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

SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model 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.

Proposes SmartThinker, a GRPO-based efficient-reasoning method that estimates a per-prompt optimal CoT length via a Gaussian model of correct vs. all response lengths and dynamically scales the length-reward coefficient so correct trajectories keep non-negative advantages.

Problem

Static GRPO length rewards for large reasoning models use fixed targets and fixed coefficients regardless of problem difficulty, causing over-compression of correct long reasoning and accuracy loss in overthinking LRMs.

Approach

Model response lengths of all and correct trajectories as Gaussians to derive a closed-form per-prompt optimal length maximizing P(correct | length), penalize only correct trajectories exceeding it, and set the length-reward coefficient dynamically from the in-group error ratio so correct trajectories' advantages stay non-negative.

Key results

• On DeepSeek-R1-Distill-Qwen-1.5B, SmartThinker reaches 61.9% average accuracy with 5169 average tokens (vs. base 60.7% and 9980 tokens) using 150 RL steps.
• On DeepSeek-R1-Distill-Qwen-7B, SmartThinker achieves 74.5% average accuracy vs. 73.1% base, with AIME25 accuracy rising from 35.0% to 40.8%.
• On Qwen3-4B-Thinking-2507, average length drops from 13040 to 7747 tokens (about 40.6% reduction) while accuracy rises from 88.5% to 89.0%, with AIME25 accuracy 69.2% to 71.7%.

Surprising findings

• Using the optimal-length target with a symmetric |l - l_opt| penalty gives weaker results than penalizing only overlong correct responses, suggesting that shortening only excess length is better.
• The Gaussian assumption is idealized, but the authors report that over half of length distributions pass a KS test with shape tolerance, and that the optimal length is not monotonic in pass rate.

Limitations

• Only outcome-based rewards; no process reward supervision.
• Evaluated only with GRPO; applicability to other GRPO variants and open-ended tasks is untested.
• Gaussian distribution assumption is idealized, and a fraction of length distributions fail the Gaussian tests.

Experimental setup

Datasets: DeepScaleR-Preview Olympiad (length distribution test)

Benchmarks: MATH500 AIME25 AMC23 MMLU MathQA LiveCodeBench HumanEval

Metrics: accuracy (Pass@1) average output length AE score

Baselines: Base Model Truncated Think ShorterBetter ThinkPrune-4k LASER-DE-4096 AutoThink

Builds on: GRPO ShorterBetter DeepScaleR ThinkPrune LASER

Closest papers by meaning

• LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models (NeurIPS 2026 · similarity 0.92)
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning (NeurIPS 2025 · similarity 0.91)
• VeriThinker: Learning to Verify Makes Reasoning Model Efficient (NeurIPS 2025 · similarity 0.90)
• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression (ICML 2026 · similarity 0.90)
• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.90)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.90)