Skip to content

Decoupled Physical Modeling and Execution for Physics Reasoning

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/yzhg9/PhysicsModeling
Area: LLM Reasoning
Keywords: Multimodal Physics Reasoning, Physical System Modeling, Decoupled Reasoning, Reinforcement Learning, Rubric-based Rewards

TL;DR

Addressing conceptual breakdowns in multimodal LLMs caused by conflating physical modeling with mathematical calculations, this paper proposes an explicit two-stage framework that establishes structured physical modeling via curriculum SFT and refines physical grounding using rubric-guided GRPO.

Background & Motivation

Multimodal Large Language Models (MLLMs) have demonstrated impressive proficiency across structured symbolic reasoning domains such as competition mathematics and automated code synthesis. However, when deployed on multimodal physics problem solving, these models consistently stumble upon seemingly elementary scenarios. The foundational cognitive mechanism of physics reasoning diverges fundamentally from pure algebraic manipulation: human experts never dive directly into equations upon inspecting an unfamiliar physical setup; rather, they first construct a coherent internal mental model of the physical system, identifying interacting bodies, boundary conditions, reference frames, and governing conservation laws. In stark contrast, standard chain-of-thought methods prompt MLLMs to generate monolithic reasoning paths that inextricably entangle physical system modeling with symbolic calculations, frequently leading to fatal conceptual errors before numerical computation even commences.

A granular diagnostic of 500 error instances produced by Qwen2.5-VL-7B corroborates this structural failure: the vast majority of incorrect responses stem directly from erroneous or absent physical modeling rather than algebraic miscalculations. Specifically, errors are dominated by violations of physical law applicability conditions (19.8%), incorrect governing principle selection (11.4%), missing system representations (10.8%), and incomplete constraint modeling (9.0%). Furthermore, an intervention experiment holding a downstream mathematical executor fixed reveals that feeding it an accurate physical modeling trace from a stronger model boosts end-to-end solving accuracy from 20.0% to 38.5%. This demonstrates that the underlying bottleneck in multimodal scientific reasoning does not reside in equation solving, but rather in establishing the formal physical representation that defines the system.

While recent reasoning post-training approaches incorporate step-by-step planning or general rubric-based process rewards, they evaluate generic structural or grammatical coherence rather than checking whether the underlying physical model faithfully respects physical laws. The core idea is to explicitly decouple physics reasoning into structured physical modeling and symbolic mathematical execution, establishing a standardized modeling representation through text-to-multimodal curriculum SFT, and subsequently guiding policy optimization via GRPO with fine-grained, physics-specific rubrics distilled directly from verified modeling traces.

Method

Overall Architecture

The framework decomposes the conditional joint probability of solving a physics problem into two explicit, sequential stages: given a multimodal problem \(x\), the model first produces an explicit physical system modeling trace \(\tau\), and subsequently derives the mathematical execution process \(z\) and final answer \(y\) conditioned on \(\tau\), formulated as \(\pi_\theta(\tau, z, y \mid x) = \pi_\theta(z, y \mid \tau, x) \pi_\theta(\tau \mid x)\). The post-training pipeline consists of two complementary phases: a supervised fine-tuning (SFT) cold start that instills structured modeling habits, followed by Group Relative Policy Optimization (GRPO) reinforced with fine-grained rubric-based process rewards that penalize physically inconsistent system representations.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multimodal Physics Problem and Diagram Input"] --> B["Text-to-Multimodal Curriculum SFT Cold Start<br/>Establish standardized system/variable/constraint/process modeling"]
    B --> C["Physical Modeling Trace-Guided Symbolic Execution<br/>Translate abstract physical representations into solvable equations"]
    C --> D["Atomic Fine-Grained Physics Rubric Derivation<br/>Distill visual/modeling/planning/execution criteria from verified traces"]
    D --> E["Rubric-Guided Multi-Reward GRPO Policy Optimization<br/>Combine outcome accuracy, format validity, and weighted rubric rewards"]
    E --> F["Physically Grounded System Modeling and Robust Execution"]

Key Designs

1. Text-to-Multimodal Curriculum SFT Cold Start: Resolving Lack of Standardized Physical System Representation

Standard MLLMs frequently bypass system-level analysis in favor of premature formula matching. To enforce deliberate physical modeling, this paper defines a canonical five-component structure for intermediate trace \(\tau\): target quantity identification (System Understanding), explicit specification of known and unknown symbols (Variables), physical and geometric boundary conditions (Constraints), step-by-step decomposition of physical evolution (Physical Modeling Process), and governing mathematical equations (Equations). To prevent visual grounding complexities from interfering with the acquisition of this structured syntax, training follows a two-stage curriculum: models first undergo SFT on 12K high-level text-only physics problems sourced from SCP-116K and MegaScience to establish symbolic modeling habits, followed by tuning on 8,880 multimodal instances grounded with Qwen3-VL-32B image captions. This curriculum ensures that the model learns the structural logic of physical modeling before aligning it with multimodal visual features.

2. Physical Modeling Trace-Guided Symbolic Execution: Bridging the Gap Between System Formulation and Numerical Execution

Having established the system representation \(\tau\), the execution phase \(z\) must strictly ground every numerical step and algebraic substitution in the modeled constraints rather than drifting into ungrounded guessing. During data construction, candidate modeling traces generated via Best-of-N sampling (\(N=5\)) undergo a rigorous two-stage validation gate: first, Gemini-2.5-Flash screens for internal logical coherence, visual-textual consistency, and modeling specificity; second, the candidate trace is passed to a downstream executor model to confirm whether following the trace faithfully yields the correct final answer. At inference time, the model executes the problem by referencing the verified governing equations and boundary constraints identified in \(\tau\), effectively eliminating common algebraic failures such as sign errors, dimension mismatches, and ill-posed systems.

3. Atomic Fine-Grained Physics Rubric Derivation: Overcoming Sparse and Physics-Agnostic Process Supervision

Generic process reward models evaluate broad stylistic or chain-of-thought properties, failing to detect nuanced violations of physical principles. In this work, fine-grained rubrics \(\mathcal{C} = \{(d_k, w_k)\}_{k=1}^{|\mathcal{C}|}\) (comprising 8 to 12 criteria per instance) are automatically extracted from verified high-quality physical modeling traces. Each criterion \(c_k\) pairs an atomic, verifiable natural language requirement \(d_k\) with an adaptive importance weight \(w_k\), spanning four core dimensions: (1) Visual Criteria (diagram reading, geometry, annotations), (2) Physical Modeling Criteria (applicability of laws, conservation principles, sign conventions), (3) Strategic Planning Criteria (subproblem decomposition and state transitions), and (4) Execution Criteria (equation formulation and algebraic consistency). An LLM judge evaluates candidate responses against each criterion to yield binary satisfaction scores \(s_k \in \{0, 1\}\), producing a normalized rubric process reward:

\[R_{\text{rubric}}(q, (\tau, z, y)_i) = \frac{\sum_{k=1}^{|\mathcal{C}|} w_k s_k(q, (\tau, z, y)_i)}{\sum_{k=1}^{|\mathcal{C}|} w_k}\]

This formulation injects dense, physically grounded gradient signals into reinforcement learning.

4. Rubric-Guided Multi-Reward GRPO Policy Optimization: Preventing Unconstrained Policy Drift and Instability

Building on the SFT checkpoint, the model is refined using Group Relative Policy Optimization (GRPO) to explore and optimize physical reasoning trajectories. For a given question \(x\), the policy samples a group of \(G\) reasoning trajectories \(\{o_i\}_{i=1}^G\), where each trajectory encapsulates the decoupled modeling, execution, and solution \(o_i = (\tau_i, z_i, y_i)\). The composite reward function integrates outcome correctness \(R_{\text{outcome}} \in \{0, 1\}\), format adherence \(R_{\text{format}} \in \{0, 0.5\}\), and the dense process score \(R_{\text{rubric}}\):

\[R_{\text{total}} = R_{\text{outcome}} + R_{\text{format}} + R_{\text{rubric}}\]

Advantages \(\hat{A}_i\) are normalized across the group, guiding policy updates to favor trajectories that not only reach the correct answer but do so by constructing physically consistent system models. This effectively prunes ungrounded reasoning paths that happen to arrive at the correct answer through flawed physical assumptions.

Loss & Training

The framework executes in two stages: 1. Supervised Fine-Tuning: Minimizes the negative log-likelihood over full trajectories \(\mathcal{L}_{\text{SFT}} = -\mathbb{E}_{(x, \tau, z, y)\sim \mathcal{D}} [\log \pi_\theta(\tau, z, y \mid x)]\). Implemented with LLaMA-Factory using LoRA (rank 64, \(\alpha=128\), learning rate \(1 \times 10^{-4}\)), training spans 2 epochs on the 12K text dataset followed by 2 epochs on the 8,880 multimodal dataset. 2. GRPO Reinforcement Learning: Initializes the policy from the multimodal SFT checkpoint and trains on the multimodal physics pool using the verl framework (LoRA rank 64, \(\alpha=128\), learning rate \(5 \times 10^{-5}\), 1 epoch). The composite reward \(R_{\text{total}}\) computes relative group advantages \(\hat{A}_i\), driving simultaneous optimization of modeling validity and algebraic accuracy.

Key Experimental Results

Main Results

Evaluation spans three in-domain multimodal benchmarks (1,000 strictly held-out problems: 467 from PhysReason, 412 from PhyX, 121 from SeePhys) and two out-of-domain benchmarks: PhysUniBench (undergraduate-level physics) and MMK-12 (physics subset). Models are evaluated across numerical, symbolic, textual, and multiple-choice modalities using a hybrid rule-based and LLM-judge verification harness.

Model Backbone Method Variant PhysReason (P@1 / P@5) PhyX (P@1 / P@5) SeePhys (P@1 / P@5) PhysUniBench (P@1 / P@5) MMK-12phys (P@1 / P@5)
Qwen2.5-VL-3B-Instruct Baseline 13.3 / 33.6 10.9 / 23.5 9.1 / 24.7 17.1 / 50.9 35.0 / 73.4
Vanilla 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 Baseline 25.0 / 47.8 17.0 / 35.7 14.0 / 28.9 19.3 / 53.7 45.4 / 79.4
Vanilla GRPO 28.7 / 52.5 19.2 / 39.3 18.2 / 33.1 20.9 / 55.4 54.2 / 87.4
Physics-Modeling + GRPO 32.1 / 58.2 21.4 / 40.0 19.8 / 34.7 23.5 / 57.8 56.8 / 89.6
Qwen3-VL-8B-Instruct Baseline 51.8 / 70.7 43.2 / 61.4 27.3 / 40.5 32.4 / 60.2 67.6 / 84.8
Vanilla GRPO 54.0 / 73.9 46.1 / 64.1 30.6 / 45.5 34.9 / 63.4 73.8 / 90.0
Physics-Modeling + GRPO 53.3 / 74.7 45.1 / 64.8 32.2 / 47.9 36.6 / 65.5 74.4 / 91.2
Gemma3-4B-IT Baseline 7.3 / 17.1 7.5 / 16.3 5.0 / 16.3 17.8 / 48.9 27.0 / 61.8
Vanilla GRPO 10.9 / 19.3 9.5 / 17.0 9.9 / 18.2 21.6 / 49.9 32.6 / 64.0
Physics-Modeling + GRPO 13.1 / 22.7 11.2 / 18.7 11.6 / 19.0 22.4 / 51.1 38.0 / 69.4

Ablation Study

Component ablations on Qwen2.5-VL-7B-Instruct dissect the contributions of SFT cold start, rubric rewards, and the text-first curriculum (Pass@1 accuracy reported):

Configuration Variant PhysReason (PR) PhyX SeePhys (SP) PhysUniBench (PUB) MMK-12phys (MMK) Key Analysis
Full Model 32.1 21.4 19.8 23.5 56.8 Best performance across all benchmarks
w/o Rubric Reward 30.4 20.9 19.8 21.9 54.0 Drops 1.3% on average (-2.8% on MMK-12)
w/o SFT Cold Start 28.1 18.2 15.7 20.5 52.8 Drops 3.7% on average; severe drop across all tasks
w/o Textual SFT (Multimodal SFT only) 22.4 (P@1) / 43.8 (P@5) 15.1 / 29.9 14.0 / 25.6 — — PR score drops sharply from 27.2/54.4 to 22.4/43.8

Key Findings

  • SFT cold start forms the foundational reasoning scaffolding: Eliminating the SFT cold start reduces average Pass@1 by 3.7 percentage points. Furthermore, removing the initial text-only SFT phase causes performance on PhysReason to plummet from 27.2/54.4 to 22.4/43.8, proving that decoupling modeling patterns in symbolic text space is essential before multimodal integration.
  • Rubrics enforce physical realism and consistency: Removing rubric rewards lowers overall accuracy by 1.3 points on average, confirming that dense physics criteria prevent models from exploiting spurious reasoning shortcuts to reach correct numerical answers.
  • Drastic suppression of wasteful self-revision and improved token efficiency: Analyzing 500 generations from Qwen3-VL-8B-Instruct reveals that the baseline frequently encounters sign or conceptual mistakes mid-reasoning, triggering reflection keywords ("wait", "rethink", "verify") in 61.8% (309/500) of generations. In contrast, after SFT cold start, explicit reflection drops to 1.0% (5/500). At a constrained budget of 1k tokens, the SFT model dramatically outperforms the baseline by establishing the correct system formulation upfront.
  • Mechanics and Electromagnetism exhibit greatest gains: Breakdown by physical subfields shows performance improvements over baseline of +7.3% in Mechanics and +6.7% in Electrodynamics, compared to +4.2% in Optics and +3.5% in Thermodynamics. Problems in mechanics and fields depend fundamentally on identifying boundary interactions, coordinates, and force vectors, areas where structured modeling provides decisive advantages.

Highlights & Insights

  • Explicit decoupling of modeling and execution: Demonstrates that treating physics as a two-stage cognitive pipeline (system modeling followed by mathematical execution) resolves the root cause of MLLM physics failures far more effectively than scaling unstructured chain-of-thought tokens.
  • Automated extraction of grounded physics rubrics: Establishes an end-to-end data curation pipeline that converts validated physical trajectories into atomic, verifiable rubric criteria, enabling scalable process-supervised RL without manual annotation.
  • Elimination of unproductive reflection loops: Demonstrates that a structured mental model constructed upfront prevents mid-stream sign errors and restarts, rendering reasoning far more efficient under low-token budgets.

Limitations & Future Work

  • Reliance on advanced synthesis models: Generating candidate modeling traces and grading rubric satisfaction relies heavily on Gemini-2.5-Flash and Qwen3.5-Flash, potentially inheriting subtle systematic modeling biases from teacher models.
  • Marginal Pass@1 gains on larger backbones: On Qwen3-VL-8B-Instruct, the gains in Pass@1 are relatively muted compared to smaller 3B and 7B backbones, suggesting that larger models already possess partial implicit modeling capabilities.
  • Evaluation cost of rubric-based RL: Evaluating 8-12 rubric criteria per candidate trajectory via an LLM judge incurs significant computational overhead during reinforcement learning exploration.
  • vs PLAN-TUNING / Trajectory Planning (Parmar et al., 2025; Jiao et al., 2024): Generic planning techniques order actions at a meta-level without verifying physical admissibility; this work explicitly checks physical grounding, system boundaries, and governing conservation laws.
  • vs Rubrics as Rewards (Gunjal et al., 2026; Viswanathan et al., 2025): Standard rubric RL relies on broad rubrics for open-ended prose; this paper introduces physics-grounded atomic criteria verifying exact variable grounding and equation formulation.
  • vs Math-LLaVA / Mint-CoT (Shi et al., 2024; Chen et al., 2025): Mathematical reasoning methods focus heavily on algebraic manipulation and geometric OCR; this work shows that physics requires unique intermediate abstraction of system dynamics before algebra can succeed.

Rating

  • Novelty: ⭐⭐⭐⭐ [Decoupling physics reasoning into explicit physical modeling and execution, coupled with trace-derived atomic rubrics, is an intuitive and highly effective paradigm]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluated across 4 VLM architectures, 5 multimodal physics benchmarks, complete component ablations, subfield breakdowns, and token-budget analyses]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Clear motivation, rigorous mathematical formulation, and exceptionally thorough appendices detailing data pipelines and error analyses]
  • Value: ⭐⭐⭐⭐ [Provides a practical and reproducible blueprint for training small-to-medium multimodal models on complex scientific reasoning tasks]