Research map / Vision-language and multimodal / Visual question answering
MMAgent-R2: Learning to Rerank and Reject for Agentic mRAG
Tao Zhang, Ziqi Zhang, Zongyang Ma, Yuxin Yang, Bing Li, Chunfeng Yuan, Kang Rong, Fengyun Rao, Jing LYU, Weiming Hu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add visual reranking (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 40.2 -> 44.3 | Improved |
| add active rejection on top of reranking (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 44.3 -> 45.5 | Improved |
| increase candidate batch size K from 3 to 5 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 44.7 -> 45.5 | Improved |
| increase candidate batch size K from 5 to 7 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 45.5 -> 45.7 | No clear effect |
| increase max rejection count from 0 to 2 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 44.3 -> 45.5 | Improved |
| increase max rejection count from 2 to 3 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 45.5 -> 44.8 | Hurt |
| increase rejection reward delta from 0.1 to 0.2 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 44.9 -> 45.5 | Improved |
| increase rejection reward delta from 0.2 to 0.3 (ablation) | MMAgent-R2-8B · InfoSeek (26K train) | accuracy (All): 45.5 -> 45.2 | Hurt |
| add visual reranking and rejection to retriever top-1 | MMAgent-R2-8B · E-VQA | entity identification accuracy: 15.4 -> 29.5 | Improved |
| use MMAgent-R2 instead of MMhops-R1 (no visual verification) | agentic mRAG · InfoSeek all | accuracy: 33.2 -> 50.2 | Improved |
| use MMAgent-R2-8B vs ReAG | mRAG · E-VQA all | accuracy: 47.0 -> 54.2 | Improved |
| use MMAgent-R2-8B vs Gemini-2.5-Pro | multi-hop VQA · MMhops Comparison | accuracy: 29.4 -> 39.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ReAG — ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering (CVPR 2026) · better on 4 of 4 shared results
- MMhops-R1 — MMhops-R1: Multimodal Multi-hop Reasoning (AAAI 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger (ICML 2025 · similarity 0.89)
- MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models (ICLR 2025 · similarity 0.89)
- Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent (ICLR 2025 · similarity 0.89)
- Retrieval-Augmented Visual Question Answering via Built-in Autoregressive Search Engines (AAAI 2025 · similarity 0.89)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.88)
- Visual Haystacks: A Vision-Centric Needle-In-A-Haystack Benchmark (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
