New framework evaluates vision-language models by tracing input influence

Researchers have introduced a new framework to evaluate vision-language models (VLMs) by analyzing how different inputs influence their output generation process. This temporal causal drive evaluation framework uses interventions to measure the impact of visual information, question text, and generated prefixes on the model's autoregressive decoding. Experiments with models like Qwen3-VL-8B-Instruct and InternVL2-8B demonstrated a shift from initial reliance on visual and question inputs to a greater dependence on generated prefixes as the output progresses. The framework showed significant improvements in reducing recovery error compared to observational methods and proved effective in distinguishing visually grounded generations from those driven by prior knowledge. AI

IMPACT This framework offers a novel way to diagnose VLM behavior, potentially leading to more robust and interpretable models.

RANK_REASON The cluster describes a new evaluation framework for vision-language models presented in an academic paper.

Read on Hugging Face Daily Papers →

• alphaXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• InternVL2-8B
• LLaVA-Video-178K
• MAVIS
• MiraData
• Qwen3-VL-8B-Instruct
• ScienceCast
• Temporal Causal Drive Evaluation Framework
• VLMBias
• arXiv
• Prefix Causal Drive
• Question Causal Drive
• Vision--Language Models
• Visual Causal Drive

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →