P1-VL: Bridging Visual Perception and Scientific Reasoning in Physics Olympiads

February 11, 2026 · View on GitHub

Paper Blog P1-VL-30B-A3B P1-VL-235B_A22B Leaderboard

HiPhO Scores

Overview

Physics demands a rigorous synthesis of abstract laws and causal logic to model the behavior of the universe. Consequently, the ability to solve complex physics problems serves as a critical benchmark, distinguishing models that merely retrieve information from those capable of genuine, first-principles reasoning. Olympiad-level competitions, such as the International Physics Olympiad (IPhO) serve as a significant surrogate to evaluate the physics reasoning capacity of language models, where problems demand deep conceptual understanding and precise system decomposition.

P1-VL is a family of open-source vision-language models engineered for advanced scientific reasoning. Our methodology 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, ranking No.3 in the model leaderboard, effectively solving tasks that require precise diagram-to-logic alignment. Notably, the standalone P1-VL outperforms previous agent-enhanced text baselines, while our agent-augmented system achieves the No. 2 overall rank globally, trailing only Gemini-3-Pro.

  • P1-VL-30B-A3B: A 30B parameter model that surpasses larger closed-source models, demonstrating exceptional efficiency
  • P1-VL-235B-A22B: A 235B parameter model achieving gold medal performance on IPhO 2025, rivaling top closed-source models

Results

P1-VL models demonstrate top-tier physics reasoning across all HiPhO contests.

HiPhO Leaderboard


FrontierScience-Olympiad Benchmark

P1-VL models achieve significant gains over their base counterparts across all three scientific domains, yielding a total score improvement of 8.0 and 9.1 points. Remarkably, even on this predominantly text-based benchmark, the multimodal P1-VL-235B-A22B outperforms its text-only sibling (P1-235B-A22B) by a margin of 2.3 points. Furthermore, when augmented with the PhysicsMinions agent framework, P1-VL-235B-A22B+PhysicsMinions attains a total score of 67.1, securing state-of-the-art performance among all evaluated open-source models.

ModelBiology/10Chemistry/40Physics/50Total/100
GPT-5.243.589.074.377.1
Gemini-3-Pro41.085.675.576.1
Claude-Opus-4.524.081.872.571.4
GPT-5.133.582.467.470.0
GPT-535.581.567.069.7
P1-VL-235B-A22B+PhysicsMinions26.377.267.367.1
Grok-433.073.267.266.2
DeepSeek-V3.2-Thinking26.374.167.365.9
P1-235B-A22B+PhysicsMinions30.071.068.065.4
Kimi-K2-Thinking20.076.665.065.1
GLM-4.720.070.669.565.0
P1-VL-235B-A22B30.071.365.564.3
o330.076.159.062.9
P1-235B-A22B22.567.265.862.0
o4-mini41.571.557.861.7
GPT-OSS-120B (high)38.863.461.560.0
Qwen3-VL-235B-A22B-Thinking26.361.957.856.3
Qwen3-235B-A22B-Thinking-250726.358.157.354.5
P1-30B-A3B15.061.956.354.4
P1-VL-30B-A3B20.058.854.052.5
Qwen3-VL-30B-A3B-Thinking18.849.443.543.4
Qwen3-30B-A3B-Thinking-250710.047.845.342.8
o120.050.940.342.5
GPT-4o3.012.414.112.3

STEM Benchmarks

Below summarizes the comparative results, illustrating that our models consistently surpass their base counterparts across both text-only and multi-modal benchmarks.

BenchmarkP1-VL-235B-A22BQwen3-VL-235B-A22B-ThinkingP1-VL-30B-A3BQwen3-VL-30B-A3B-Thinking
AIME2493.893.390.490.0
AIME2592.190.887.983.7
HMMT-Feb83.372.973.370.0
HMMT-Nov88.384.285.480.8
IMO-Answerbench70.662.365.360.3
AMOBench47.539.044.537.0
BeyondAIME70.668.565.963.8
Brumo93.390.089.283.8
CMICC83.181.679.173.4
GPQA81.477.176.573.1
LiveBench79.979.472.771.3
HLE15.913.913.412.3
MMMU78.077.273.674.8
MMMU-Pro70.269.763.462.3
EMMA-Mini71.369.664.861.4
MathVista-Mini83.982.679.479.2

🧮 HiPhO Benchmark

HiPhO (High School Physics Olympiad) is the first benchmark focused on recent Physics Olympiads from 2024-2025 with human-aligned evaluation.

hipho overview

✨ Key Features

  • Up-to-date Coverage: Includes 13 Olympiad exam papers from 2024–2025 across international and regional competitions.
  • Mixed-modal Content: Supports four modality types, spanning from text-only to diagram-based problems.
  • Professional Evaluation: Uses official marking schemes for answer-level and step-level grading.
  • Human-level Comparison: Maps model scores to medal levels (Gold/Silver/Bronze) and compares with human performance.

Coevolutionary Multimodal Multi-Agent System: PhysicsMinions

To overcome the limitations of single-model inference, P1-VL is augmented with PhysicsMinions, a coevolutionary multimodal multi-agent system that enhances test-time reasoning through structured perception, dual-stage verification, and iterative refinement.

ModuleRole
Visual StudioPerforms structured visual perception by observing, validating, and refining diagrams or plots into symbolic representations (optional when no visual input is present).
Logic StudioGenerates an initial solution and iteratively improves it via self-reflection and targeted revision.
Review StudioPerforms dual-stage verification with two dedicated verifiers: a Domain-Verifier (e.g., Physics-Verifier) that checks scientific consistency, and a General-Verifier that validates logical coherence, derivations, and numerical correctness.

Beyond physics Olympiad tasks, PhysicsMinions naturally extends to broader scientific scenarios. By dynamically selecting domain-specific verifiers and solver prompts, the agentic framework supports heterogeneous problems in physics, chemistry, and biology while maintaining a unified coevolutionary structure.


Acknowledgements

We are grateful to the open-source community for their invaluable contributions. Special thanks to:

  • Qwen3-VL - for providing the foundational base models that powered our research
  • verl - for the versatile reinforcement learning framework that enabled our training pipeline
  • vLLM - for the efficient LLM serving and inference infrastructure
  • Megatron-LM - for the large-scale model training framework

We also thank colleagues and collaborators who supported the development of P1 models, the accompanying datasets and visual assets.

🧾 Citation

If you find this work useful, please cite:

@misc{p1vl2025,
  title={P1-VL: Bridging Visual Perception and Scientific Reasoning in Physics Olympiads},
  author={P1 Team},
  year={2026},
  url={https://arxiv.org/abs/2602.09443}
}