Research map / Vision-language and multimodal / Multimodal LLMs

Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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 PEARL, an RLVR framework for VLMs that derives a perception checklist of verifiable sub-questions per reasoning instance, uses its reward to gate and reweight reasoning updates, and improves Qwen2.5-VL reasoning benchmarks over GRPO/DAPO.

Problem

Outcome-only RLVR for vision-language models verifies only the final answer, ignoring upstream visual perception errors, which leads to visual hallucinations, spurious reasoning, and reward hacking.

Approach

Build a perception checklist of QA-anchored, verifiable sub-questions per training instance; use auxiliary perception rollouts to produce a perception reward that both trains perception directly and gates/reweights reasoning updates (skip reasoning rollouts when perception reward is zero), combined with GRPO via a dual objective.

Key results

• On Qwen2.5-VL-7B, PEARL reaches 47.9 average vs 44.2 for GRPO and 40.1 for the base model; MathVerse 50.8 (+9.7 over base 41.1).
• On Qwen2.5-VL-3B, PEARL averages 39.8, above PAPO_D (39.2) and DAPO (36.6).
• On Qwen3-VL-8B, GRPO degrades average (62.0 to 61.1) while PEARL reaches 63.5.

Surprising findings

• Standard GRPO degrades Qwen3-VL-8B performance, while PEARL still improves it.
• Training only on perception probes (Perception Rewarding) does not consistently beat reasoning-only training, though it still gives gains over the base model.

Limitations

• The number of perception probes per instance is limited by information in the source QA pair and cannot cover the full visual context.
• Large-model experiments used a 2048-token response limit and 5 rollouts, causing frequent truncation for Qwen3-VL-8B and Qwen2.5-VL-32B.
• Lambda was tuned over only three values due to compute; results are sensitive to other hyperparameters.

Experimental setup

Datasets: ViRL39K MMK12 Geo3K

Benchmarks: MathVerse MathVision MathVista LogicVista WeMath DynaMath MMMU dev AI2D RealWorldQA BLINK MMMU Pro V CountBench

Metrics: accuracy

Baselines: GRPO DAPO PAPO Vision-SR1 VLAA-Thinker Visionary-R1 WeThink MM-EUREKA-Qwen VL-Rethinker

Builds on: GRPO DAPO EasyR1

Closest papers by meaning

• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.94)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.93)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.93)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.92)
• Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.91)
• Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning (ICLR 2026 · similarity 0.91)