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

Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 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 RPDI-EE, a training-free early-exit method that monitors the ratio of local to global average token entropy during long CoT reasoning and inserts an end-of-thinking marker when the ratio exceeds a threshold.

Problem

Large reasoning language models overthink, generating redundant reasoning that degrades accuracy and wastes compute; existing early-exit methods need proxy models or answer probing, which add overhead or cause over-truncation.

Approach

Uses the average token entropy in a sliding window (LTF) relative to the global average entropy (GTF) as a Reasoning Path Deviation Index, treating entropy as a soft count of high-entropy transition tokens; exits when RPDI exceeds a threshold, checked only at boundary symbols, with no probing or proxy model.

Key results

• RPDI-EE improves average accuracy over Vanilla CoT on all eight tested LRLMs, with an average gain of 3.9% and 5.1% on distillation models per the paper's text.
• On DeepSeek-R1-Distill-Qwen-32B, RPDI-EE reaches 75.5 average accuracy versus 67.6 for vanilla CoT and 72.5 for DEER.
• On DeepSeek-R1-Distill-Qwen-7B, RPDI-EE gets 61.8 average accuracy versus 56.2 for vanilla CoT, but its average token length (7186) is only modestly below vanilla (7558) and above DEER's 6378 on the average.

Surprising findings

• RPDI-EE reduces accuracy on GPQA-Diamond for DeepSeek-R1-Distill-Qwen-7B (24.8 vs 32.3) and on Qwen3-235B (80.3 vs 80.8), so gains are not uniform.
• Dynasor-CoT and NoThinking cut accuracy substantially on several models despite shorter outputs.

Limitations

• Gains in token length reduction are modest compared with aggressive early-exit baselines.
• Hyperparameters (window W=512, threshold lambda=2.0) are tuned on the evaluation setup, and the paper notes sensitivity to them.
• Experiments are limited to math and science benchmarks with greedy decoding; baselines were only run on a subset of models due to compute cost.

Experimental setup

Datasets: GSM8K AMC23 MATH500 AIME2024 AIME2025 OlympiadBench GPQA-Diamond

Benchmarks: GSM8K AMC23 MATH500 AIME2024 AIME2025 OlympiadBench GPQA-Diamond

Metrics: accuracy generated token length

Baselines: Vanilla CoT NoThinking ThinkLess Dynasor-CoT DEER

Builds on: DEER Dynasor-CoT NoThinking ThinkLess

Closest papers by meaning

• Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.92)
• Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.91)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.91)
• Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.91)
• Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.91)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.90)