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

Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models

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.

Proposes ASAG, a training-free early-stopping method for large reasoning models that combines answer confidence with attention-entropy dynamics at action transition points, plus logits injection and jump prompts to escape thinking traps.

Problem

Large reasoning models overthink, producing redundant chain-of-thought tokens that raise cost and sometimes reduce accuracy; existing training-free early-exit methods rely on unreliable confidence signals that cause premature or delayed exits.

Approach

Uses the normalized attention entropy of the last four layers, tracked across action transition points, as a complementary convergence signal alongside probed answer confidence; when entropy drops and confidence is high it exits, when entropy is high it keeps reasoning, and when converged but low-confidence it boosts the converged answer's logits; a jump prompt redirects when attention indicates a

Key results

• On Qwen3-8B, ASAG improves average accuracy by 3.2% (72.5 to 75.1) while reducing generated tokens to 64.2% of vanilla (about 36% reduction) across the five main benchmarks.
• On DeepSeek-R1-Distill-Qwen-7B, ASAG attains the largest average gain: 62.3 vs 58.2 accuracy with a 64.7% compression rate.
• Under sampling (T=0.6), ASAG on Qwen3-4B raises average accuracy from 71.5 to 74.0 with 37.8% fewer tokens.

Surprising findings

• Attention entropy at correct intermediate answers drops sharply, with over 70% of later ΔH values below -0.1, while the model's token confidence alone is an unreliable exit signal.
• On challenging problems, the confidence-only baseline DEER often exits prematurely on cases where vanilla decoding would have been correct.

Limitations

• Relies on greedy decoding in main results; sampling results only reported for a subset of models and benchmarks.
• Hyperparameters (lambda, alpha, jump attempts, logits mixing weight) were chosen empirically on the evaluated benchmarks.
• Theoretical lemma on entropy-stability bound is loosely connected to the empirical exit rule.

Experimental setup

Datasets: DAPO-MATH-17K

Benchmarks: GSM8K MATH-500 AMC 2023 AIME 2024 AIME 2025 OlympiadBench GPQA Diamond HumanEval LiveCodeBench

Metrics: pass@1 accuracy average generated tokens compression rate

Baselines: Vanilla NoThinking TALE Dynasor DEER

Builds on: DEER Dynasor TALE NoThinking Speculative Rejection

Closest papers by meaning

• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.90)
• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.90)
• LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.90)
• Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.90)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.90)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)