Research map / Vision-language and multimodal / Multimodal LLMs
Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning
Vishwas Sathish, Viresh Ranjan, Xinliang Zhu, Arnab Dhua, Douglas Gray
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 search-aware factor to the reward (ablation) | SVRL (ablation chain) · FVQA-test | accuracy (%): 57.7 -> 59.9 | Improved |
| add search-aware factor on top of query alignment (ablation) | SVRL · FVQA-test | search ratio (%): 97.6 -> 80.3 | Improved |
| add train-time self-verification objective (vs external test-time verifier) (ablation) | SVRL · FVQA-test | accuracy (%): 60.6 -> 64.2 | Improved |
| add self-verification objective (ablation) | SVRL · InfoSeek | accuracy (%): 58.3 -> 64.1 | Improved |
| use external test-time verification to filter queries and snippets (ablation) | SVRL-search-aware · FVQA-test | accuracy (%): 59.9 -> 60.6 | Improved |
| replace exact-match reward with LLM-judge reward (ablation) | SVRL · FVQA-test | accuracy (%): 58.1 vs 57.7 | Mixed |
| SVRL vs MMSearch-R1++ baseline | Qwen2.5-VL-7B · FVQA-test | accuracy (%): 56.4 -> 65.3 | Improved |
| SVRL vs MMSearch-R1++ baseline | Qwen2.5-VL-7B · InfoSeek | accuracy (%): 55.1 -> 64.7 | Improved |
| SVRL vs MMSearch-R1++ baseline | Qwen2.5-VL-7B · SimpleVQA | accuracy (%): 57.4 -> 58.3 | Improved |
| SVRL reduces search ratio vs MMSearch-R1++ | Qwen2.5-VL-7B · FVQA-test | search ratio (%): 80.3 -> 61.4 | Improved |
| apply SVRL to Qwen3-VL-8B | Qwen3-VL-8B · FVQA | accuracy (%): 56.4 -> 64.9 | Improved |
| SVRL-7B vs GPT-5.2-thinking under adaptive search | Qwen2.5-VL-7B · FVQA | accuracy (%) GPT-5.0 judge: 60.8 -> 65.3 | Improved |
Closest papers by meaning
- Diving into Self-Evolving Training for Multimodal Reasoning (ICML 2025 · similarity 0.92)
- CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning (CVPR 2026 · similarity 0.91)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.90)
- Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.90)
- Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.90)
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.
