Paper 2602.11124
PhyCritic: Multimodal Critic Models for Physical AI
- Published
- Feb 2026
- Research lab
- NVIDIA
- Citations
- 9
- GitHub
- Not linked
01 In brief
Summary
PhyCritic is a multimodal critic model designed for physical AI tasks, addressing the lack of physics-aware evaluation in existing judge models.
It uses a two-stage RLVR pipeline: first, a physical skill warmup stage with GRPO on physical QA pairs to enhance perception and reasoning; second, self-referential critic finetuning, where the critic generates its own prediction before judging candidate responses, grounding its evaluation in its own understanding.
The model is trained on a dataset built from embodied datasets (RoboVQA, BridgeData V2, HoloAssist, AgiBot World) with questions from Cosmos-Reason1, and evaluated on the new PhyCritic-Bench benchmark.
Results show PhyCritic-7B outperforms open-source 7B/8B baselines on PhyCritic-Bench (68.0 vs.
56.0 for Eagle-2.5-8B), generalizes to general reward benchmarks (VL-RewardBench, Multimodal RewardBench), and improves physical reasoning on CosmosReason1-Bench, CV-Bench, and EgoPlanBench2.
Ablations confirm the necessity of both stages and the self-referential mechanism.
PhyCritic also serves as an effective judge for best-of-N sampling and as a reward for DPO training, improving downstream policy performance.
02 From the paper
Abstract
With the rapid development of large multimodal models, reliable judge and critic models have become essential for open-ended evaluation and preference alignment, providing pairwise preferences, numerical scores, and explanatory justifications for assessing model-generated responses. However, existing critics are primarily trained in general visual domains such as captioning or image question answering, leaving physical AI tasks involving perception, causal reasoning, and planning largely underexplored. We introduce PhyCritic, a multimodal critic model optimized for physical AI through a two-stage RLVR pipeline: a physical skill warmup stage that enhances physically oriented perception and reasoning, followed by self-referential critic finetuning, where the critic generates its own prediction as an internal reference before judging candidate responses, improving judgment stability and physical correctness. Across both physical and general-purpose multimodal judge benchmarks, PhyCritic achieves strong performance gains over open-source baselines and, when applied as a policy model, further improves perception and reasoning in physically grounded tasks.