RLAWM defines 'modeling drift' and penalizes divergence from physical axioms to fix LLM physics reasoning
Xinye Yang · hf · 2026-10-22
The paper identifies a failure mode in research-level LLM physics reasoning: models silently alter given quantities, fabricate laws, or inject unstated assumptions to force problems into familiar patterns. The authors define this as modeling drift — the gap between a model's internal axiomatic world model and the true physical world in the problem.
- RLVR and PRM reward shapes don't directly supervise world modeling in axiomatic domains like physics
- RLAWM improves physics reasoning by penalizing divergence between the model's internal world model and true physical axioms, via hierarchical modeling and solving rewards
- From Northwestern and University of Florida researchers
More from Research
- Turing-complete malware found runnable in PyTorch's backward pass — cephaloform · 2026-10-22
- One researcher pretrains coding model Mirendil-Base in 3.5 months for under $10M — bneyshabur · 2026-10-22
- Nature paper converts subword LLMs to byte-level with under 1% of pretraining cost, yielding Bolmo — Justgototheeffinmoon · 2026-10-22
- Tsinghua paper: letting agents manage their own KV cache cuts cost by 27.7% — dair_ai · 2026-10-22
- Effective receptive field in DiTs mirrors 2017 CNN finding, CS professor notes — CSProfKGD · 2026-10-22
- Flint cuts bike-demand validation RMSLE 30.8% in 3 hours of autonomous model iteration — cong_ml · 2026-10-22