Label-Free Bias Identification in Vision Models via Gradient Probes on Concept Decompositions
This paper introduces a post-hoc, label-free method for identifying spurious correlations in frozen vision classifiers without requiring bias annotations, group labels, or retraining. The approach applies non-negative matrix factorization to intermediate activations to extract interpretable concept vectors, then ranks them using a gradient-based bias estimator derived from misclassified examples. On Colored MNIST, Waterbirds, and CelebA benchmarks, the method recovers known spurious cues and improves worst-group accuracy by up to 17.9 percentage points on Waterbirds by suppressing top-ranked concepts at inference time. Notably, the method surfaces decision-relevant directions that do not always coincide with annotated attributes, offering both an auditing tool and a debiasing handle for deployed models.
Related guides (2)
Related events (8)
Vision-Language Models Suppress Female Representations Under Ambiguous Input
This paper investigates gender bias in vision-language models (VLMs) when inputs are ambiguous (e.g., workers in full gear or seen from behind), finding that models default to male outputs even for strongly female-stereotyped occupations. The authors introduce LALS (Latent Association Leaning Score), a zero-shot metric that probes internal visual-token activations to measure concept associations across layers. Across 15 occupations, 800+ ambiguous images, and four VLMs, they find a systematic decoupling: models internally encode female associations but suppress them before generation, with male signals amplifying end-to-end while female signals peak mid-network and are filtered out. Cultural visual cues like clothing color further modulate these internal associations.
Beyond Prediction Accuracy: Target-Space Recovery Profiles for Evaluating Model-Brain Alignment
This paper introduces a framework for evaluating alignment between artificial vision models and the human visual cortex that goes beyond scalar prediction accuracy. Using repeated fMRI data from the Natural Scenes Dataset, the authors decompose brain response spaces into reproducible dimensions and measure which of these dimensions are recovered by model predictions. A key finding is that pretrained and randomly initialized models can achieve similar prediction accuracy while showing distinct recovery profiles, revealing that accuracy alone can mask fundamental model-brain mismatches. The framework also enables brain-to-brain comparisons as a diagnostic human reference baseline.
Information-theoretic formalization of the binding problem in Vision Transformers
Researchers introduce a formal information-theoretic framework for the binding problem — the challenge of associating features (color, shape) with the correct objects in multi-object scenes. They develop a probing method to measure binding information in model representations and apply it to several pre-trained Vision Transformers, examining components like the [CLS] token and spatial tokens across datasets with feature sharing, occlusion, and natural features. Results position binding information as a key factor in visual recognition and reasoning quality, and suggest current ViT architectures have limited binding capability, consistent with known failure modes.
Vision-OPD: On-Policy Self-Distillation for Fine-Grained Visual Understanding in MLLMs
Vision-OPD addresses a 'regional-to-global perception gap' in multimodal LLMs, where models answer fine-grained visual questions more accurately when given cropped evidence regions than full images. The method instantiates a crop-conditioned teacher and full-image-conditioned student from the same MLLM, minimizing token-level divergence along on-policy rollouts to transfer regional perception to the full-image policy. This self-distillation requires no external teacher models, ground-truth labels, reward verifiers, or inference-time tools. Benchmarks show competitive or superior performance against larger open-source, closed-source, and agentic 'Thinking-with-Images' models.
MedFocus: Causal Visual Attribution Framework for Chest X-ray Reasoning in Large Vision-Language Models
This paper addresses the faithfulness of visual attribution methods in Large Vision-Language Models (LVLMs) applied to chest X-ray (CXR) analysis. The authors develop a causal evaluation framework using counterfactual editing to verify whether expert-annotated regions are causally responsible for model predictions, testing 11 attribution methods across six open-source LVLMs. Finding that existing attribution methods frequently fail to identify the actual visual evidence used by models, they propose MedFocus, a concept-based attribution method using unbalanced optimal transport to localize anatomical regions and measure their causal effect on outputs. MedFocus substantially outperforms prior methods and provides spatial, concept-level, and token-level attributions.
Real Images, Worse Judgments: Evaluating VLMs on Concreteness and Imagery
This paper evaluates whether vision-language models (VLMs) benefit from real image context when making lexical judgments about word concreteness and imagery. The authors find that real-image contexts frequently hurt alignment with human ratings, especially when visual evidence is least relevant to the word being judged. Probing and canonical correlation analysis reveal that real images cause representational shifts and increased sensitivity to spurious visual cues. Instructing models to focus on text-only content at inference time partially mitigates this degradation.
The Abstraction Gap in Vision-Language Causal Reasoning
Researchers introduce a dual-probe methodology and the CAGE benchmark (49,500 questions across 5,500 images) to distinguish linguistic plausibility from faithful causal reasoning in vision-language models. An Abstraction Gap (AG) metric quantifies the normalized performance difference between text-only and chain-of-reasoning probes. Evaluating eight VLMs, seven exhibit AG exceeding 0.50—generating fluent causal text but failing structured causal chain tasks—while one model achieves near-zero AG, suggesting architectural and pretraining choices are decisive. Fine-tuning on 45,000 chain-annotated examples fails to close the gap, pointing to a fundamental capability distinction.
Social Gaze Consistency as a Semantic Cue for AI-Generated Image Detection
This paper introduces Social Gaze Consistency (SGC), a high-level semantic detection axis based on the mutual coherence of gaze direction, head-eye alignment, and pupil placement between interacting individuals in images. The authors construct a controlled diagnostic dataset with region-specific gaze perturbations and a Block-Compositional Caption Supervision scheme to train detectors without generator-fingerprint memorization shortcuts. Cross-architecture validation shows +3.7 pp improvement on the COCOAI Interaction subset when applied to FakeVLM, with gains transferring from a single inpainter (FLUX.1-Fill) to multi-generator suites. The work argues that diffusion models share a spectral weakness in periocular structure, making gaze coherence a robust, backbone-agnostic detection signal orthogonal to existing low-level artifact methods.