State commitment learning: training language models to distinguish computation from memory

In three linesNew training method to distinguish temporary computation from persistent state in language models. Counterfactual Erasure RL (CERL) rewards models when answers remain correct after erasing intermediate thoughts. Evaluation on mathematics, logic, and scientific QA shows reduced dependence on hidden computations without accuracy loss.

Summary generated by Claude — human-verified