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

Scalable Chain of Thoughts via Elastic Reasoning

ICLR 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 Elastic Reasoning, which splits generation into thinking and solution budgets and trains with budget-constrained rollout inside GRPO so reasoning models stay accurate under tight token budgets.

Problem

Large reasoning models generate uncontrolled-length chains of thought, making inference under strict token, latency, or compute budgets unreliable; naive truncation often destroys the solution segment.

Approach

Separate budgeting allocates independent token budgets to thinking (forcibly closed with </think>) and solution phases, and budget-constrained rollout trains the model with GRPO under a fixed (1K thinking, 1K solution) budget so it learns to answer from truncated reasoning and generalizes to other budgets.

Key results

• E1-Math-1.5B reaches 35.0% Pass@1 on AIME2024 under a budget, versus 27.1% for L1-Max, 24.2% for L1-Exact, and 41.0% for the original model.
• E1-Code-14B achieves a Codeforces rating of 1987 (96.0 percentile), comparable to O1-2024-12-17 (Low) at 1991 (96.1 percentile).
• E1-Math-1.5B cuts average tokens by more than 30% versus DeepScaleR-1.5B, including 32.1% on AIME2024; E1-Code-14B cuts tokens 37.4% on LiveCodeBench (17,815 to 11,145).

Surprising findings

• A second round of training with a larger thinking budget (3K) slightly degraded AIME2024 accuracy rather than improving it.
• Trained E1 models are shorter even without budget constraints and match or slightly exceed the base models' accuracy on LiveCodeBench.

Limitations

• Assumes reasoning can be cleanly segmented into thinking and solution phases, which may not hold for interleaved reasoning or dialogue.
• Evaluation limited to math and code; other domains such as scientific QA are untested.

Experimental setup

Datasets: AIME (1984-2023) AMC Omni-Math STILL TACO SYNTHETIC-1 LiveCodeBench (2023/05/01-2024/07/31)

Benchmarks: AIME 2024 MATH500 AMC Olympiad-Bench Minerva Math LiveCodeBench (2024/08/01-2025/02/01) Codeforces HumanEval+

Metrics: Pass@1 codeforces rating codeforces percentile average tokens

Baselines: S1 L1-Exact L1-Max DeepScaleR-1.5B-Preview DeepCoder-14B-Preview DeepSeek-R1-Distill-Qwen-14B O1-2024-12-17 O3-Mini-2025-1-31 DeepSeek-R1

Builds on: GRPO S1 budget forcing L1 DeepScaleR DeepCoder

Closest papers by meaning

• Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.90)
• Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization (ICML 2025 · similarity 0.89)
• A*-Thought: Efficient Reasoning via Bidirectional Compression for Low-Resource Settings (NeurIPS 2025 · similarity 0.89)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.89)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.89)