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

OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping

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.

Frames chain-of-thought early stopping as an optimal stopping problem and trains a lightweight plug-in stopping policy (last two layers plus linear head) on a frozen reasoning model to trade accuracy against length via a scalar lambda.

Problem

Reasoning models overthink, generating redundant CoT tokens that add latency and cost without accuracy gains; existing early-exit methods use fixed thresholds on classifiers, which are suboptimal for deciding when to stop.

Approach

Model CoT stopping as an optimal stopping problem with per-prefix reward accuracy minus lambda times length, computed exactly by forcing answers after each step; train a stopping policy (last 2 transformer layers plus linear head) with the base model frozen, and sweep lambda for a accuracy-length frontier. A theorem shows threshold classifiers can lose arbitrarily large value relative to the optim

Key results

• On DeepSeek-R1-Distill-Qwen-7B, OS-Pruner reduces length by 59.3% on GSM8K (88.0 vs 88.7 acc) and 52.8% on MATH-500 with +2.7 accuracy points at best-AES points.
• Average AES of 0.40 on DeepSeek-R1-Distill-Qwen-7B versus 0.29 for the best baseline (FlashThink).
• On AIME with DeepSeek-R1-Distill-Qwen-7B, length is reduced by only 6.9% at the best-AES point, versus larger reductions on easier datasets.

Surprising findings

• On DRPO-7B, OS-Pruner's AIME accuracy drops from 44.4 to 36.9 at its best-AES operating point, a notable accuracy loss despite a high AES.
• Forcing final answers without the boxed prefix leads to models thinking in the answer section, weakening accuracy-length control on DeepSeek-R1-Distill-Qwen-7B and DRPO-7B.

Limitations

• Evaluated only on math reasoning tasks; coding and scientific reasoning untested.
• Scaling beyond 20B parameters not explored.
• Accuracy control weaker on models with poor instruction following when final answer is forced (DeepSeek-R1-Distill-Qwen-7B, DRPO-7B).

Experimental setup

Datasets: DeepScaleR-Preview-Dataset

Benchmarks: GSM8K MATH-500 AIME24/25

Metrics: Pass@1 average generated reasoning tokens Accuracy Efficiency Score (AES)

Baselines: Classification Probe Answer Convergence FlashThink DRPO-7B

Builds on: DRPO FlashThink Answer Convergence vLLM Math-Verify

Closest papers by meaning

• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.91)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.90)
• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression (ICML 2026 · similarity 0.90)
• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.90)
• Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.90)