On Language Drift during RLVR Post-Training
The paper studies why LLM reasoning models trained with RLVR develop weird or nonstandard chain-of-thought language, arguing that the drift is not just a quirk but a consequence of the optimization setup. Its main contribution is tying theory and experiments together to show that RLVR can incentivize unbounded drift on genuinely novel reasoning tasks while SFT does not, and that preventing this drift may inherently cost reward/performance, though the claim’s strength depends on how broadly their formal assumptions and empirical settings carry over to real frontier training.
Abstract
Recent advances in LLM reasoning models---driven primarily by the paradigm of post-training via reinforcement learning with verifiable reward (RLVR)---have enabled them to accomplish impressively complex tasks. However, in parallel with their rising capabilities, LLMs have increasingly displayed signs of language drift in their chains of thought (CoTs): unusual, non-standard, and seemingly nonsensical language use. Although it is well-documented---and can potentially impair CoT monitorability---the causes of language drift are thus far poorly understood. In this paper, we identify the conditions under which language drift occurs: we prove theoretically that RLVR optimization pressure permits unbounded language drift, while supervised fine-tuning does not. We then show empirically that language drift specifically arises during RLVR on novel reasoning tasks---i.e. when the target behavior cannot be drawn out of the base model. Finally, we prove that it is not possible to constrain language drift without constraining expected reward, suggesting that CoT monitorability cannot be improved without harming performance during RLVR post-training at the frontier.