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

Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

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 DTO, a dynamic thinking-pattern optimization framework that segments LRM reasoning traces, prunes redundant segments and finalizes early, then applies SimPO on the optimized vs. unoptimized pairs to cut overthinking on math benchmarks.

Problem

Large reasoning models overthink, producing unnecessarily long and complex reasoning paths that waste computation and can hurt accuracy. The paper asks how to select beneficial reasoning segments (thinking patterns) at the right points to improve efficiency without losing accuracy.

Approach

Segments reasoning traces into thinking patterns, uses Monte Carlo exit probes to find the earliest point where the answer is reliably correct, inserts a finalize pattern, and uses an auxiliary LLM plus validation decoding to prune redundant segments. The optimized trajectories form preference pairs for SimPO, targeting segment-level rather than length-level efficiency.

Key results

• On DeepSeek-R1-Distill-Qwen-1.5B, DTO reaches 85.48% on MATH with 1936 tokens versus 79.80% and 3543 tokens for the baseline (efficiency 1.960).
• Per-question analysis reports attention FLOPs reduced by 47% for originally correct responses on DeepSeek-R1-Distill-Qwen-1.5B, and accuracy on originally incorrect responses improved by 15.6%.
• On AMC2023 with DeepSeek-R1-Distill-Qwen-1.5B, DTO reaches 70.25% versus 58.25% baseline with 3377 vs 5339 tokens.

Measured effects reported in the paper

Surprising findings

• Optimized trajectories transform a non-trivial share of initially incorrect responses into correct ones, even when no correct sample exists for the problem.
• On DeepScaleR-1.5B-Preview the gains are narrower, and on GSM8K DTO's accuracy is slightly below baseline for DeepSeek-R1-Distill-Qwen-1.5B (83.91 vs 82.13 is actually higher; DTO is lower than FCS+Ref. on that set).

Limitations

• Validation is limited to math and one MMLU-Pro subset with 100 questions per domain; open-ended tasks are not evaluated.
• Requires multiple sampled responses per problem and Monte Carlo exit probing, adding optimization-phase cost.
• Relies on an auxiliary 70B LLM to judge segment utility, so the pruning depends on that judge's quality.

Experimental setup

Datasets: MATH training set (5,000 samples) MATH GSM8K Gaokao AMC2023 AIME2024 AIME2025 MMLU-Pro

Benchmarks: MATH GSM8K Gaokao AMC2023 AIME2024 AIME2025 MMLU-Pro

Metrics: accuracy average tokens efficiency score eta attention FLOPs

Baselines: Fast Prompt SFT O1-Pruner DAST FCS + Ref. Instruct version

Builds on: SimPO O1-Pruner DAST DeepScaleR

Closest papers by meaning

• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.92)
• Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.92)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.92)
• Rethinking Thinking Tokens: LLMs as Improvement Operators (ICML 2026 · similarity 0.92)
• Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.91)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.91)