SeedThink: Test-Time Control via Seed-Thought Initialization

ICLR 2026 Conference Submission25520 Authors

Keywords: Large Reasoning Models, Speculative Decoding, Efficient Reasoning

Abstract: Large reasoning models (LRMs) achieve impressive performance via extended chains of thought, but this substantially increases inference overhead, making efficiency a critical bottleneck. In this paper, we first show that initializing the reasoning process with high-quality seed thoughts can steer the model away from unproductive "overthinking'' and produce more efficient reasoning trajectories. Critically, we find that the optimal granularity of this seed --- from a high-level outline to a detailed solution --- depends on problem difficulty. Motivated by this, we propose SeedThink, a novel framework that adaptively selects the seed granularity based on an estimate of problem difficulty. Specifically, SeedThink features two core innovations: (1) a \textbf{difficulty-aware seeding policy that dynamically generates seed thoughts} to reduce repetitive verification and prune unproductive branches; and (2) \textbf{seamless integration with enhanced speculative decoding}, where seed thoughts are reused as a model-free draft corpus to achieve dual-path acceleration --- shorter reasoning traces and faster token generation. Our experiments show that {SeedThink} significantly reduces inference costs while largely preserving performance. Notably, our method achieves up to 4.1× end-to-end speedup and a 68\% reduction in generation length with minimal accuracy degradation, highlighting the promise of adaptive initialization for balancing reasoning quality and efficiency.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 25520