Reasoning Can Be Restored by Correcting a Few Decision Tokens

Large reasoning models (LRMs) substantially outperform their base LLM counterparts on challenging reasoning benchmarks, yet it remains poorly understood where base models go wrong during token-by-token generation and how to narrow this gap efficiently. We study the base–reasoning gap by quantifying token-level distributional disagreement between a base model and a stronger reasoning model using likelihood-based divergences. Across benchmarks, we find that the reasoning advantage is highly sparse and concentrates on a small set of early, planning-related decision tokens. For instance, on Qwen3-0.6B, only $\sim$8\% of generated tokens account for the salient disagreement; these tokens concentrate early in the response, are strongly enriched in planning-related decisions ($17\times$), and coincide with high base-model uncertainty—suggesting that base models fail mainly at early planning points that steer the subsequent reasoning trajectory. Building on these findings, we propose disagreement-guided token intervention, a simple inference-time delegation scheme that performs a one-token takeover by the reasoning model only at high-disagreement positions and immediately switches back to the base model. With a small intervention budget, this sparse delegation substantially recovers and can even surpass the performance of a same-size reasoning model on challenging reasoning tasks. Code is available at \url{https://github.com/AlphaLab-USTC/RRTokenIntervention}.

Modern AI models specially trained for reasoning solve hard math problems much better than ordinary models of the same size. It has been unclear why, and where in the answer this advantage appears. We compared a small base model and a stronger reasoning model word by word, looking at how each would continue the same partial answer. We found the advantage is highly concentrated: only about 8 percent of the words account for most of the gap, and they tend to appear very early in the answer. These are mostly planning moments, where the model is deciding how to approach the problem and is also most uncertain about what to write next. Based on this, we built a simple system that lets a small model write most of the answer by itself, and only asks a stronger model to fill in those few critical planning words. With help on just a few percent of the writing, a tiny model can recover most of the stronger model's reasoning ability, and sometimes match or surpass a reasoning model of its own size.