Capability $\neq$ Interpretability: Human Interpretability of Vision Foundation Models

• URL: http://arxiv.org/abs/2605.20337v1
• Date: Tue, 19 May 2026 18:00:06 GMT
• Title: Capability $\neq$ Interpretability: Human Interpretability of Vision Foundation Models
• Abstract summary: We build a framework for measuring and comparing the human interpretability of vision models.<n>A chance-anchored scoring function places every model on a common scale.<n>Foundation models are consistently *less* interpretable than their supervised counterparts.
• Score: 24.250649573452666
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: How interpretable are the features of leading vision models? The question is increasingly pressing as these models move from research benchmarks into high-stakes deployments, yet existing methods cannot answer it reliably. We close this gap with a framework for measuring and comparing the human interpretability of vision models, built around two complementary psychophysics protocols: (1) localizability -- can an observer predict where a feature fires on a novel image? -- and (2) nameability -- can an observer accurately describe what the feature represents? Features are recovered via sparse autoencoders, and a chance-anchored scoring function places every model on a common scale. Applying the framework to six vision transformers -- two supervised ViTs and four foundation models (DINOv2, DINOv3, CLIP, SigLIP) -- we collected more than $15{,}000$ behavioral responses, analyzing the $13{,}400$ responses from the $377$ participants who passed our pre-specified quality checks. Foundation models are consistently *less* interpretable than their supervised counterparts, and the gap is not a capability tradeoff: interpretability does not correlate with downstream task performance on any benchmark we examine. What does correlate is the locality of a feature's activations and coarse-grained semantic alignment with humans -- models with focal activations and representations that reflect the world's broad categorical structure produce more interpretable features, whereas fine-grained perceptual alignment does not. The two protocols yield strongly correlated rankings and share the same predictors, establishing interpretability as an independent, measurable dimension of representation quality -- and, surprisingly, one on which every foundation model we tested falls below the supervised baselines that came before. Capability alone cannot close that gap; locality and coarse-grained alignment can.

Related papers

• MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents [74.83535434786145]
  MNIST-PRO is a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints.<n>We evaluate ten multimodal models across four memory representations, including raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas.
  arXiv  Detail & Related papers  (2026-08-31T16:05:39Z)
• Paying More Attention to Visual Tokens in Self-Evolving Large Multimodal Models [54.097381112138]
  Current self-evolving LMMs struggle on vision-language understanding tasks such as image captioning and visual question answering.<n>We propose VISE, a purely unsupervised self-evolving framework that directly regularizes the model's visual conditioning policy.<n>VISE operates within a single model without specialist roles, external reward models, or annotations.
  arXiv  Detail & Related papers  (2026-06-25T17:59:55Z)
• S$^2$COPE: Self-Supervised Concept Discovery via Preference Learning [15.951097715837799]
  We introduce Self-Supervised Concept discOvery via Preference lEarning (model), a label-free framework that resolves this dilemma.<n>Our work suggest that interpretability can emerge through a model's autonomous interaction with incidental visual structures, without any human supervision.
  arXiv  Detail & Related papers  (2026-06-12T16:02:44Z)
• Epistemic Observability in Language Models [0.0]
  We find that models report highest confidence precisely when they are fabricating.<n>We prove, under formal assumptions, that this is not a capability gap but an observational one.<n>We construct a tensor interface that escapes the impossibility by exporting computational byproducts.
  arXiv  Detail & Related papers  (2026-03-20T21:59:34Z)
• Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models [50.99097734404912]
  We show that RL-friendly models exhibit intra-class compactness and inter-class separation in their probability assignments to correct vs. incorrect responses.<n>Experiments across six mathematical benchmarks show consistent improvements across all model families, with gains up to 5.9 points on AIME24.
  arXiv  Detail & Related papers  (2026-01-11T13:34:44Z)
• A4-Agent: An Agentic Framework for Zero-Shot Affordance Reasoning [29.682282730123234]
  Affordance prediction, which identifies interaction regions on objects based on language instructions, is critical for embodied AI.<n>We propose A4-Agent, a training-free agentic framework that decouples affordance prediction into a three-stage pipeline.<n>Our framework significantly outperforms state-of-the-art supervised methods across multiple benchmarks.
  arXiv  Detail & Related papers  (2025-12-16T14:27:47Z)
• Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent [58.90049897180927]
  We introduce an automated framework for detecting unintended reliance on visual features in vision models.<n>A self-reflective agent generates and tests hypotheses about visual attributes that a model may rely on.<n>We evaluate our approach on a novel benchmark of 130 models designed to exhibit diverse visual attribute dependencies.
  arXiv  Detail & Related papers  (2025-10-24T17:59:02Z)
• When Does Pruning Benefit Vision Representations? [6.306016476757605]
  Pruning is widely used to reduce the complexity of deep learning models, but its effects on interpretability and representation learning remain poorly understood.<n>This paper investigates how pruning influences vision models across three key dimensions: (i) interpretability, (ii) unsupervised object discovery, and (iii) alignment with human perception.
  arXiv  Detail & Related papers  (2025-07-02T13:57:49Z)
• Test-Time Consistency in Vision Language Models [26.475993408532304]
  Vision-Language Models (VLMs) have achieved impressive performance across a wide range of multimodal tasks.<n>Recent benchmarks, such as MM-R3, highlight that even state-of-the-art VLMs can produce divergent predictions across semantically equivalent inputs.<n>We propose a simple and effective test-time consistency framework that enhances semantic consistency without supervised re-training.
  arXiv  Detail & Related papers  (2025-06-27T17:09:44Z)
• Unsupervised Model Diagnosis [49.36194740479798]
  This paper proposes Unsupervised Model Diagnosis (UMO) to produce semantic counterfactual explanations without any user guidance. Our approach identifies and visualizes changes in semantics, and then matches these changes to attributes from wide-ranging text sources.
  arXiv  Detail & Related papers  (2024-10-08T17:59:03Z)
• Intrinsic User-Centric Interpretability through Global Mixture of Experts [31.738009841932374]
  InterpretCC is a family of intrinsically interpretable neural networks that optimize for ease of human understanding and explanation faithfulness.<n>We show that InterpretCC explanations are found to have higher actionability and usefulness over other intrinsically interpretable approaches.
  arXiv  Detail & Related papers  (2024-02-05T11:55:50Z)
• DualFair: Fair Representation Learning at Both Group and Individual Levels via Contrastive Self-supervision [73.80009454050858]
  This work presents a self-supervised model, called DualFair, that can debias sensitive attributes like gender and race from learned representations. Our model jointly optimize for two fairness criteria - group fairness and counterfactual fairness.
  arXiv  Detail & Related papers  (2023-03-15T07:13:54Z)
• Exploring the Trade-off between Plausibility, Change Intensity and Adversarial Power in Counterfactual Explanations using Multi-objective Optimization [73.89239820192894]
  We argue that automated counterfactual generation should regard several aspects of the produced adversarial instances. We present a novel framework for the generation of counterfactual examples.
  arXiv  Detail & Related papers  (2022-05-20T15:02:53Z)
• DoLFIn: Distributions over Latent Features for Interpretability [8.807587076209568]
  We propose a novel strategy for achieving interpretability in neural network models. Our approach builds on the success of using probability as the central quantity. We show that DoLFIn not only provides interpretable solutions, but even slightly outperforms the classical CNN and BiLSTM text classification.
  arXiv  Detail & Related papers  (2020-11-10T18:32:53Z)
• Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision [57.14468881854616]
  We propose an auxiliary training objective that improves the generalization capabilities of neural networks. We use pairs of minimally-different examples with different labels, a.k.a counterfactual or contrasting examples, which provide a signal indicative of the underlying causal structure of the task. Models trained with this technique demonstrate improved performance on out-of-distribution test sets.
  arXiv  Detail & Related papers  (2020-04-20T02:47:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.