Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning
For LLM practitioners, this provides a theoretically grounded method to reduce wasted compute from incorrect reasoning chains, outperforming existing empirical abstention approaches.
The paper introduces a principled framework for dynamic abstention in LLM reasoning, where the model decides to stop generating at each token position. By modeling abstention as an action in a regularized RL framework, they derive a rule to abstain when the value function falls below a reward parameter, achieving improved selective accuracy on math reasoning and toxicity avoidance tasks.
Large language models (LLMs) using chain-of-thought reasoning often waste substantial compute by producing long, incorrect responses. Abstention can mitigate this by withholding outputs unlikely to be correct. While most abstention methods decide to withhold outputs before or after generation, dynamic mid-generation abstention considers early termination of unpromising reasoning traces at each token position. Prior work has explored empirical variants of this idea, but principled guidance for the abstention rule remains lacking. We present a formal analysis of dynamic abstention for LLMs, modeling abstention as an explicit action within a regularized reinforcement learning framework. An abstention reward parameter controls the trade-off between compute and information. We show that abstaining when the value function falls below this reward strictly outperforms natural baselines under general conditions. We further derive a principled and efficient method to approximate the value function. Empirical results on mathematical reasoning and toxicity avoidance tasks support our theory and demonstrate improved selective accuracy over existing methods.