Reproduce “Seeing Before Agreeing: Aligning Multi-Agent Consensus with Visual Evidence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuhan Wang, Shuochen Chang, Yalin Feng, Dongsheng Ma, Yuanzi Li, Zhengren Wang, Yinglong Yang, Yufei Chen and 3 more
A training-free multi-agent framework named EAGLE enhances Visual Question Answering by explicitly aligning agents' visual evidence through grounding regions and claims, rather than solely on textual rationales. This approach led to the highest average accuracy across six diverse VQA benchmarks, outperforming strong baselines by 2.54 percentage points and achieving the best efficiency in VLM agent calls.
Key insights
- Reliable multi-agent consensus in VQA necessitates aligning agents on explicit visual evidence (specific image regions and their interpretations) in addition to agreeing on the final answer.
- An LLM-based Evidence Router can effectively categorize the evidence focus required for a question, leading to more consistent and task-appropriate grounding across diverse VQA types.
- Diagnosing both answer and visual evidence consistency allows for an evidence-gated workflow, enabling efficient early exits when consensus is visually supported and guiding revisions when discrepancies are detected.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.