physical-reasoning

This paper evaluates frozen video representations and finds that latent prediction errors in video world models reflect appearance and motion rather than reliably measuring physical surprise, presenting a framework that separates representation content from error signals across IntPhys2, Physion++, and Kubric benchmarks.

Psi-0.6 is a unified vision-language world model that adaptively composes perception, language, geometry, and simulation via 'streams of thought', achieving zero-shot gains on multi-view spatial reasoning and supporting counterfactual simulation optimized with RL from task feedback.

PhysEvolve 提出一个验证器驱动的自演化框架，利用物理引擎验证模型自生成的可执行世界与反事实仿真，以 GRPO 训练视觉语言模型的因果物理推理能力，在 PhysBench、CausalPhys 和 ContPhy 上取得显著提升且无需人类干预数据。

Google DeepMind 发布 Gemini 4 Argon，一款面向机器人全栈的通用物理推理模型，在 ERQA 基准上以 87% 领先，可支持机器人控制、Real2Sim、物理理解与自改进等工作流。

Axiom AI engineers linked OpenAI's GPT-6 Astra to a 2024 Toyota Corolla's cameras and power steering, letting the general-purpose language model drive to an In-N-Out drive-thru, highlighting growing (if rudimentary) physical-world understanding in LLMs. The article also covers related efforts like Elorian AI and Scale AI's Humanity's Sixth Sense benchmark for physical scene understanding.

该论文提出 LMBuild，一个用于评估 LLM 智能体生成可建造、功能性结构能力的基准，包含交互式装配环境、基于 CAD 数据集构建的评测集以及涵盖结构可靠性、功能可供性和物理可实现性的评估框架，在 30 个系统上的评测揭示了功能推理仍是前沿模型的主要瓶颈。

This paper introduces the World Embedding Benchmark, with 8,000 controlled simulation cases spanning fluid/solid mechanics, dynamics, and optics, to evaluate how video representations encode physical information. It finds pretrained multimodal embeddings struggle with physical alignment, probes recover physics from frozen embeddings, contrastive training trades off alignment vs. recoverability, and retrieved reference videos improve physical fidelity of MiniMax-H3-generated videos.

PhysVista 提出了一个基于感知-推理-评估闭环认知框架的基准，用于评测视觉语言模型的物理智能，涵盖事件级与尺度级推理以及真实和AI生成视频，并揭示了VLM在物理推理和合理性评估上的显著局限。

The paper introduces Physis-Lang, a self-evolving framework that uses physical language as a shared, optimizable representation across data curation, training, and video generation, alongside the PhysCapBench benchmark and an agentic loop for refining physical captions. With open-source Cosmos-Nano backbones, the enhanced models surpass proprietary Veo3.1 on physical plausibility benchmarks.

CausalWM is a causal world model that uses Chain-of-Thought to predict future frames by capturing physical dependencies step-by-step, ranking top on benchmarks like TriWorldBench and PAI-Bench.

Principia is a benchmark that evaluates video models on Newtonian physics using relational consistency between paired objects, revealing significant gaps in current models' physical reasoning.

Code-as-World has been open-sourced, introducing a framework that uses code as executable representations for the physical world to enhance AI physical reasoning, along with pre-trained models and evaluation tools.

Code-as-World introduces a paradigm that represents physical environments as executable code to enable quantitative reasoning for vision-language models, using an agentic discovery loop for scalable supervision and achieving state-of-the-art performance.

This paper proposes Answer-Level Trust Selection (ATS), a post-hoc, model-agnostic framework for assessing the reliability of individual predictions from vision-language models in quantitative physical reasoning tasks.

VAORA proposes a novel reward design for vision-language models that aligns reasoning with visual contexts and action outcomes, improving physical reasoning and generalization across unseen tasks and environments.

Introduces EVLA, a framework that enhances vision-language driving assistants with real-time awareness of electrified powertrain states, enabling energy-optimal and physically grounded decisions.

This paper identifies a failure mode called PhysHack in LLM-based LEGO assembly generation and proposes PVPO, a sample-efficient reinforcement learning method with model-based data selection that improves physical and semantic alignment using only a small fraction of training data.

WorldOlympiad presents a comprehensive benchmark for evaluating video-based world models across physical faithfulness, geometric consistency, and interaction fidelity, revealing significant gaps in current generative models.

Dream.exe proposes an evaluation framework that uses robotic manipulation tasks to assess video generation models' understanding of physical reality, finding that visual quality does not predict executable motion accuracy.

BilliardPhys-Bench is a new benchmark that tests multimodal LLMs on physical reasoning using synthetic billiards scenarios, requiring predictions of collisions and final ball positions. The paper finds that current models struggle with longer simulations and exhibit a 'stasis bias' of predicting no interaction when uncertain.