On Language Drift during RLVR Post-Training

arXiv:2610.02015v1 Announce Type: cross 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.

Related

• Unlearnable, or Unmeasured? On the Reliability of Difficulty Labels in RLVR
• Quantifying Empirical Compute-Supervision Tradeoffs in RLVR
• Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models
• How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients
• LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training

Source: arXiv cs.AI | 2026-10-02