A unified framework is introduced 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.
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.
Video-language models (VLMs) have achieved remarkable performance on video understanding and visual question answering, yet they remain unreliable in reasoning about physical plausibility, where understanding object interactions, causal dynamics, and fundamental physical principles is essential. This limitation is particularly evident on challenging physical reasoning benchmarks, revealing a persistent gap in physical commonsense reasoning. To address this challenge, we propose PhysMRV, a training-free physical memory and verification framework for physical plausibility reasoning. Unlike retrieval-augmented VLMs that retrieve semantically similar videos as additional context, PhysMRV transforms training videos into a Hierarchical Memory Bank of structured physical knowledge comprising three complementary levels: scene descriptions capturing visual context, physical-event graphs modeling object interactions and causal structure, and physics-rule summaries distilling reusable physical principles and cues. During inference, PhysMRV retrieves physically relevant memories and leverages their structured physical evidence to guide a frozen VLM in verifying physical plausibility, requiring neither fine-tuning nor parameter updates. We evaluate PhysMRV on three challenging physical reasoning benchmarks, ImplausiBench, IntPhys2, and GRASP Level 2, across multiple state-of-the-art VLMs. Experimental results demonstrate consistent improvements over direct prompting across diverse VLMs and evaluation benchmarks, showing that structured physical memories provide an effective and scalable means of enhancing physical plausibility reasoning without additional training.
Wenyuan Wang, Lianyu Hu, Hao Wang et al.· 0 citations
Physics problems in textbooks are typically presented as static diagrams accompanied by brief textual descriptions, requiring learners to infer dynamic physical behaviors through mental visualization. This process often imposes high cognitive demands and limits learners'ability to form accurate mental models. In this paper, we present \textbf{LivePhys}, a framework that enables a \emph{Scan-to-Play} paradigm for mechanics learning by transforming static textbook physics problems into executable, interactive simulations. LivePhys decouples multimodal perception from physics-aware reasoning and deterministic simulation. Given a problem diagram and its accompanying text, LivePhys performs text extraction, geometric segmentation, and cross-modal grounding to construct a structured, physics-aware intermediate representation. A multimodal large language model is then used as a reasoning controller to infer entities, parameters, and constraints, which are executed by a physics engine to generate spatially consistent and interactive simulations that allow learners to explore and manipulate problem conditions dynamically. Our evaluation results show that LivePhys significantly outperforms general-purpose multimodal models in simulation executability, spatial accuracy, and interaction fidelity. In addition, a user study demonstrates that interacting with LivePhys-generated simulations reduces learners'perceived cognitive load compared to static textbook materials.
Xiaowei Dai, Ziyu Luo, Xiangwen Zhang et al.· 0 citations
The demand for high-fidelity, physically-based animation has traditionally been met by sophisticated solvers rooted in elastodynamics and finite element analysis (FEM). Recently, the emergence of neural physics has led to a paradigm shift, transforming neural networks into solvers with memory that dramatically increase the scale and speed of digital environments. Despite its reputation, physics-based simulation does not have to be intimidating. This course aims to demystify the field, proving that these complex systems are accessible and intuitive when approached correctly. We provide a unified journey from classical formulations to modern neural techniques, grounding the audience in the fundamentals of elastostatics and dynamics. We demonstrate how physical problems are discretized via linear finite elements and solved through the elegant lens of optimization. Transitioning into neural physics, we showcase how traditional simulation knowledge translates directly into machine learning loss functions and neural architectures. We analyze strategies for modeling latent spaces for a system’s equilibrium states and to create truly controllable, real-time frameworks. Designed for a broad audience—including students, engineers, researchers, and artists—this course balances theory with practice. To ensure these concepts are immediately actionable, we provide comprehensive reference code for all discussed methods. By the end of the session, attendees will possess the tools to quickly and easily implement their own physics solvers, empowering them to build the next generation of physics-enhanced frameworks and interactive worlds.
D. Corigliano, Otman Benchekroun, J. Barbič· Proceedings of the Special I...· 0 citations
The results show that state-of-the-art LLMs can solve much of the standard introductory physics problems, but that their performance remains uneven and constrained by problem modality and problem difficulty.
IMBENCH is introduced, a benchmark designed to evaluate intuitive manipulation as an integrated capability spanning perception, physical reasoning, action generation, and iterative execution, and position IMBENCH as a step toward evaluating and enabling more integrated, adaptive physical intelligence.
Anurag Maurya, Sukhvansh Jain, Prajwal Avhad et al.· 0 citations
Large reasoning models trained via reinforcement learning (RL) have been increasingly shown to outperform their supervised fine-tuned (SFT) counterparts on mathematical reasoning tasks; Yet the mechanistic basis for this advantage remains unclear. We therefore ask, what internal representational differences enable RL models'superior performance? Our work presents two converging lines of evidence: First, linear probes trained on layer-wise hidden states reveal that RL models tend to achieve higher accuracy in predicting answer correctness compared to SFT models, indicating more linearly separable and structured representations. Second, mean ablation studies show that RL models develop a hierarchical architecture where deeper layers become progressively more critical, whereas SFT models distribute importance uniformly across layers. Together, these findings demonstrate that RL training fundamentally restructures how models represent and process reasoning problems. Finally, we analyze token-count variability under repeated sampling across problems to assess adaptive compute allocation. While we observe higher variability in some RL-tuned models than in their SFT counterparts, we see strong consistency in others, suggesting that token allocation may depend more on the overall training pipeline than on RL versus SFT alone. We believe this token-allocation variability reveals the spread of plausible on-policy reasoning, highlighting which models exhibit stable policies versus those that are under-determined, potentially non-identifiable solution behaviour.
Antyabha Rahman, Akshaj Gurugubelli, Omar Ankit et al.· 0 citations