Large reasoning models generate their outputs token by token, but not every plausible next token is equally likely to lead to a correct final answer. A new paper on arXiv investigates whether the internal representations of these models encode the relative long-horizon value of candidate tokens—that is, how much each candidate contributes to reaching the right outcome.

The authors frame the problem around the observation that at a given prefix, multiple next-token candidates may be plausible, yet they shape subsequent reasoning and differ in their chances of leading to the correct answer. The study proposes to decode this relative candidate value from the model's representations, rather than relying solely on the model's own probability estimates.

While the abstract does not specify the method or results in detail, the significance is clear: understanding how reasoning models evaluate candidate tokens internally could improve decoding strategies and help explain why some reasoning paths succeed while others fail. The work points toward a more fine-grained view of reasoning models' decision-making processes.