RLAWM penalizes physics modeling drift during LLM training
Original title: Axiomatic World Modeling for Physics Reasoning
Authors report gains of up to 12.2 points on PhysReason and 5.2 on PHYSICS, with no added inference cost.
The paper defines modeling drift as an LLM silently changing quantities, inventing laws, or adding assumptions before solving a physics problem. Reinforcement Learning with Axiomatic World Modeling adds separate training rewards for capturing the stated physical world and for solving consistently with it. The method is applied only during training. Authors also claim zero-shot gains on olympiad-level physics and math.
Unconfirmed: Benchmark gains are reported by the paper's authors and are not independently confirmed.
- Modeling and solving rewards are scored separately during training.
- Authors report a 2.45x gain and 10 extra points on olympiad physics and math.
- The framework adds no extra cost at inference time.
From the original article
1 min read at huggingface.co
Recent research-level physics-reasoning benchmarks for LLMs reveal a failure mode that lies not in knowledge or execution, but in modeling the world being reasoned about. Models often silently alter given quantities, fabricate laws, or inject unstated assumptions to force a problem into familiar patterns, and then reason inside this corrupted world.
By Xinye Yang, Zhenyang Liu, Yuxuan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.