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

LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models

NeurIPS 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 LEAD, an RL method for length-efficient reasoning that adapts reward weights online via a Potential-Scaled Instability controller and sets per-problem target lengths from correct rollouts with a symmetric length reward.

Problem

Reasoning models become verbose; static length-reward weights and global length budgets trade off accuracy against compression and ignore that per-problem reasoning needs differ.

Approach

Combines decoupled per-reward group normalization with online PSI-based reward weighting (coefficient of variation times headroom potential), and a per-prompt target length equal to the mean length of the model's correct rollouts with a symmetric penalty penalizing both overthinking and over-compression.

Key results

• On DeepSeek-R1-Distill-Qwen-1.5B (4K), LEAD reaches 53.36 average accuracy and 0.68 AES, improving over DRPO by 2.62 accuracy points and 0.18 AES, and is the only method above base accuracy (51.85) with shorter length.
• On DeepSeek-R1-Distill-Qwen-7B (4K), LEAD reaches 65.17 accuracy and -0.11 AES, the best AES among RL methods though not the shortest.
• At 8K budget on 1.5B, LEAD achieves 54.44 accuracy and 0.54 AES, the highest among trained methods.

Surprising findings

• Scalarized GRPO with length reward yields AES at most 0.08 across reward ratios, with most ratios negative.
• ShorterBetter's minimum-of-correct target causes over 3-point accuracy drop due to a single short outlier setting an overly tight target.

Limitations

• Applicable mainly where correctness is verifiable with multiple rollouts; evaluated only on math.
• Uses a single scalar target length per prompt, which may not capture multiple valid solution styles.
• Compresses at training time rather than enforcing an inference-time token budget.

Experimental setup

Datasets: MATH (level 3-5)

Benchmarks: AIME 2024 AIME 2025 AMC 2023 MATH-500 OlympiadBench

Metrics: pass@n accuracy average length Accuracy-Efficiency Score

Baselines: GRPO GDPO DRPO ShorterBetter

Builds on: GRPO GDPO DAPO ShorterBetter DRPO

Closest papers by meaning

• SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning (ICML 2026 · similarity 0.92)
• Learn to Reason Efficiently with Adaptive Length-based Reward Shaping (ICLR 2026 · similarity 0.92)
• Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.91)
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning (NeurIPS 2025 · similarity 0.91)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.90)