DigestAI news desk
Research updated

PhysMent Benchmark Tests LLMs on Interactive Physics Reasoning

Researchers have introduced PhysMent, a new benchmark designed to evaluate how large language models reason about the physical world through active experimentation. Unlike traditional static tests that provide all necessary data upfront, PhysMent requires models to interact with a MuJoCo physics simulator. Agents must apply forces, query object states, and modify scene geometry to discover…

1 source primary source

Key points

  • PhysMent evaluates LLMs via iterative interaction with a MuJoCo physics simulator rather than static data.
  • Models achieve up to 80% on qualitative tasks but fall below 30% on hard quantitative problems.
  • Failures stem from procedural errors like inefficient exploration, not conceptual misunderstandings.

The benchmark consists of 105 classical mechanics scenes organized by difficulty and modality, including tasks with hidden objects or required object creation. A six-dimensional scoring framework assesses performance across these scenarios. Results indicate that while current models achieve up to 80% accuracy on qualitative, single-concept tasks, they struggle significantly with quantitative problems. Most models drop below 30% accuracy on the hardest categories, where the challenge lies in executing precise experimental procedures.

Across seven evaluated models, accuracy ranged from 25% to 67%. The primary failure modes were not conceptual gaps but procedural issues, such as premature answer submission, inefficient exploration strategies, and inconsistent grounding in simulator feedback. This highlights a significant gap in current AI capabilities regarding dynamic, interactive physical reasoning.

Read the original at arXiv cs.CL · by Joseph Chan, Utkarsh Jha, Xiyin Yang, Abhinav Jarajapu, Anik Sahai, Eddie Hu, Robin Jeshua Deepak, Stefano Saravalle, Aditya Shah primary source Open source ↗

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories