Decoupled Physical Modeling and Execution for Physics Reasoning

Ye Zhang; Xuehang Guo; Rui Pan; Pengfei Yu; Denghui Zhang; Manling Li; Qingyun Wang

cs.LGAug 22, 2026

Decoupled Physical Modeling and Execution for Physics Reasoning

Authors: Ye Zhang, Xuehang Guo, Rui Pan, Pengfei Yu, Denghui Zhang, Manling Li, Qingyun Wang

Organizations: University of Pennsylvania · William & Mary · University of Illinois Urbana-Champaign · Amazon ·

Abstract

Physics reasoning requires constructing a consistent model of the underlying physical system rather than relying solely on symbolic or formula-based manipulation. Although large language models have shown strong ability in solving math and coding problems, they still struggle with physics problems, as these problems entangle the physical modeling process with mathematical calculations. Humans approach physics by first building a representation of the system before performing calculations. Inspired by this, we introduce a unified framework that distills intermediate representations that explicitly encode the physical modeling process and adopt a two-stage post-training strategy, where supervised fine-tuning establishes structured modeling, and reinforcement learning with rubric-based feedback improves the quality of the modeling process. Experiments on multiple multimodal physics benchmarks show that our approach generally improves physical reasoning performance across different models and datasets. Across PhysReason, PhyX, and SeePhys, physical modeling outperforms GRPO by ~3% on average. showing that explicit physical modeling is an effective strategy for improving physics reasoning in small VLMs.

Figures & tables

Figure 1: Physics reasoning errors often arise from incorrect physical modeling. Explicit modeling separates system representation from mathematical execution.

Figure 2: Overview of our physical-modeling post-training framework. Explicit physical modeling provides both structured SFT trajectories and rubric-based process rewards for GRPO, enabling the model to separate system modeling from mathematical execution.

Qwen2.5-VL-3B-Instruct
Method	PhysReason		PhyX		SeePhys		PhysUni mc		MMK-12 phys.
	P@1	P@5	P@1	P@5	P@1	P@5	P@1	P@5	P@1	P@5
Baseline	13.3	33.6	10.9	23.5	9.1	24.7	17.1	50.9	35.0	73.4
GRPO	17.3	36.8	12.6	27.0	12.4	29.8	19.1	52.7	42.4	82.4
Physics-Modeling + GRPO	22.9	47.1	13.6	34.7	16.5	33.9	21.2	53.7	46.2	84.0
Qwen2.5-VL-7B-Instruct

Table 1: Main results on PhysReason ( Zhang et al., 2025 ) , PhyX ( Shen et al., 2025 ) , and SeePhys ( Xiang et al., 2025 ) for in-domain evaluation, and on PhysUniBench ( Wang et al., 2026 ) and the physics subset of MMK-12 ( Meng et al., 2025 ) for out-of-domain evaluation. We report Pass@1 and Pass@5 for the baseline and GRPO variants. Bold indicates the best result within each model block.

Qwen2.5-VL-7B
Method	PhysReason		PhyX		SeePhys
	P@1	P@5	P@1	P@5	P@1	P@5
Baseline	25.0	47.8	17.0	35.7	14.0	28.9
SFT Cold Start	27.2	54.4	18.2	36.0	15.7	28.9
w/o Textual	22.4	43.8	15.1	29.9	14.0	25.6
Qwen3-VL-8B

Table 2: Effect of Physical Modeling SFT Cold Start

Figure 3: Pass@1 accuracy of Qwen3-VL-8B-Instruct under different token budgets.

Method	Samples w/ Reflection	Ratio
Baseline	309/500	0.618
Modeling-SFT	5/500	0.010

Table 3: Frequency of reflective reasoning patterns in 500 sampled Qwen3-VL-8B-Instruct generations.

Method	PR	PhyX	SP	PUB	MMK
w/o SFT	28.1	18.2	15.7	20.5	52.8
w/o Rubric	30.4	20.9	19.8	21.9	54.0
Full	32.1	21.4	19.8	23.5	56.8

Table 4: Component ablations on Qwen2.5-VL-7B-Instruct. We report Pass@1 accuracy on PhysReason (PR), PhyX, SeePhys (SP), PhysUniBench (PUB), and MMK-12 (MMK).

Figure 4: Performance across physics subfields for Qwen2.5-VL-7B-Instruct.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Distribution of physics reasoning error types from a preliminary analysis using a Qwen2.5-VL-7B-Instruct.

Modeling Trace	Correct	Accuracy
No modeling trace	72/400	18.0%
Self-generated trace	80/400	20.0%
Strong-model trace	154/400	38.5%

Appendix

Table 5: Performance of a fixed executor under different sources of physical modeling traces on the same 400 problems.

Data	Cat.	Orig.	Synth.	Total
PhysReason	Electro.	962	523	1485
	Mech.	567	476	1043
	Optics	353	207	560
	Thermo.	239	89	328
	Other	19	24	43
PhyX	Electro.	418	454	872

Appendix

Table 6: Category-level statistics of the constructed physical modeling trace dataset.

Dataset	Size	Count	Rate	Count	Rate
		Similarity $>0.90$		Near-Duplicate
PhysReason	467	12	2.57%	6	1.28%
PhyX	412	6	1.46%	4	0.97%
SeePhys	121	5	4.13%	0	0.00%
Total	1000	23	2.30%	10	1.00%

Appendix

Table 7: Textual overlap between the training and held-out evaluation sets

Model	Reasoning Excerpt
Baseline	“By conservation of energy: $(m_{1}-m_{2})gh+\frac{1}{2}(m_{1}+m_{2})v^{2}=0$ Solve for $v^{2}$ : $v^{2}=\frac{2(m_{2}-m_{1})gh}{m_{1}+m_{2}}$ Wait — this gives a negative value for $v^{2}$ , which is impossible. That means we have the sign wrong. The net potential energy loss is $m_{1}gh-m_{2}gh$ , which is positive, and this energy is converted to kinetic energy.”
SFT
cold start	“The system consists of two buckets connected by a lightweight, inextensible rope passing over a frictionless, massless pulley. Since the rope is inextensible, both buckets move with the same magnitude of acceleration $a$ and reach the same final speed $v$ . Determine Acceleration: Apply Newton’s Second Law to the system. The net force is the difference between the weights of the two buckets. Relate this net force to the total mass to solve for acceleration $a$ . $a=\frac{(m_{1}-m_{2})g}{m_{1}+m_{2}}$ ”

Appendix

Table 8: Reasoning excerpts from an Atwood-machine example.

Physical Reasoning

Physical reasoning studies how AI systems infer object properties, forces, and causal dynamics to predict and act in the physical world.

See all 66 papers

Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models
Jul 6, 2026Raj Jaiswal, Dhruv Jain, Rishabh Dhawan +4LLM Self-Correction Physical Reasoning
IIIT Delhi · IIT BHU · NII Tokyo +2
Solving Physics Olympiad via Reinforcement Learning on Physics Simulators
Apr 13, 2026Mihir Prabhudesai, Aryan Satpathy, Yangmin Li +6Scientific Reasoning in Language Models Synthetic Data Generation for Language Models
Carnegie Mellon University
Reward Design for Physical Reasoning in Vision-Language Models
Apr 15, 2026Derek Lilienthal, Manisha Mukherjee, Sameera HorawalavithanaVLM Reasoning Physical Reasoning
Pacific Northwest National Laboratory, Richland, WA 99352, USA

Jul 6, 2026J/K move · Enter open · S save

Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models

Raj Jaiswal, Dhruv Jain, Rishabh Dhawan, Sree Krishna Uppalapati+3

IIIT Delhi · IIT BHU · NII Tokyo+2

LLM Self-Correction Physical Reasoning RL for Language Model Reasoning Small Language Models Error Feedback

Open paper AskPDF arXiv

Physics reasoning fails structurally in small language models: an error at any step propagates forward, corrupting every inference that follows. Limited domain knowledge, hallucination under multi-step derivation, and distributional sensitivity compound this failure. We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization, without exposing it to ground truth solutions as generation targets. Unlike annotation-dependent step-level methods, no preference data construction is required and the external verifier operates exclusively at training time. Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases. Conceptual errors reduce from 89.7% to 68.7%, yet persist as the hardest failure mode across all conditions.

Abstract

Physical Reasoning

Explore similar work

Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models

Solving Physics Olympiad via Reinforcement Learning on Physics Simulators

Reward Design for Physical Reasoning in Vision-Language Models