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript Work: Which layer, token, subspace or circuit changes, and does the intervention preserve unrelated capabilities?
Exit evidence: A probe may measure accessibility rather than semantic understanding; intervention and natural distribution shift are different tests.
2024–2025
Sparse dictionaries: activity budgets, thresholds, and granularity
TopK selects a limited set of active features; JumpReLU learns discontinuous activation thresholds; Matryoshka trains nested dictionary prefixes. These are alternative ways to shape a sparse decomposition, not interchangeable evidence that a learned feature is a unique human concept.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 11.01 Scaling and evaluating sparse autoencoders TopK SAE replaces a magnitude-shrinking sparsity penalty with an explicit activation budget: keep the largest k encoder scores and decode only those latents. Gao et al. combine this older k-sparse principle with scalable training, dead-latent recovery, scaling analysis and feature-quality evaluations. A wider dictionary does not by itself make a model more understandable. Official code · Author manuscript (2024) 11.02 Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders JumpReLU SAE learns a separate positive threshold for every latent: coefficients below it vanish, while coefficients above it retain their full magnitude. Training combines reconstruction error with a direct L0 penalty, and uses a carefully derived straight-through estimator to learn discontinuous thresholds. The gain is improved sparsity–fidelity behavior, not a guarantee of one biological concept per latent. Author PDF with implementation pseudocode 11.03 Learning Multi-Level Features with Matryoshka Sparse Autoencoders Matryoshka SAE trains one dictionary at several nested prefix widths: each prefix must reconstruct the same activation without relying on later features. Early features therefore face pressure to retain broad concepts while later features can specialize. This reduces feature absorption in the tested settings, with a modest reconstruction and training-cost trade-off. Author code · Toy-model code · Author manuscript Work: What changes when the active-feature budget, dictionary width, or reconstruction objective changes?
Exit evidence: These foundational experiments concern language-model activations. Match actual sparsity and model/data settings; natural-image or pathology transfer requires new evidence.
2025
Evaluate fidelity, concept isolation, and useful interventions separately
SAE stitching and meta-SAEs test whether dictionaries are complete or atomic; SAEBench compares reconstruction, interpretation, and practical task measurements. A useful dictionary need not be canonical, and a method can improve one proxy while losing on another objective.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 12.01 Sparse Autoencoders Do Not Find Canonical Units of Analysis A useful SAE dictionary is not necessarily a unique, complete set of irreducible concepts. Leask et al. test two failure modes: stitching exposes information missing from smaller dictionaries, and a meta-SAE decomposes large-dictionary directions into sparse combinations. BatchTopK makes the number of components flexible per example while fixing the batch-average budget. Author manuscript · BatchTopK component implementation · Author meta-SAE dashboard 12.02 SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability SAEBench evaluates sparse autoencoders along several distinct axes instead of equating low reconstruction error with interpretability. Its eight evaluations cover concept detection, readable feature descriptions, preservation of model behavior and separation of concepts under interventions. The central lesson is to choose an SAE using the task-relevant profile of evidence, not a single aggregate leaderboard score. Official evaluation code · Author manuscript · Interactive evaluations Work: Does a better sparsity–reconstruction curve also improve the intended downstream use?
Exit evidence: The benchmark is for language-model SAE evaluation. Its measurement principles transfer; its numerical results are not a pathology benchmark.
2025
Locate visual concepts and align representations across models
PatchSAE examines spatial visual features and how adaptation changes their use. Universal SAE learns cross-model sparse correspondences from representations of shared images. The first audits adaptation; the second aligns feature spaces, rather than training a distilled student.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 13.01 Sparse autoencoders reveal selective remapping of visual concepts during adaptation PatchSAE turns intermediate CLIP image-token activations into a sparse concept dictionary with spatial locations. It then uses the same dictionary to compare CLIP with prompt-adapted MaPLe: much of the observed adaptation benefit comes from remapping already recognized concepts to task labels, rather than discovering an entirely new visual vocabulary. Proceedings PDF · Code and demo 13.02 Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment Universal Sparse Autoencoders learn a shared sparse concept space across multiple frozen vision models. Each model keeps its own encoder and decoder, but a code obtained from any one model must reconstruct the activations of all the others on the same input. This aligns concept coordinates during learning instead of matching independently discovered dictionaries afterward. Proceedings PDF · Code · Project Work: Which concepts persist across image locations, adaptation settings, and encoders?
Exit evidence: Readable examples and cross-decoding do not prove a complete or unique concept basis. Preserve the source-specific activation unit and the actual downstream intervention.
2025–2026
Test semantic coherence, concept supervision, and hierarchical steering
Hierarchical CLIP SAEs and VLM SAE studies connect feature interpretation with retrieval, bias analysis, or interventions on generated descriptions. Concept Bottleneck SAE adds supplied concept structure; SAE++ builds a hierarchy over dictionary atoms. Human alignment, control strength, and preserving unrelated content are separate questions.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 14.01 Interpreting CLIP with Hierarchical Sparse AutoencodersZaigrajew et al. multi-TopK CLIP method, distinct from Bussmann et al. fixed-prefix Matryoshka SAE. Author repository discloses a post-acceptance text-standardization correction; review distinguishes proceedings measurements from revised code. Zaigrajew and colleagues train a CLIP sparse autoencoder at several activation-ranked TopK budgets simultaneously. Every budget uses the same encoder and decoder, but retains a different number of the current input’s strongest activations. This multi-budget objective seeks useful reconstructions from coarse through richer sparse descriptions; it is not the fixed-prefix Matryoshka SAE of Bussmann and colleagues. Proceedings PDF · Author code and correction notice · Original ICML implementation · arXiv 14.02 Sparse Autoencoders Learn Monosemantic Features in Vision-Language ModelsNeurIPS 2025 proceedings volume 38; the repository BibTeX uses publication year 2026. Catalog year follows the conference edition. This work evaluates whether sparse vision features are semantically coherent and whether manipulating them changes a VLM output. Its MonoSemanticity score measures activation-weighted image similarity and is checked against human pairwise judgments; separate CLIP-feature interventions steer LLaVA. Interpretability and controllability are related experiments, not interchangeable definitions. Proceedings PDF · Code and human-comparison benchmark · arXiv 14.03 Interpretable and Steerable Concept Bottleneck Sparse Autoencoders CB-SAE combines an unsupervised sparse dictionary with a user-specified concept bottleneck. It scores existing neurons for interpretability and downstream steerability, prunes low-utility neurons, then trains a parallel concept autoencoder while freezing retained SAE weights. This recovers useful reconstruction and supplies concepts missing from the retained dictionary. Proceedings PDF · Supplement · Author repository (code not yet released) · arXiv 14.04 SAE++: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMsNeurIPS 2026 Accept (poster) verified from official program event 151133, OpenReview cl4vXMku1D, matching all five authors. Program title omits SAE++ and joins LearnMulti-Level; review uses arXiv v3 revised 7 October 2026. Proceedings version not yet verified. SAE++ learns a hierarchy by training a second SAE on the decoder atoms of a first SAE. Level 1 decomposes model activations into sparse features; Level 2 groups the learned feature directions into higher-level concepts. It is not a residual-error cascade and does not feed Level-1 sparse sample codes into a conventional stacked bottleneck. Reviewed manuscript v3 · Official NeurIPS poster · OpenReview · Code Work: Can the target concept change while unrelated visual content and answer quality remain intact?
Exit evidence: A semantic-coherence score is not a causal or clinical certificate. Report coverage, duplicate features, judge dependence, and both intended and collateral intervention effects.
2024–2026
Pathology concepts: distinguish representation from classifier reliance
Early pathology SAE work relates features to biological annotations and encoder depth. PATH-DISSECT adds depth-resolved selectivity checks and controlled interventions inside the frozen encoder, measured through a frozen linear probe. A represented tissue concept need not be used by every downstream head.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 15.01 Learning biologically relevant features in a pathology foundation model using sparse autoencodersReview uses arXiv:2407.10785v3, including its supplement. A later NeurIPS 2025 Imageomics workshop paper has the title Evaluating the Utility of Sparse Autoencoders for Interpreting a Pathology Foundation Model; its inaccessible full PDF was not used to replace this version's claims or publication status. A sparse dictionary trained on frozen PLUTO embeddings exposes cell- and tissue-associated directions that are easier to inspect than individual dense coordinates. The study links selected SAE activations to estimated cell counts, compares layers and stains, and checks transfer to different cancer cohorts. Its main contribution is an empirical interpretability study, not a new foundation encoder or proof that every sparse feature is one biological concept. Versioned full manuscript · Authors' later Imageomics workshop publication 15.02 PATH-DISSECT: Interpreting Concept Representation and Classifier Reliance in Pathology Foundation Models with Sparse AutoencodersOfficial MICCAI 2026 Satellite Events page supplies the 11-page camera-ready; Springer DOI and final pages were pending at inspection. This is MI4MedFM workshop work on tissue patches, not main-conference or patient-level WSI outcome validation. Causal claims concern interventions on a frozen encoder plus linear readout, not biological causation. PATH-DISSECT separates a concept being encoded from a classifier relying on it. It trains sparse autoencoders at multiple blocks of frozen pathology transformers, validates class-associated features, then edits those features inside the forward pass while measuring a frozen linear probe. Its intervention is on model activations, not a causal experiment on tissue biology. Official camera-ready PDF · Author implementation · Workshop OpenReview record Work: Is a tissue concept encoded, and does the specified classifier actually depend on it?
Exit evidence: Tile-level annotations and linear-probe interventions are not WSI outcome validation or causal tissue biology. Preserve patch sampling, patient/site splits, feature-selection scope, and error-preserving controls.
2025–2026
Sparse features for slide prototypes and weakly supervised localization
ProtoMIL uses SAE-derived patch prototypes in slide learning; SAE-1N selects a discriminative latent using slide supervision. These methods connect sparse features to WSI tasks, but localization quality and slide-level predictive performance can move in different directions.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 16.01 Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image ClassificationOfficial MICCAI 2025 record and 11-page accepted manuscript checked. PANDA is evaluated as binary tumor classification in this study, not the original six-class ISUP grading task. Artifact concepts are masked before MIL retraining; this is not a universal AUC improvement. ProtoMIL discovers sparse visual concepts from frozen CONCH patch embeddings, then expresses a WSI class logit as a sum of weighted concept contributions. Pathologists can inspect maximally activating patches and disable unwanted concepts before retraining the slide classifier. The point is an inspectable decision decomposition, not a claim that attention alone explains a prediction. Official MICCAI open-access record · Official accepted manuscript · Author implementation · arXiv manuscript 16.02 Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddingsPSB 2026 volume 31, pages 708–721, with source supplement checked. Neuron discovery uses slide labels, but the reported localization thresholds use patch-level validation annotations. Single-neuron localization does not preserve all multi-feature slide-classification performance. SAE-1N selects one sparse feature using only slide labels, then uses its patch activation as a tumor-localization signal. A BatchTopK SAE decomposes frozen image embeddings; each feature is max-pooled over a slide and ranked by its tumor-versus-normal association. The selected coordinate is not normalized separately within each slide, unlike softmax attention. The reported thresholded localization still uses annotated validation patches. Official PSB proceedings PDF · Author implementation and supplement Work: Which labels select the feature, tune its threshold, and evaluate localization or slide prediction?
Exit evidence: Do not hide patch-annotation use in threshold selection or assume improved localization implies better slide AUC. Distinguish retraining after an intervention from a fixed-classifier causal test.
2026
Identify domain features without erasing task-relevant morphology
EXPOSE trains an SAE on pathology embeddings, identifies domain-associated latent coordinates, and masks selected coordinates before downstream prediction. It studies whether this intervention reduces acquisition-domain dependence while retaining useful outcome information.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 17.01 EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse AutoencodersIndependently verified on the official eXCV 2026 Proceedings Track list, not ECCV main conference. Scientific review uses arXiv:2608.28191v1; a final typeset proceedings version was not verified. Evaluation is prostate tissue-microarray spots, not a whole-slide benchmark, and includes a thin-section failure case. EXPOSE learns a sparse basis of frozen pathology features, identifies dimensions that distinguish two acquisition domains, and masks selected dimensions before training a relapse classifier. It is an automatic domain-feature selection strategy, not a new image encoder. The evidence concerns prostate tissue-microarray spots and controlled preparation/scanner variants, not unrestricted WSI deployment. Official workshop accepted-paper list · Author implementation · Versioned full manuscript and appendix Work: Does domain-feature suppression preserve prediction in every held-out acquisition condition?
Exit evidence: The inspected study uses prostate tissue-microarray spots, not a general full-WSI benchmark. Results vary by domain; removing detectable acquisition information does not prove that only nuisance was removed.
Research branches
Post-hoc feature attribution
Contribution to which output under which perturbation or coalition distribution?
“Why Should I Trust You?”: Explaining the Predictions of Any Classifier · A Unified Approach to Interpreting Model Predictions
Gradient and spatial evidence
Does a heatmap locate necessary evidence or merely correlate with it?
Axiomatic Attribution for Deep Networks · Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization
Explanation validation
Does the explanation change when the learned parameters or input–label relationship required by the claimed task is destroyed?
Mechanistic interpretation
Which internal components causally mediate the behavior?
A Mathematical Framework for Transformer Circuits · Sparse Autoencoders Find Highly Interpretable Features in Language Models
Concept directions, sparse codes, and causal interventions
Is the claimed concept merely readable, compositionally usable, or causally responsible for a behavior under a declared intervention?
Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) · Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalization · Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability
KD mechanism diagnostics
Which student features changed because of distillation, and are they task-relevant or merely correlated?
On Explaining Knowledge Distillation: Measuring and Visualising the Knowledge Transfer Process
Interpretable pathology distillation
Do transferred relations produce stable, clinically meaningful attention or merely plausible heatmaps?
Readable prompt ensembles, local concepts and selective adaptation
Is the adaptation unit a prompt bank, a concept dictionary, an attention head, or another trainable branch?
CFM: Language-aligned Concept Foundation Model for Vision · Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrations
Counterexamples for spatial, semantic and action grounding
Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Recovering, steering and testing visual concepts
Which layer, token, subspace or circuit changes, and does the intervention preserve unrelated capabilities?
“World Knowledge” in the Weights: Reading Concept Circuits of Vision Transformers · Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention
Sparse dictionaries: activity budgets, thresholds, and granularity
What changes when the active-feature budget, dictionary width, or reconstruction objective changes?
Scaling and evaluating sparse autoencoders · Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders · Learning Multi-Level Features with Matryoshka Sparse Autoencoders
Evaluate fidelity, concept isolation, and useful interventions separately
Does a better sparsity–reconstruction curve also improve the intended downstream use?
Sparse Autoencoders Do Not Find Canonical Units of Analysis · SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
Locate visual concepts and align representations across models
Which concepts persist across image locations, adaptation settings, and encoders?
Sparse autoencoders reveal selective remapping of visual concepts during adaptation · Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment
Test semantic coherence, concept supervision, and hierarchical steering
Can the target concept change while unrelated visual content and answer quality remain intact?
Interpreting CLIP with Hierarchical Sparse Autoencoders · Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models · Interpretable and Steerable Concept Bottleneck Sparse Autoencoders · SAE++: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs
Pathology concepts: distinguish representation from classifier reliance
Is a tissue concept encoded, and does the specified classifier actually depend on it?
Learning biologically relevant features in a pathology foundation model using sparse autoencoders · PATH-DISSECT: Interpreting Concept Representation and Classifier Reliance in Pathology Foundation Models with Sparse Autoencoders
Sparse features for slide prototypes and weakly supervised localization
Which labels select the feature, tune its threshold, and evaluate localization or slide prediction?
Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image Classification · Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddings
Identify domain features without erasing task-relevant morphology
Does domain-feature suppression preserve prediction in every held-out acquisition condition?
Comparison table
| Family or stage | Scientific distinction | Main risk | Anchors |
|---|---|---|---|
| Surrogate / additive | Approximates or decomposes a prediction in an interpretable input basis. | Neighborhood and background choices can reverse the story. | 2016 · 2017 |
| Gradient-based | Measures local sensitivity or its path integral. | Saturation, baseline, layer, and preprocessing shape the map. | 2017 · 2017 |
| Randomization test | Destroys learned weights or label structure and measures whether an explanation retains an incompatible invariance. | Passing is necessary for some faithfulness tasks but does not prove completeness, causal mediation, localization accuracy, or human usefulness. | 2018 |
| Internal mechanism | Decomposes activations and tests components with interventions. | Readable features need not be complete, unique, or causal. | 2021 · 2024 |
| Concept directions, sparse codes, and causal tests | TCAV measures directional concept sensitivity; compositional coding audits expose probe/SAE amortization limits; certified intervention tests attach sequential uncertainty to a model-level causal estimand. | Concept-set bias, basis choice, probe capacity, intervention artifacts, adaptive stopping, or a readable label can be mistaken for a faithful mechanism. | 2018 · 2026 · 2026 |
| Measure transferred and residual student features | Adds an explicit post-hoc measurement framework to a literature dominated by new transfer losses. | Layer correspondence and dependence scores do not establish causal mediation or completeness. | 2025 |
| Pathologist-attention-inspired relation transfer | Adds a pathology-specific attention-consistency audit to general post-hoc KD feature diagnostics. | Tiny selected reader audits, patch-level splits, and heatmap plausibility can be mistaken for faithful explanation. | 2025 |
| Readable prompt ensembles, local concepts and selective adaptation | TPB reweights mistakes across discrete prompt classifiers, CFM exposes spatially grounded sparse concepts, and Robotic Steering selects action-predictive attention heads. Compare what is selected or trained, how frozen features are reused, and whether the final intervention preserves unrelated capabilities. | Frozen backbones do not imply label-free or zero-cost adaptation. Probe predictability is not causal proof, and Robotic Steering also trains FFN LoRA and its action head. | 2026 · 2026 |
| Counterexamples for spatial, semantic and action grounding | Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable. | A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations. | 2026 |
| Recovering, steering and testing visual concepts | Intermediate features can expose information hidden by the final representation. Compare frozen-backbone negation adaptation, language-directed representations, circuit interventions and geometry-subspace probes; distinguish information accessibility from causal control. | A probe may measure accessibility rather than semantic understanding; intervention and natural distribution shift are different tests. | 2026 · 2026 |
| Sparse dictionaries: activity budgets, thresholds, and granularity | TopK selects a limited set of active features; JumpReLU learns discontinuous activation thresholds; Matryoshka trains nested dictionary prefixes. These are alternative ways to shape a sparse decomposition, not interchangeable evidence that a learned feature is a unique human concept. | These foundational experiments concern language-model activations. Match actual sparsity and model/data settings; natural-image or pathology transfer requires new evidence. | 2025 · 2024 · 2025 |
| Evaluate fidelity, concept isolation, and useful interventions separately | SAE stitching and meta-SAEs test whether dictionaries are complete or atomic; SAEBench compares reconstruction, interpretation, and practical task measurements. A useful dictionary need not be canonical, and a method can improve one proxy while losing on another objective. | The benchmark is for language-model SAE evaluation. Its measurement principles transfer; its numerical results are not a pathology benchmark. | 2025 · 2025 |
| Locate visual concepts and align representations across models | PatchSAE examines spatial visual features and how adaptation changes their use. Universal SAE learns cross-model sparse correspondences from representations of shared images. The first audits adaptation; the second aligns feature spaces, rather than training a distilled student. | Readable examples and cross-decoding do not prove a complete or unique concept basis. Preserve the source-specific activation unit and the actual downstream intervention. | 2025 · 2025 |
| Test semantic coherence, concept supervision, and hierarchical steering | Hierarchical CLIP SAEs and VLM SAE studies connect feature interpretation with retrieval, bias analysis, or interventions on generated descriptions. Concept Bottleneck SAE adds supplied concept structure; SAE++ builds a hierarchy over dictionary atoms. Human alignment, control strength, and preserving unrelated content are separate questions. | A semantic-coherence score is not a causal or clinical certificate. Report coverage, duplicate features, judge dependence, and both intended and collateral intervention effects. | 2025 · 2025 · 2026 · 2026 |
| Pathology concepts: distinguish representation from classifier reliance | Early pathology SAE work relates features to biological annotations and encoder depth. PATH-DISSECT adds depth-resolved selectivity checks and controlled interventions inside the frozen encoder, measured through a frozen linear probe. A represented tissue concept need not be used by every downstream head. | Tile-level annotations and linear-probe interventions are not WSI outcome validation or causal tissue biology. Preserve patch sampling, patient/site splits, feature-selection scope, and error-preserving controls. | 2024 · 2026 |
| Sparse features for slide prototypes and weakly supervised localization | ProtoMIL uses SAE-derived patch prototypes in slide learning; SAE-1N selects a discriminative latent using slide supervision. These methods connect sparse features to WSI tasks, but localization quality and slide-level predictive performance can move in different directions. | Do not hide patch-annotation use in threshold selection or assume improved localization implies better slide AUC. Distinguish retraining after an intervention from a fixed-classifier causal test. | 2025 · 2026 |
| Identify domain features without erasing task-relevant morphology | EXPOSE trains an SAE on pathology embeddings, identifies domain-associated latent coordinates, and masks selected coordinates before downstream prediction. It studies whether this intervention reduces acquisition-domain dependence while retaining useful outcome information. | The inspected study uses prostate tissue-microarray spots, not a general full-WSI benchmark. Results vary by domain; removing detectable acquisition information does not prove that only nuisance was removed. | 2026 |
Concept lineage
“Why Should I Trust You?”: Explaining the Predictions of Any ClassifierA Unified Approach to Interpreting Model Predictions
Local approximation is connected to a constrained additive attribution family with explicit coalition semantics.
Axiomatic Attribution for Deep NetworksSanity Checks for Saliency Maps
An axiomatic attribution construction is subjected to parameter and label randomization, revealing that signed rankings can change even when absolute-value pictures still look input-like.
Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based LocalizationSanity Checks for Saliency Maps
Grad-CAM moves from an attractive localization method to one explicitly tested for dependence on learned parameters and training labels.
Axiomatic Attribution for Deep NetworksA Mathematical Framework for Transformer Circuits
Input sensitivity motivates deeper component-level hypotheses that require causal intervention inside the network.
Failure modes to audit
- Calling attention or a heatmap an explanation without a target claim.
- Choosing baselines after seeing the desired result.
- Evaluating plausibility but not faithfulness.
- Ignoring instability across seeds, layers, and equivalent inputs.
Questions for every paper
- What exact output is explained?
- Which intervention defines importance?
- Is the explanation stable and complete?
- Would the explanation support a real decision?
Complete maintained paper route
Section titled “Complete maintained paper route”33 canonical readings · 2016–2026· 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 33 readings in maintained order
“Why Should I Trust You?”: Explaining the Predictions of Any Classifier
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin
Official artifacts: Official code
Metadata checked
A Unified Approach to Interpreting Model Predictions
Scott M. Lundberg and Su-In Lee
Official artifacts: Official implementation
Metadata checked
Axiomatic Attribution for Deep Networks
Mukund Sundararajan, Ankur Taly, and Qiqi Yan
Metadata checked
Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization
Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra
Official artifacts: Official code
Metadata checked
Sanity Checks for Saliency Maps
Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim
Official artifacts: Official code and errata · Official supplement
Metadata checked
Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)
Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viégas, and Rory Sayres
Published in the ICML 2018 proceedings, PMLR 80.
Official artifacts: Official implementation
Metadata checked
A Mathematical Framework for Transformer Circuits
Nelson Elhage et al.
Author-published technical article; not presented as a peer-reviewed conference paper.
Metadata checked
Learning biologically relevant features in a pathology foundation model using sparse autoencoders
Nhat Minh Le, Ciyue Shen, Neel Patel, Chintan Shah, Darpan Sanghavi, Blake Martin, Alfred Eng, Daniel Shenker, Harshith Padigela, Raymond Biju, Syed Ashar Javed, Jennifer Hipp, John Abel, Harsha Pokkalla, Sean Grullon, and Dinkar Juyal
Review uses arXiv:2407.10785v3, including its supplement. A later NeurIPS 2025 Imageomics workshop paper has the title Evaluating the Utility of Sparse Autoencoders for Interpreting a Pathology Foundation Model; its inaccessible full PDF was not used to replace this version's claims or publication status.
Official artifacts: Versioned full manuscript · Authors' later Imageomics workshop publication
Metadata checked
Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders
Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda
Official artifacts: Author PDF with implementation pseudocode
Metadata checked
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Hoagy Cunningham, Aidan Ewart, Logan Riggs Smith, Robert Huben, and Lee Sharkey
Metadata checked
Human Visual Attention-Inspired Knowledge Distillation Underlying Interpretable Computational Pathology
Muzhou Yu, Zihan Zhong, Xingang Zhou, Yuekun Wang, Tingyu Liang, Jiamin Chen, Hongmin Huang, Junhan Zhou, Dachun Zhao, Bo Lei, Yu Wang, Wenbin Ma, and Kaisheng Ma
Peer-reviewed open-access journal article 15:42124. Low- and high-magnification language motivates the design, but HVisKD transfers sample- and region-relation features between teacher and student given the same pathology patches; it is not cross-magnification distillation.
Official artifacts: Open-access article · Official code
Metadata checked
On Explaining Knowledge Distillation: Measuring and Visualising the Knowledge Transfer Process
Gereziher Adhane, Mohammad Mahdi Dehshibi, Dennis Vetter, David Masip, and Gemma Roig
Published in the WACV 2025 proceedings, pages 3467–3476.
Official artifacts: Official implementation
Metadata checked
Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image Classification
Susu Sun, Dominique van Midden, Geert Litjens, and Christian F. Baumgartner
Official MICCAI 2025 record and 11-page accepted manuscript checked. PANDA is evaluated as binary tumor classification in this study, not the original six-class ISUP grading task. Artifact concepts are masked before MIL retraining; this is not a universal AUC improvement.
Official artifacts: Official MICCAI open-access record · Official accepted manuscript · Author implementation · arXiv manuscript
Metadata checked
Sparse Autoencoders Do Not Find Canonical Units of Analysis
Patrick Leask, Bart Bussmann, Michael Pearce, Joseph Bloom, Curt Tigges, Noura Al Moubayed, Lee Sharkey, and Neel Nanda
Official artifacts: Author manuscript · BatchTopK component implementation · Author meta-SAE dashboard
Metadata checked
Scaling and evaluating sparse autoencoders
Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu
Official artifacts: Official code · Author manuscript (2024)
Metadata checked
Learning Multi-Level Features with Matryoshka Sparse Autoencoders
Bart Bussmann, Noa Nabeshima, Adam Karvonen, and Neel Nanda
Official artifacts: Author code · Toy-model code · Author manuscript
Metadata checked
SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
Adam Karvonen, Can Rager, Johnny Lin, Curt Tigges, Joseph Isaac Bloom, David Chanin, Yeu-Tong Lau, Eoin Farrell, Callum Stuart Mcdougall, Kola Ayonrinde, Demian Till, Matthew Wearden, Arthur Conmy, Samuel Marks, and Neel Nanda
Official artifacts: Official evaluation code · Author manuscript · Interactive evaluations
Metadata checked
Interpreting CLIP with Hierarchical Sparse Autoencoders
Vladimir Zaigrajew, Hubert Baniecki, and Przemyslaw Biecek
Zaigrajew et al. multi-TopK CLIP method, distinct from Bussmann et al. fixed-prefix Matryoshka SAE. Author repository discloses a post-acceptance text-standardization correction; review distinguishes proceedings measurements from revised code.
Official artifacts: Proceedings PDF · Author code and correction notice · Original ICML implementation · arXiv
Metadata checked
Sparse autoencoders reveal selective remapping of visual concepts during adaptation
Hyesu Lim, Jinho Choi, Jaegul Choo, and Steffen Schneider
Official artifacts: Proceedings PDF · Code and demo
Metadata checked
Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment
Harrish Thasarathan, Julian Forsyth, Thomas Fel, Matthew Kowal, and Konstantinos G. Derpanis
Official artifacts: Proceedings PDF · Code · Project
Metadata checked
Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
Mateusz Pach, Shyamgopal Karthik, Quentin Bouniot, Serge Belongie, and Zeynep Akata
NeurIPS 2025 proceedings volume 38; the repository BibTeX uses publication year 2026. Catalog year follows the conference edition.
Official artifacts: Proceedings PDF · Code and human-comparison benchmark · arXiv
Metadata checked
Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalization
Vitória Barin-Pacela, Shruti Joshi, Isabela Camacho, Simon Lacoste-Julien, and David Klindt
Published in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337.
Metadata checked
Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability
Amir Asiaee
Published in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337.
Metadata checked
Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention
Weichen Zhou, Yawen Zou, Chunzhi Gu, Ran Dong, Haoran Xie, and Chao Zhang
ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7101. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript
Metadata checked
“World Knowledge” in the Weights: Reading Concept Circuits of Vision Transformers
Yanlin Chen, Tang Li, and Xi Peng
ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #8168. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript
Metadata checked
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, Yu Su, Jonathan Tremblay, Chan Hee Song, and Jaesik Park
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #10791 checked on 2026-09-12. Publisher version of record was not independently verified. The project page describes an oral; the catalog follows the official Spotlight record.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked
Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrations
Chancharik Mitra, Yusen Luo, Raj Saravanan, Dantong Niu, Anirudh Pai, Jesse Thomason, Trevor Darrell, Abrar Anwar, Deva Ramanan, and Roei Herzig
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #13378 checked on 2026-09-12. Publisher version of record was not independently verified.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked
CFM: Language-aligned Concept Foundation Model for Vision
Kai Wittenmayer, Sukrut Rao, Amin Parchami-Araghi, Bernt Schiele, and Jonas Fischer
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #4902 checked on 2026-09-12. Publisher version of record was not independently verified.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked
Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddings
Keita Tamura, Yao-zhong Zhang, Yohei Okubo, and Seiya Imoto
PSB 2026 volume 31, pages 708–721, with source supplement checked. Neuron discovery uses slide labels, but the reported localization thresholds use patch-level validation annotations. Single-neuron localization does not preserve all multi-feature slide-classification performance.
Official artifacts: Official PSB proceedings PDF · Author implementation and supplement
Metadata checked
EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders
Anja Witte, Maximilian Lennartz, Jan Baumbach, Guido Sauter, Stefan Bonn, Patrick Fuhlert, and Marina Zimmermann
Independently verified on the official eXCV 2026 Proceedings Track list, not ECCV main conference. Scientific review uses arXiv:2608.28191v1; a final typeset proceedings version was not verified. Evaluation is prostate tissue-microarray spots, not a whole-slide benchmark, and includes a thin-section failure case.
Official artifacts: Official workshop accepted-paper list · Author implementation · Versioned full manuscript and appendix
Metadata checked
PATH-DISSECT: Interpreting Concept Representation and Classifier Reliance in Pathology Foundation Models with Sparse Autoencoders
Mayur Parvatikar, Xingjian Li, and Min Xu
Official MICCAI 2026 Satellite Events page supplies the 11-page camera-ready; Springer DOI and final pages were pending at inspection. This is MI4MedFM workshop work on tissue patches, not main-conference or patient-level WSI outcome validation. Causal claims concern interventions on a frozen encoder plus linear readout, not biological causation.
Official artifacts: Official camera-ready PDF · Author implementation · Workshop OpenReview record
Metadata checked
Interpretable and Steerable Concept Bottleneck Sparse Autoencoders
Akshay Kulkarni, Tsui-Wei Weng, Vivek Narayanaswamy, Shusen Liu, Wesam A. Sakla, and Kowshik Thopalli
Official artifacts: Proceedings PDF · Supplement · Author repository (code not yet released) · arXiv
Metadata checked
SAE++: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs
Yusong Zhao, Hengyi Wang, Tanuja Ganu, Akshay Nambi, and Hao Wang
NeurIPS 2026 Accept (poster) verified from official program event 151133, OpenReview cl4vXMku1D, matching all five authors. Program title omits SAE++ and joins LearnMulti-Level; review uses arXiv v3 revised 7 October 2026. Proceedings version not yet verified.
Official artifacts: Reviewed manuscript v3 · Official NeurIPS poster · OpenReview · Code
Metadata checked