Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning

2026-04-20

AI Summary

This work addresses the inefficiency of large language models (LLMs) in reasoning tasks, where they often expend excessive computational resources generating lengthy and erroneous responses due to the absence of effective early termination mechanisms. The paper introduces the first formal theoretical framework for dynamic abstention, modeling it as an explicit action within regularized reinforcement learning: at each generation step, the model decides whether to terminate early by comparing an approximate value function against a predefined abstention reward threshold. The authors rigorously prove that this strategy outperforms baseline approaches under general conditions and propose an efficient algorithm for approximating the value function. Empirical results demonstrate that the method significantly improves selective accuracy on mathematical reasoning and toxicity avoidance tasks while achieving a superior trade-off between computational cost and informational gain.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

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.

Problem

• dynamic abstention
• LLM reasoning
• compute efficiency
• selective accuracy
• value function

Innovation

• dynamic abstention
• value function
• reinforcement learning
• chain-of-thought reasoning
• selective accuracy