Research map / Vision-language and multimodal / Multimodal LLMs

PhyCritic: Multimodal Critic Models for Physical AI

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces PhyCritic, a physics-focused multimodal critic trained with a two-stage GRPO pipeline (physical skill warmup, then self-referential critic finetuning) on Qwen2.5-VL-7B, plus a new PhyCritic-Bench for physical AI judging.

Problem

Existing multimodal critic/judge models are trained mainly on general visual domains and do not reliably judge physical perception, causal reasoning, and planning responses.

Approach

The critic first solves the query itself (self-prediction) and then judges candidate response pairs with explicit reference to that self-prediction; trained with GRPO using self-prediction and critic-accuracy rewards after a physical RLVR warmup.

Key results

• PhyCritic-7B reaches 68.0 overall on PhyCritic-Bench, the best among open-source 7B/8B models (Qwen2.5-VL-7B: 51.6).
• On CosmosReason1-Bench PhyCritic-7B scores 63.9, the best open-source result, vs Cosmos-R1-7B at 63.0.
• Ablation shows removing self-referential critic finetuning drops PhyCritic-Bench from 68.0 to 64.4.

Surprising findings

• Physical RL alone gives limited judging gains (+2.0 on PhyCritic-Bench) even though it improves physical reasoning.
• Removing detailed critic prompt criteria causes clear drops, contrasting with prior work that uses simple critic prompts.

Limitations

• Self-referential training requires ground-truth answers for prompts, limiting use in fully open-ended settings.
• Critic training data is small (3,258 pairs) and labels rely on GPT-4o answer verification.
• PhyCritic-Bench is small (225 samples).

Experimental setup

Datasets: Cosmos-Reason1 RL dataset RoboVQA BridgeData V2 HoloAssist AgiBot World LingoQA VL-RewardBench Multimodal RewardBench CV-Bench EgoPlanBench2

Benchmarks: PhyCritic-Bench VL-RewardBench Multimodal RewardBench CosmosReason1-Bench CV-Bench EgoPlanBench2

Metrics: pairwise judgment accuracy overall accuracy macro accuracy

Baselines: Qwen2.5-VL-7B Eagle-2.5-8B Cosmos-Reason1-7B RoboBrain2.0-7B UnifiedReward-Think-Qwen-7B Gemini-2.5-Pro GPT-4o

Builds on: GRPO DeepSeek-R1 Cosmos-Reason1 Qwen2.5-VL JudgeBench

Closest papers by meaning

• Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following (CVPR 2026 · similarity 0.88)
• LLaVA-Critic: Learning to Evaluate Multimodal Models (CVPR 2025 · similarity 0.88)
• Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning (CVPR 2025 · similarity 0.87)
• PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.87)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.87)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.86)