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

DyCon: Dynamic Reasoning Control via Evolving Difficulty Modeling

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.

Proposes DyCon, a training-free method that fits a ridge regressor on LRM step-level embeddings to estimate evolving problem difficulty and uses it to dynamically down-weight reflection-keyword logits, reducing overthinking token usage.

Problem

Large reasoning models overthink, producing redundant reflection even on easy problems; existing efficiency methods use static or sample-level difficulty estimates or require task-specific training.

Approach

Difficulty evolves during reasoning and is linearly decodable from step-boundary hidden states (remaining-length proxy, log-normalized). A lightweight linear regressor estimates step-level difficulty online, and the estimate modulates a logit bias that suppresses reflection tokens more when difficulty is low, requiring no LRM weight updates.

Key results

• On Qwen3-4B-Thinking-2507 AIME2024, DyCon reaches 86.7 Pass@1 vs 83.3 baseline while cutting tokens from 21493 to 18867 (about 12%).
• Across math benchmarks DyCon achieves up to 40.6% token reduction (QwQ-32B MMLU-algebra) and up to 6.7 absolute accuracy gain (QwQ-32B AIME25).
• On non-math benchmarks, DyCon reduces tokens by up to 52.5% (e.g., TriviaQA on R1-Qwen-7B) with accuracy gains up to 8.6 points on GPQA-D.

Surprising findings

• A static (non-adaptive) suppression coefficient drops AIME2024 accuracy from 50.0 to 33.3 on R1-Qwen-7B, showing that uniform suppression harms hard problems.
• Replacing the suppression vocabulary with SEAL's reflection list yields comparable or slightly better results, suggesting insensitivity to the exact keyword list.

Limitations

• Uses remaining reasoning length as a proxy for difficulty, which is noisy for redundant reasoning traces.
• Some baselines (TrimR, FlashThink, ThinkPilot) were not publicly released and were re-implemented from descriptions, and several table entries are missing for some baselines.
• Evaluation is limited to text reasoning; extension to multimodal settings is left for future work.

Experimental setup

Datasets: MATH GPQA

Benchmarks: MATH-500 AIME2024 AIME2025 GSM8K AMC23 MMLU-algebra GPQA-Diamond StrategyQA CommonSenseQA LiveCodeBench TriviaQA OlympiadBench

Metrics: Pass@1 token count

Baselines: CoD NoThinking ThinkPilot DEER SEAL Manifold Steering Controlling Thinking Speed NoWait TrimR FlashThink Dynasor-CoT

Builds on: SEAL NoWait NoThinking DEER

Closest papers by meaning

• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
• SuCo: Sufficiency-guided Continuous Adaptive Reasoning (ICML 2026 · similarity 0.89)
• DeepCompress: A Dual Reward Strategy for Dynamically Exploring and Compressing Reasoning Chains (ICLR 2026 · similarity 0.89)
• DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference (ICLR 2026 · similarity 0.89)
• Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.89)
• Promoting Efficient Reasoning with Verifiable Stepwise Reward (AAAI 2026 · similarity 0.89)