InftyThink: Breaking Length Limits of Long-Context Reasoning (2026)

InftyThink is a novel reasoning paradigm for large language models that breaks the length limits of traditional long-context reasoning. Instead of processing entire reasoning chains in a single pass, InftyThink interleaves short reasoning segments with concise progress summaries, creating an iterative process with a characteristic sawtooth memory pattern. This approach enables unbounded reasoning depth while keeping computational costs bounded, avoiding quadratic scaling with sequence length and performance degradation beyond pre-training context windows. The method is model-agnostic and has been demonstrated on Qwen2.5-Math-7B and other architectures, achieving 3-11% improvements on MATH500, AIME24, and GPQA_diamond benchmarks. The authors also developed a pipeline to reconstruct long-context reasoning datasets into the iterative format, transforming OpenR1-Math into 333K training instances. InftyThink is open-source, with code and project page available, and was accepted at ICLR 2026.