Language Models Can Explain Visual Features via Steering

• URL: http://arxiv.org/abs/2603.22593v2
• Date: Wed, 25 Mar 2026 08:24:41 GMT
• Title: Language Models Can Explain Visual Features via Steering
• Abstract summary: Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge.<n>We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image.<n>We prompt the language model to explain what it sees'', effectively eliciting the visual concept represented by each feature.
• Score: 8.36855091633073
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Related papers

• Decoupling the Declarative from the Procedural in Vision-Language-Action Models [65.22639791239932]
  Policy trained to clone a behavior from object-specific demonstrations must generalize beyond that object.<n>We propose w$2$VLA, a new Vision-Language-Action model with restructured information flow.<n>Unlike popular, state-of-the-art VLAs, we show that our modular approach successfully decouples knowledge representations.
  arXiv  Detail & Related papers  (2026-06-19T14:43:58Z)
• Towards Understanding Multimodal Fine-Tuning: Spatial Features [25.349396112139214]
  Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model.<n>We present the first mechanistic analysis of VLM adaptation using stage-wise model diffing.
  arXiv  Detail & Related papers  (2026-02-06T18:48:18Z)
• Vision-Grounded Machine Interpreting: Improving the Translation Process through Visual Cues [0.0]
  Vision-Grounded Interpreting (VGI) is a novel approach designed to address the limitations of unimodal machine interpreting.<n>We present a prototype system that integrates a vision-language model to process both speech and visual input from a webcam.<n>To evaluate the effectiveness of this approach, we constructed a hand-crafted diagnostic corpus targeting three types of ambiguity.
  arXiv  Detail & Related papers  (2025-09-28T16:25:33Z)
• An Iterative Feedback Mechanism for Improving Natural Language Class Descriptions in Open-Vocabulary Object Detection [0.08974531206817744]
  We present an approach for improving non-technical users' natural language text descriptions of their desired targets of interest.<n>We quantify the improvement that our feedback mechanism provides by demonstrating performance with multiple publicly-available open-vocabulary object detection models.
  arXiv  Detail & Related papers  (2025-03-21T16:34:04Z)
• Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment [57.0121616203175]
  We propose FiSAO, a novel self-alignment method that utilizes the model's own visual encoder as a fine-grained verifier to improve vision-language alignment.<n>By leveraging token-level feedback from the vision encoder, FiSAO significantly improves vision-language alignment, even surpassing traditional preference tuning methods that require additional data.
  arXiv  Detail & Related papers  (2024-10-18T03:34:32Z)
• Enhancing Large Vision Language Models with Self-Training on Image Comprehension [131.14381425260706]
  We introduce Self-Training on Image (STIC), which emphasizes a self-training approach specifically for image comprehension. First, the model self-constructs a preference for image descriptions using unlabeled images. To further self-improve reasoning on the extracted visual information, we let the model reuse a small portion of existing instruction-tuning data.
  arXiv  Detail & Related papers  (2024-05-30T05:53:49Z)
• Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvement [102.22911097049953]
  Large vision-language models (LVLMs) have achieved impressive results in visual question-answering and reasoning tasks.<n>Existing methods often depend on external models or data, leading to uncontrollable and unstable alignment results.<n>We propose SIMA, a self-improvement framework that enhances visual and language modality alignment without external dependencies.
  arXiv  Detail & Related papers  (2024-05-24T23:09:27Z)
• Calibrated Self-Rewarding Vision Language Models [27.686545023186852]
  Large Vision-Language Models (LVLMs) have made substantial progress by integrating pre-trained large language models (LLMs) and vision models through instruction tuning. LVLMs often exhibit the hallucination phenomenon, where generated text responses appear linguistically plausible but contradict the input image. We propose the Calibrated Self-Rewarding (CSR) approach, which enables the model to self-improve by iteratively generating candidate responses, evaluating the reward for each response, and curating preference data for fine-tuning.
  arXiv  Detail & Related papers  (2024-05-23T14:30:33Z)
• Multimodal Large Language Model is a Human-Aligned Annotator for Text-to-Image Generation [87.50120181861362]
  VisionPrefer is a high-quality and fine-grained preference dataset that captures multiple preference aspects. We train a reward model VP-Score over VisionPrefer to guide the training of text-to-image generative models and the preference prediction accuracy of VP-Score is comparable to human annotators.
  arXiv  Detail & Related papers  (2024-04-23T14:53:15Z)
• Diffexplainer: Towards Cross-modal Global Explanations with Diffusion Models [51.21351775178525]
  DiffExplainer is a novel framework that, leveraging language-vision models, enables multimodal global explainability. It employs diffusion models conditioned on optimized text prompts, synthesizing images that maximize class outputs. The analysis of generated visual descriptions allows for automatic identification of biases and spurious features.
  arXiv  Detail & Related papers  (2024-04-03T10:11:22Z)
• Chain-of-Spot: Interactive Reasoning Improves Large Vision-Language Models [81.71651422951074]
  Chain-of-Spot (CoS) method is a novel approach that enhances feature extraction by focusing on key regions of interest. This technique allows LVLMs to access more detailed visual information without altering the original image resolution. Our empirical findings demonstrate a significant improvement in LVLMs' ability to understand and reason about visual content.
  arXiv  Detail & Related papers  (2024-03-19T17:59:52Z)
• Interpreting and Controlling Vision Foundation Models via Text Explanations [45.30541722925515]
  We present a framework for interpreting vision transformer's latent tokens with natural language. Our approach enables understanding of model visual reasoning procedure without needing additional model training or data collection.
  arXiv  Detail & Related papers  (2023-10-16T17:12:06Z)

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.