Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Formalizes mid-generation abstention as a KL-regularized RL action and proposes dynamic value-thresholding, abstaining when an MLP probe's estimated value of the current prefix falls below a fallback reward.

Problem

LLM chain-of-thought reasoning wastes compute generating long incorrect responses; existing abstention methods decide before or after generation or at fixed positions, and lack principled rules for when to abstain mid-generation.

Approach

Model abstention as a terminal action with fallback reward r_bot in a KL-regularized MDP; prove that abstaining when the value function falls below r_bot dominates no-abstention and fixed-position rules; estimate the value with a probe trained by binary cross-entropy at all token positions of completed traces.

Key results

• On OlympiadBench at 90% abstention, Phi-3 dynamic abstention reaches 0.64 selective accuracy vs 0.34 for best input-processing baseline.
• On OlympiadBench at 90% abstention, Qwen dynamic abstention reaches 0.91 selective accuracy vs 0.75 for best fixed-position baseline.
• On GSM8K dynamic abstention reaches 0.92 selective accuracy at 10% abstention for both models.

Surprising findings

• Dynamic abstention retains only 63-86% of input-only token savings at 10% abstention rates.

Limitations

• Theoretical guarantees assume an oracle value function and, for strict optimality, zero KL regularization and an optimal base policy.
• Fallback reward r_bot is fixed rather than state-dependent.
• Token savings are lower than input-only methods at low-to-mid abstention rates.

Experimental setup

Datasets: GSM8K OlympiadBench RealToxicityPrompts

Benchmarks: GSM8K OlympiadBench RealToxicityPrompts

Metrics: selective accuracy abstention rate estimated reward token savings non-toxic response rate

Baselines: Constant Step Probe Self-Assessment LoRA Abstention no abstention

Builds on: KL-regularized RL selective classification Constant Step Probe P(True)

Closest papers by meaning

• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.91)
• Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.90)
• $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.90)
• AbstRaL: Augmenting LLMs' Reasoning by Reinforcing Abstract Thinking (ICLR 2026 · similarity 0.90)
• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.90)
• In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback (AAAI 2026 · similarity 0.89)