ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs
Summary
ReactHuman is a physics-grounded benchmark that evaluates multimodal large language models for human-like reactive decision-making in simulated humanoid robots facing household hazards, revealing persistent safety failures and offering a scalable diagnostic tool.
View Cached Full Text
Cached at: 09/15/26, 02:40 AM
Paper page - ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs
Source: https://huggingface.co/papers/2609.10895
Abstract
ReactHuman is a physics-grounded benchmark that evaluates multimodal large language models as decision cores for simulated humanoid robots facing sudden household hazards, revealing persistent reactive safety failures.
Reacting to sudden physical hazards (catching a slipping plate, dodging a falling knife) is both a meaningful test ofembodied intelligenceand a hard requirement for deployingmultimodal large language models(MLLMs) as the decision coreof household robots. Existing evaluations, however, probe intuitive physics passively through question answering over videos, or target deliberate, long-horizon tasks such as navigation and rearrangement; none measure whether a model can turn physical understanding into immediate,safety-critical action. We introduce ReactHuman, the firstphysics-grounded benchmarkfor human-likereactive decision-making, in which the evaluated MLLM acts as the brain of a simulated humanoid facing sudden household hazards; it spans 17 event families and over 1,000 bit-for-bit reproducible scenes with exact, annotation-free ground truth derived from 240 Hzrigid-body simulation, includingadversarial objectswhose appearance contradicts their physics (a foam anvil, a steel apple). We further design a five-metric suite that scores each reaction along three axes: reasonable, safe, and physically grounded. We physically execute every committed plan so that decisions have observable consequences. With this harness we evaluate seven representative MLLMs. Results show that reactive safety is far from solved: models mishandle roughly one hazard in three, act from fixed dispositions rather than the observed scene, trust appearance over motion, and miss interception points at meter scale even when the chosen action is correct; none of these failures shrink with model scale. ReactHuman thus offers both a fine-grained diagnosis and a scalable training signal toward physically grounded, safety-aware embodied agents. The benchmark can be found here: https://huggingface.co/datasets/Alan123/reacthuman-benchmark-scaled
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2609\.10895
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.10895 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.10895 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.10895 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Words to World: Benchmarking LLMs in Interactive Physical Environments
该论文提出了一个用于评估大语言模型在闭环物理交互中表现的基准,模型需通过连接机械臂、摄像头和传感器的系统在真实环境中执行任务并根据物理反馈调整行动。实验揭示了LLM在物理行动中的常见失败模式,如幻觉状态变化、糟糕的异常恢复和不稳定的长程规划。
Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs
This paper introduces PatternEval, a diagnostic benchmark for evaluating response-pattern failures in hybrid-thinking multimodal large language models, and proposes PatternRL for aligning these patterns through reinforcement learning with specific penalties.
H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models
H2R-Bench is a new benchmark that evaluates video world models on transforming human manipulation videos into robot-centric demonstrations, testing embodiment constraints and interaction fidelity across six manipulation families and two robot embodiments.
Mr.LHDR: A Benchmark for Multimodal Real-World Long-Horizon Deep Research Agents
The paper introduces Mr.LHDR, a benchmark for evaluating multimodal real-world long-horizon deep research agents, showing that current models struggle with dependency-consistent evidence integration in complex reasoning chains.
DailySafe: Benchmarking Dynamic Safety Awareness in Everyday Multi-Turn Multimodal Interactions
DailySafe is a new benchmark of 195 naturalistic multi-turn, multi-image interactions across 10 everyday domains for evaluating Dynamic Safety Awareness in multimodal large language models, testing whether they can track evolving risks revealed across conversational turns and visual observations.