When2Think Teaches Reasoning Models When to Think

Large reasoning models have made extended internal deliberation a central route to better performance. Yet longer reasoning is not automatically better. A model may spend substantial computation on a straightforward problem while failing to allocate enough effort to a genuinely difficult one. When2Think, proposed by a Microsoft research team, treats this as a computation-allocation problem that should be solved at the instance level.

Moving beyond uniform length penalties

Many efficiency strategies try to shorten every response or rely on fixed routing rules. Such approaches can save computation on easy examples, but they may also remove useful reasoning from hard cases. When2Think takes a different approach: the model is trained to decide whether an instance should receive a direct answer or an extended reasoning trace. The paper frames these two behaviors as System 1, or NoThink, and System 2, or Think.

The main mechanism is Instance-level Difficulty-Aware Control, or IDAC. It shapes the training reward with reference statistics computed in advance, including accuracy and token usage for individual instances. In practical terms, the objective is not simply to reward shorter outputs. It attempts to encourage short solutions when additional deliberation brings little value, while protecting the reasoning budget for problems where deeper computation is useful.

Critic-free post-training design

IDAC is combined with verifier-based rewards and batch-wise standardized advantages. According to the paper summary, this combination supports stable optimization without a learned reward model or online queries to a reference model. The design therefore targets two problems at once: choosing the appropriate reasoning depth and keeping the post-training procedure relatively lightweight.

Reported benchmark results

The reported mathematics experiments show an improved accuracy-efficiency balance. Relative to the base model, When2Think raises Pass@3 by 10.0% on AIME24 while reducing token usage by 27.9%. On AIME25, it reaches 40.0% Pass@3 and outperforms the compression and routing-only baselines described in the paper.

The broader implication is that reasoning efficiency may depend less on making every answer shorter than on deciding where computation matters. A hybrid model that can skip unnecessary deliberation may reduce waste without applying a blanket penalty to difficult tasks. At the same time, the available material only reports mathematics benchmarks and high-level method details. It does not establish performance across other domains, provide a full cost analysis, or resolve how robust the learned difficulty estimates are. Those questions will require further evaluation. Still, When2Think offers a useful post-training perspective: “whether to think” and “how long to think” can be learned as a single adaptive control problem.

Source: Hugging Face Daily Papers