New RECAP method boosts vision-language model reasoning

Researchers have developed a new training method called RECAP to improve vision-language models (VLMs) by addressing temporal grounding failures. This method enhances how models utilize image evidence during multi-step reasoning processes. RECAP achieves this by using revisit-conditioned advantage estimation and other techniques to propagate rewards through visually critical reasoning steps without requiring additional rewards or inference-time changes. Across several VLM backbones, RECAP demonstrated consistent performance gains on benchmarks like HallusionBench, improving image-grounded reasoning capabilities. AI

IMPACT Enhances image-grounded reasoning in VLMs, potentially improving their ability to interpret and act on visual information.

RANK_REASON The cluster contains a research paper detailing a new training method for vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• GRPO
• HallusionBench
• Hugging Face
• InternVL3-8B
• LLaVA-OV-7B
• Qwen2.5-VL-7B
• RECAP
• vision-language models
• VPPO

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →