AI Paper Atlas

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

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
add search-aware factor to the reward (ablation)SVRL (ablation chain) · FVQA-testaccuracy (%): 57.7 -> 59.9Improved
add search-aware factor on top of query alignment (ablation)SVRL · FVQA-testsearch ratio (%): 97.6 -> 80.3Improved
add train-time self-verification objective (vs external test-time verifier) (ablation)SVRL · FVQA-testaccuracy (%): 60.6 -> 64.2Improved
add self-verification objective (ablation)SVRL · InfoSeekaccuracy (%): 58.3 -> 64.1Improved
use external test-time verification to filter queries and snippets (ablation)SVRL-search-aware · FVQA-testaccuracy (%): 59.9 -> 60.6Improved
replace exact-match reward with LLM-judge reward (ablation)SVRL · FVQA-testaccuracy (%): 58.1 vs 57.7Mixed
SVRL vs MMSearch-R1++ baselineQwen2.5-VL-7B · FVQA-testaccuracy (%): 56.4 -> 65.3Improved
SVRL vs MMSearch-R1++ baselineQwen2.5-VL-7B · InfoSeekaccuracy (%): 55.1 -> 64.7Improved
SVRL vs MMSearch-R1++ baselineQwen2.5-VL-7B · SimpleVQAaccuracy (%): 57.4 -> 58.3Improved
SVRL reduces search ratio vs MMSearch-R1++Qwen2.5-VL-7B · FVQA-testsearch ratio (%): 80.3 -> 61.4Improved
apply SVRL to Qwen3-VL-8BQwen3-VL-8B · FVQAaccuracy (%): 56.4 -> 64.9Improved
SVRL-7B vs GPT-5.2-thinking under adaptive searchQwen2.5-VL-7B · FVQAaccuracy (%) GPT-5.0 judge: 60.8 -> 65.3Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy