Research map / Vision-language and multimodal / Multimodal LLMs

Evidence-RL: Towards Evidence-intensive Visual Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Counterfactual Evidence Disentanglement (CED), a training-time counterfactual audit that rewards GRPO-trained VLM answers whose likelihood depends on object-centric evidence regions more than on matched non-evidence regions, yielding Evidence-RL.

Problem

VLMs often give correct-looking answers from language priors or shortcuts rather than image evidence; existing perception-aware RL post-training uses global perturbations or attention proxies that do not test whether a specific answer causally depends on local supporting evidence.

Approach

For each sampled response, neutralize (mean-replace in feature space) an object-centric Evidence Region and compare the log-likelihood drop against area-matched non-evidence regions; a tanh-bounded contrastive margin gates the correctness reward inside GRPO so correct answers that rely on the evidence are favored.

Key results

• Ours (Qwen2.5-VL-7B) reaches 67.13 average across nine benchmarks vs 61.19 for the base and 63.99 for the best RL baseline (SophiaVL-R1), and is the only RL method with non-negative deltas on all nine.
• On Qwen3.5-9B, Answer-CED improves average by +11.34 over the frozen base (3 seeds, +11.83 mean) and exceeds correctness-only by +12.58 average.
• Cross-backbone validation shows positive mean gains on all four backbones, e.g. +4.59 on Qwen2.5-VL-3B and +1.33 on Qwen3-VL-8B.

Surprising findings

• PAPO's global-perturbation KL reward became unstable on Qwen3.5-9B, causing repetitive reasoning loops (25% repeated-segment rate) and budget exhaustion.
• CoT-level counterfactual reward led to reward hacking by shortening reasoning chains (mean chain 3.6 tokens on counting), so the answer-only variant transferred better despite lower training reward.

Limitations

• Evidence Regions rely on weak COCO object-box proposals, so the signal depends on proposal quality and is object-centric only.
• Blank-image ScienceQA items show the reward suppresses prior-following answers, a modest ScienceQA gain.
• Multiple benchmark results for baselines are drawn from retrained or re-cited numbers with some marked lite/partial comparisons.

Experimental setup

Datasets: COCO val2017 COCO instance annotations

Benchmarks: CountBench SpatialEval HallusionBench VLMsAreBlind FREAK MathVista MMBench MMMU ScienceQA

Metrics: accuracy average gain over base

Baselines: PAPO VPPO Perception-R1 VLM-R1 SophiaVL-R1 VAPO-Thinker-7B Correctness-only GRPO

Builds on: GRPO PAPO VPPO VLM-R1

Closest papers by meaning

• Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs (ICLR 2026 · similarity 0.93)
• DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.90)
• DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning (ICML 2026 · similarity 0.90)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)
• See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning (CVPR 2026 · similarity 0.89)
• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.89)