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

InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models

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

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 InftyThink, which splits long chain-of-thought reasoning into bounded segments with intermediate summaries, and fine-tunes base models on a 333K-instance reconstruction of OpenR1-Math, reporting gains over vanilla long-CoT on MATH500, AIME24 and GPQA_diamond.

Problem

Long-context chain-of-thought reasoning is computationally expensive (quadratic scaling), truncated by maximum context length, and degrades beyond pretraining context windows.

Approach

Replace a single monolithic reasoning chain with an iterative loop in which the model reasons for a bounded segment, writes a summary of progress, and continues from that summary, producing a sawtooth memory pattern; training data is reconstructed from long CoT traces by partitioning them and summarizing each segment with a 70B summarizer.

Key results

• Qwen2.5-Math-7B with InftyThink reaches 91.29% on MATH500 vs 89.51% vanilla, and 43.96% on AIME24 vs 32.92% (about 11 points gain).
• Qwen2.5-Math-1.5B average accuracy across the three main benchmarks rises from 59.54% to 65.48% with InftyThink, with lower wall-time latency (1.17s vs 1.79s average).
• Training-free prompt-based iterative reasoning with R1-distill-Qwen-1.5B drops or fails to beat the trained version on AIME24 (26.85 to 29.38 for trained) and GPQA_D (26.76 vs 35.54 trained).

Surprising findings

• Training-free prompt-based iterative summarization degrades performance relative to the fine-tuned InftyThink version, showing training on the paradigm is needed.
• Smaller summarizers (Qwen2.5-1.5B-Instruct) perform nearly as well as the 70B summarizer, suggesting summarizer size has negligible effect.

Limitations

• Quality depends on the model's summarization ability; poor summaries lose information.
• Segmenting reasoning may disrupt coherent flow for some problem types.
• Iterative inference adds steps that can increase latency in time-sensitive use.

Experimental setup

Datasets: OpenR1-Math OpenThoughts

Benchmarks: MATH500 AIME24 GPQA_diamond AIME25 Math Odyssey AMC23 HumanEval MBPP

Metrics: accuracy average generated tokens inference wall time

Baselines: Vanilla long-context CoT RoPE positional encoding interpolation CoT-Valve TokenSkip LightThinker

Builds on: OpenR1-Math DeepSeek-R1 distillation RoPE interpolation

Closest papers by meaning

• Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.91)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.91)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.90)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.90)
• LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning (ICML 2026 · similarity 0.89)
• Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.89)