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

SuCo: Sufficiency-guided Continuous Adaptive Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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.

Introduces Minimal Sufficient CoT (MSC), the shortest reasoning prefix that yields the ground-truth answer, and SuCo, a two-stage SFT-then-GRPO framework (MFT + SAPO) for continuous, problem-adaptive reasoning length control in LRMs.

Problem

Large reasoning models generate excessively long chain-of-thought even for simple queries, and existing adaptive methods rely on discrete modes or fixed budget tiers without a principled criterion for when reasoning is sufficient.

Approach

Defines reasoning sufficiency as the per-token geometric mean probability of the ground-truth answer given a CoT prefix, selects the shortest sentence-level prefix exceeding a problem-adaptive threshold (scaled by length percentile), trains on these MSC traces via SFT, then uses GRPO with a sufficiency-aware over/under-thinking reward and an EMA-tracked complexity pool.

Key results

• At 7B scale SuCo averages 72.1% across eight benchmarks vs 63.2% for DeepSeek-R1-Distill-Qwen, while using about 1,267 average tokens vs 5,239.
• At 1.5B scale SuCo reaches 53.1% average accuracy (vs 50.5% for LHRMs) with 1,483 average tokens.
• On the MFT ablation (1.5B), MFT reaches 45.7% accuracy with 1,344 tokens vs 38.4% and 5,082 tokens for full CoT SFT.

Surprising findings

• Shorter MSC-based reasoning gives higher accuracy than full CoT on MATH difficulty levels, suggesting less reasoning can yield better performance.
• Roughly 40.7% of training samples yield empty MSC, meaning the model learns to answer without explicit reasoning for many questions.

Limitations

• MSC construction requires ground-truth answers, limiting use on open-ended generation tasks.
• MFT depends on data distilled from strong LRMs and an 80B refinement model.
• Empty CoT rate and gains may depend on the chosen threshold and length-percentile design.

Experimental setup

Datasets: Llama-Nemotron Post-Training Dataset Mixture-of-Thoughts OpenR1-Math-220k OpenCodeReasoning s1K-1.1 MATH GSM8K MBPP LiveCodeBench MMLU-STEM GPQA-Diamond AIME 2025

Benchmarks: GSM8K MATH500 AMC23 AIME25 MBPP LiveCodeBench-V6 MMLU-STEM GPQA-Diamond StrategyQA CommonsenseQA AlpacaEval 2.0

Metrics: accuracy response length (tokens) length-controlled win rate

Baselines: DeepSeek-R1-Distill-Qwen AdaCoT AdaptThink S-GRPO LHRMs Qwen2.5-Math-Instruct

Builds on: GRPO DeepSeek-R1 AdaCoT AdaptThink S-GRPO LHRMs

Closest papers by meaning

• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.91)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
• Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization (NeurIPS 2025 · similarity 0.90)
• Thinkless: LLM Learns When to Think (NeurIPS 2025 · similarity 0.90)
• Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization (ICML 2025 · similarity 0.90)
• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.90)