疎で順序シャッフルされたCoTからも答えを抽出できる推論型LLM Rethinking Dense Sequential Chains: Reasoning Language Models Can Extract Answers from Sparse, Order-Shuffling Chain-of-Thoughts

Chain-of-Thought (CoT) prompting has become a default lever for boosting reasoning in large language models, and the prevailing assumption has been that dense, strictly sequential intermediate steps are essential to the technique's success. This paper pushes back on that assumption, arguing that modern reasoning-oriented LLMs can still extract correct final answers from CoT traces that are sparse, partially removed, or even shuffled out of their original logical order.

The core contribution is conceptual as much as empirical: the authors try to separate the form of a chain of thought from its informational content. If a model's accuracy survives aggressive sparsification and reordering of intermediate steps, then the contiguous, well-ordered structure that most prompting research has emphasized may not be doing as much work as previously believed. Instead, the model may be leveraging the bag of cues contained in the steps, treating much of the surrounding scaffolding as redundant. This reframes CoT less as a strict proof-like derivation and more as a pool of useful intermediate signals.

The finding sits within a broader shift in the field. Reasoning models such as OpenAI's o1 family and DeepSeek-R1 have made very long internal CoTs a centerpiece of state-of-the-art performance, but the cost is real: more tokens, higher latency, and more surface area for hallucinated steps. Parallel lines of work, including skeleton-of-thought prompting, CoT compression, and various forms of self-consistency, already hint that long linear traces are not the only viable shape for reasoning. The present paper appears to align with that trend, suggesting that optimizing the density and ordering of intermediate steps is a promising axis for efficiency gains.

There are reasons to interpret the results cautiously. Tasks differ sharply in how much they depend on strict step order. Formal mathematical derivations, multi-step program synthesis, and tightly causal planning problems may degrade much more under shuffling than knowledge-heavy question answering, where the chain often functions as a retrieval prompt rather than a deduction. The degree to which the conclusions generalize across model scales, decoding strategies, and domains is likely to require further study, and readers should treat the headline claim as a useful provocation rather than a settled rule.

If the results hold up under wider scrutiny, the practical implications could be meaningful for deployment. Inference-time systems might prune or reorder intermediate steps without harming accuracy, enabling cheaper reasoning at scale. Training pipelines that rely on synthetic CoT data could relax their dependence on perfectly ordered traces, broadening the pool of usable supervision. And evaluation practices, which often grade chains on their internal coherence, may need to distinguish more carefully between chains that look rigorous and chains that actually drive correct answers. Viewed this way, the paper is less a refutation of CoT than an invitation to rethink which parts of it are load-bearing.