Reproduce “VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set” on OpenResearch

About this paper

Shufan Shen, Junshu Sun, Qingming Huang, Shuhui Wang

VL-SAE gives image and text representations a shared set of interpretable concepts, allowing their activations to be compared. For large models, an auxiliary autoencoder aligns representations before sparse encoding. On LLaVA 1.5, it reaches 83.60% accuracy on POPE’s adversarial setting, versus 81.10% with VCD and 78.67% with regular decoding; this does not show it eliminates hallucinations or explains them alone.

Key insights

• A unified concept set derived from multi-modal representations effectively demystifies VLM alignment, providing a common semantic ground for interpreting vision and language interactions.
• Analyzing co-activated 'aligned concepts' reveals how VLMs filter irrelevant single-modal activations, offering semantically consistent interpretations of cross-modal understanding.
• Concept-level alignment can be leveraged to both diagnose VLM failures, such as identifying conceptual sources of object hallucinations, and actively enhance model performance.

Summary by alphaXiv.