Source-linked AI summary
P1-VL: Bridging Visual Perception and Scientific Reasoning in Physics Olympiads
Yun Luo, Futing Wang, Qianjia Cheng, Fangchen Yu, Haodi Lei, Jianhao Yan, Chenxi Li, Jiacheng Chen, Yufeng Zhao, Haiyuan Wan, Yuchen Zhang, Shenghe Zheng, Junchi Yao, Qingyang Zhang, Haonan He, Wenxuan Zeng, Li Sheng, Chengxing Xie, Yuxin Zuo, Yizhuo Li, Yulun Wu, Rui Huang, Dongzhan Zhou, Kai Chen, Yu Qiao, Lei Bai, Yu Cheng, Ning Ding, Bowen Zhou, Peng Ye, Ganqu Cui
TL;DR
Physics reasoning benchmarks require models to combine visual perception with physical laws, while prior approaches largely emphasize text. P1-VL combines curriculum reinforcement learning with progressive difficulty and agentic inference-time self-verification; it achieves 12 gold medals and 1 silver on HiPhO with its flagship model, and agent augmentation reaches 2nd overall. The models also generalize across broader scientific and STEM benchmarks, while evaluation excludes some competitions lacking complete official score distributions and includes timeout cases for Kimi-K2-Thinking.
Problem
Physics Olympiad problems require multimodal perception and causal reasoning because diagrams can contain essential constraints absent from text, but prior approaches predominantly focus on textual reasoning.
Method
P1-VL combines curriculum reinforcement learning with progressive difficulty expansion and stabilization mechanisms, plus agentic augmentation for iterative correction and self-verification at inference.
Results
12 gold medals and 1 silver across HiPhO made P1-VL-235B-A22B the first specialized open-source physics vision-language model to achieve this result, while agent augmentation reached 2nd overall and both variants generalized across STEM benchmarks.
Takeaways & Limitations
P1-VL provides an open-source foundation for aligning visual perception with physical laws and advancing general-purpose physical intelligence.
Takeaways & Limitations
The evaluation excludes CPhO, USAPhO, and APhO-2024 because complete official contestant score distributions were unavailable, and 9.58% of Kimi-K2-Thinking cases exceed its 2-hour inference timeout.
Abstract
from arXiv · showhide
The transition from symbolic manipulation to science-grade reasoning represents a pivotal frontier for Large Language Models (LLMs), with physics serving as the critical test anchor for binding abstract logic to physical reality. Physics demands that a model maintain physical consistency with the laws governing the universe, a task that fundamentally requires multimodal perception to ground abstract logic in reality. At the Olympiad level, diagrams are often constitutive rather than illustrative, containing essential constraints, such as boundary conditions and spatial symmetries, that are absent from the text. To bridge this visual-logical gap, we introduce P1-VL, a family of open-source vision-language models engineered for advanced scientific reasoning. Our method harmonizes Curriculum Reinforcement Learning, which employs progressive difficulty expansion to stabilize post-training, with Agentic Augmentation, enabling iterative self-verification at inference. Evaluated on HiPhO, a rigorous benchmark of 13 exams from 2024-2025, our flagship P1-VL-235B-A22B becomes the first open-source Vision-Language Model (VLM) to secure 12 gold medals and achieves the state-of-the-art performance in the open-source models. Our agent-augmented system achieves the No.2 overall rank globally, trailing only Gemini-3-Pro. Beyond physics, P1-VL demonstrates remarkable scientific reasoning capacity and generalizability, establishing significant leads over base models in STEM benchmarks. By open-sourcing P1-VL, we provide a foundational step toward general-purpose physical intelligence to better align visual perceptions with abstract physical laws for machine scientific discovery.
1. Introduction
P1-VL targets the multimodal and scientific-reasoning demands of physics Olympiads by combining curriculum reinforcement learning with agentic self-verification. On HiPhO, it achieves leading open-source results and generalizes across broader STEM evaluations.
- Motivation: Physics Olympiad problems test science-grade reasoning by requiring multimodal perception, causal logic, and mapping physical intuition to abstract laws.Diagrams can contain essential constraints absent from textual problem statements.
- Contribution: P1-VL is an open-source vision-language model family designed to bridge visual perception and scientific reasoning.The paper positions this bridge as a step toward general-purpose physical intelligence.
- Method: Curriculum RL progressively expands difficulty and uses stabilization mechanisms to sustain reasoning improvement during post-training.The framework is designed to address reward sparsity, entropy collapse, and training stagnation.
- Method: Agentic Augmentation enables iterative correction and self-verification through multi-turn reasoning, critique, and solution refinement at inference.This extends test-time problem-solving depth without adding training parameters.
- Results: 12 gold medals and 1 silver across 13 HiPhO exams made P1-VL-235B-A22B the first specialized open-source physics vision-language model to achieve this result.The benchmark aggregates recent Olympiad exams from 2024–2025.
- Results: P1-VL-30B-A3B achieved 9 gold medals, while agent augmentation raised P1-VL to 2nd overall and both variants improved over baselines across biology, chemistry, physics, mathematics, and multimodal STEM tasks.The reported gains include 8.0 and 9.1 points on FrontierScience-Olympiad for the two variants, respectively.
2. Physics Dataset
P1-VL uses a curated multimodal physics dataset built from Olympiads, textbooks, and competition guides, with automated filtering and human verification to improve reliability.
- Dataset Overview: 8,033 problems form the curated multimodal dataset, including physics Olympiads, undergraduate textbooks, and competition guides.The sources contribute 4,126, 2,968, and 939 problems, respectively.
- Dataset Overview: Physics Olympiad problems are prioritized because they require conceptual modeling, multi-step inference, and rule-verifiable outcomes.
- Dataset Construction: The dataset targets problems that support structured reasoning aligned with physical laws and empirical consistency.
- Dataset Construction: The collection pipeline combines OCR correction, answer cross-validation, task filtering, visual consistency checks, and expert review.Non-consensus answer extractions are discarded, and samples with missing referenced figures are filtered.
3. Approach
P1-VL formulates physics problem solving as reinforcement learning over multimodal reasoning trajectories, combining verifiable rewards, curriculum difficulty expansion, stabilization mechanisms, and visual-agent processing.
- RL Formulation: P1-VL models physics problem solving as an RL process over states, token actions, deterministic transitions, and trajectory rewards.The state includes the problem statement and previously generated reasoning tokens.
- RL Formulation: The learning objective maximizes expected return by optimizing the policy over trajectories sampled from the model.
- RL Formulation: GSPO performs sequence-level optimization with length-normalized likelihood ratios and sequence-level advantages.Length normalization is used to reduce variance.
- Reward and Verification: P1-VL uses binary correctness rewards and test-case-style aggregation for problems containing multiple required sub-answers.The final reward aggregates correctness indicators across sub-answers.
- Curriculum Training: The curriculum expands task difficulty progressively, filtering trivial samples and recovering or refining zero-shot failures before harder training stages.Samples with D(x_i, y_i) > 0.7 are treated as trivial, while zero-pass-rate samples undergo a recovery pipeline.
- Curriculum Training: Exploration budgets grow with curriculum difficulty through larger group sizes and longer generation windows.
- Training Stabilization: Sequence-Level Masked Importance Sampling rejects out-of-distribution trajectories to mitigate instability from training-inference engine mismatch.Geometric-mean importance weights normalize the metric across sequence lengths, with rejection controlled by threshold C.
- Training Dynamics: Curriculum training steadily improves HiPhO validation performance while response lengths and exploration space increase across stages.The authors interpret the longer responses as evidence of deeper reasoning capacity.
4. Experiment
The experiments evaluate P1-VL on HiPhO, using standardized inference and automated grading, and compare its performance with open- and closed-source models. P1-VL achieves strong Olympiad results, including 12 gold medals for its largest model and further gains from agentic augmentation.
- Experimental Setup: HiPhO aggregates 13 physics Olympiad exams administered between 2024 and 2025 across seven international and regional competition series.The benchmark prioritizes globally influential competitions with available human performance data.
- Experimental Setup: The evaluation averages eight inference runs per problem before aggregating scores across problems for each Olympiad exam.Models use temperature 0.6 and the largest permitted token limit under the HiPhO protocol.
- Results: P1-VL-235B-A22B ranks third overall with 12 gold and 1 silver medals and an average score of 39.3.It exceeds Gemini-2.5-Pro, GPT-5, and Grok-4, and improves 5.4 points over Qwen3-VL-235B-A22B-Thinking.
- Results: P1-VL-30B-A3B achieves 9 gold and 4 silver medals with an average score of 35.0, ranking third among open-source models.Despite its smaller size, it outperforms larger Qwen3-VL baselines.
- Results: Agentic augmentation raises P1-VL-235B-A22B’s average score from 39.3 to 40.9, producing the second-place global ranking.The augmented system surpasses GPT-5.2 and reaches state-of-the-art performance on three physics Olympiads.
- Evaluation Scope: The HiPhO table uses exam scores and theoretical exam sections, with model full marks constrained not to exceed human full marks.The benchmark awards gold, silver, and bronze according to stated score thresholds.
5. Discussion
The discussion examines whether physics-focused reinforcement learning generalizes beyond physics and how training choices affect stability. P1-VL improves over its base models across text-only and multimodal STEM benchmarks, while curriculum training and mixed data address training limitations.
- Generalizability of P1-VL: P1-VL-235B-A22B and P1-VL-30B-A3B improve over their base counterparts across all three FrontierScience-Olympiad scientific domains.Their total scores increase by 8.0 and 9.1 points, respectively.
- Generalizability of P1-VL: The multimodal P1-VL-235B-A22B exceeds its text-only sibling by 2.3 points on the text-only FrontierScience-Olympiad benchmark.With PhysicsMinions, it reaches a total score of 67.1 and state-of-the-art performance among evaluated open-source models.
- Out-of-Domain Evaluation: Across text-only and multimodal STEM benchmarks, P1-VL models consistently surpass their respective base counterparts.The evaluation includes mathematical reasoning and broader STEM tasks.
- Out-of-Domain Evaluation: On AMOBench, P1-VL-235B-A22B and P1-VL-30B-A3B outperform their baselines by 8.5 and 7.5 points, respectively.On EMMA-Mini, the corresponding gains are 1.7 and 3.4 points.
- Training Stability: RL training of Qwen3-VL-MoE models on physics data can catastrophically collapse because of training-inference mismatch.The discussion attributes the mismatch to discrepancies between inference and training frameworks; Masked Importance Sampling stabilizes training.
- Training Data Composition: Adding text-only data padded with blank images produces superior performance almost always rather than negative transfer in the analyzed Qwen3-VL-4B-Thinking setting.This supports using mixed multimodal training data.
- Curriculum Training: Without difficulty expansion, response length stagnates across early curriculum stages and performance gains remain negligible.Curriculum training instead increases average response length and reasoning accuracy.
6. Conclusion
The paper concludes that P1-VL addresses the visual-logical gap in physics through curriculum reinforcement learning and agentic test-time reflection. Its HiPhO results and broader scientific generalization support the model family’s potential for complex scientific reasoning.
- Conclusion: P1-VL is presented as the first open-source vision-language model family to achieve Olympiad-level proficiency in physics.The approach targets physical problem solving that combines visual constraints with causal logic.
- Conclusion: Curriculum RL trains robust reasoning patterns, while Agentic Augmentation enables test-time reflection.The conclusion connects these mechanisms to P1-VL’s physics reasoning results.
- Conclusion: P1-VL-235B-A22B secures 12 gold medals on HiPhO and demonstrates generalization across broader scientific disciplines.The paper presents these results as evidence of complex scientific reasoning capacity.
- Implications: The authors position visual-constraint synthesis with causal logic as relevant to reliable world models and embodied AI.They release P1-VL to accelerate research toward machine scientific discovery.
7. Acknowlegement
The work was supported by Shanghai AI Laboratory and a locally commissioned task from the Shanghai Municipal Government. The authors also acknowledge open-source projects that supported implementation.
- Acknowledgements: The work received support from Shanghai AI Laboratory and a locally commissioned task from the Shanghai Municipal Government.The acknowledgement also thanks developers and maintainers of open-source projects used in implementation.
- Acknowledgements: Qwen3-VL, Qwen, and slime are acknowledged for providing foundational models or implementation infrastructure.The passage identifies these projects as critical to the work’s implementation.
A.1. RL Training on Intern Series
The authors test their reinforcement-learning strategy on two Intern models and report consistent HiPhO improvements, including stronger medal outcomes with a model-based verifier.
- InternVL-3.5-30B-A3B improves from 18.3 to 22.6 on HiPhO after reinforcement-learning training.
- Intern-S1-mini improves from 8.5 to 12.4 on HiPhO after reinforcement-learning training.
- With Gemini-2.5-Flash as verifier, trained Intern-S1-mini earns one silver and ten bronze medals, versus ten bronze medals for the baseline.
- The reported improvements support the effectiveness of the reinforcement-learning strategy in eliciting latent model capabilities.
A.2. Case Study I: Text+Variable Figure
The case study evaluates P1-VL on a diagram-dependent IPhO problem involving Cox’s timepiece, requiring visual interpretation, tabular integration, and hydrostatic force analysis. P1-VL-235B-A22B achieves a perfect score while deriving the relevant physical behavior and force conditions.
- Case-study setup: P1-VL-235B-A22B achieves a perfect score of 1.0/1.0 on the 2025 IPhO Cox’s timepiece problem.The problem requires interpreting tube configurations, figure information, and table parameters.
- Reasoning capabilities: The solution translates tube schematics, liquid levels, gas volumes, and tabular parameters into hydrostatic equations and work-related constraints.
- Hydrostatic analysis: For the tube configuration in Figure 11(b), the water pressure at the top is reduced by the weight of the water column of height h.
- Force balance: The holding force must counteract the tube’s weight and the downward pressure difference between atmospheric pressure and the reduced interior pressure.
- Experiment outcomes: Behavior A occurs when h* > H, whereas Behavior B occurs when h* < H and the internal water level stops at h* as the tube continues rising.