Reasoning models carry a long-tolerated quirk: their thinking budget barely changes with problem difficulty. When2Think, posted to arXiv on Sep 17 by a Sungkyunkwan University and Microsoft team (the corresponding author is at Microsoft), offers a quantified footnote: on the easiest Level 1 problems of MATH-500, the 1.5B R1-Distill still writes an average of 1,199 tokens before answering, though those problems barely need any reasoning — while on hard problems the thinking often falls short.
The efficiency tax of one-size-fits-all
Existing token-saving approaches mostly take one of two roads: uniform length penalties, or rigid Think/NoThink routing. The paper calls their shared cost an "efficiency tax" — easy instances get cheaper, and hard-instance accuracy drops along with it. On AIME24, LC-R1 and AdaptThink both cut inference cost, yet accuracy falls by 10.0 and 1.3 points respectively; AdaptThink spends only 5,806 tokens but its Pass@3 drops to 44.7%, below the backbone model.
Allocation per instance: IDAC reward shaping
When2Think formalizes efficient reasoning as instance-adaptive computation allocation. Before training, a reference policy pre-computes two statistics for every problem: reference accuracy serving as a difficulty proxy, and average trajectory length serving as a per-problem token budget. An IDAC term then imposes stronger reward decay on trajectories that exceed the reference length. Combined with verifier-based rewards and batch-wise standardization (BWS), the PPO-style optimization needs no learned reward model, no critic, and no online reference-model queries during training. The backbone is R1-distill-1.5B — distilled from DeepSeek-R1 on a Qwen2.5-Math-1.5B base.
Results: fewer tokens, higher scores
On AIME24, Pass@3 rises from 46.0% to 56.0% — up 10.0 points — while average tokens drop by 3,959, a 27.9% reduction. On AIME25 it reaches 40.0% Pass@3, which the paper reports as beating compression-only and routing-only baselines. The behavioral evidence is finer-grained: on Level 1, average tokens fall from 1,199 to 619 with accuracy held at 95.8%; on Level 5, it saves 2,276 tokens while maintaining accuracy; on adversarially perturbed GSM-Plus instances it deliberately spends 462 more tokens and gains 6.3 points of accuracy. The Think ratio rises monotonically from about 0.2 at Level 1 to over 0.7 at Level 5 — computation genuinely follows difficulty rather than learning a new uniform cut.
So what
The limitations are stated plainly: the method depends on verifiable rewards, so its direct domain is math-style tasks with checkable answers; difficulty estimation relies on offline reference statistics and gets shaky when difficulty itself is ambiguous; and the team acknowledges that larger models are inherently more reasoning-efficient, so adaptive control matters most at the 1.5B scale. The engineering takeaway may be the most practical one — per-problem reasoning budgets need no online queries, just one offline statistics pass, which API services and on-device deployments sensitive to inference cost can adopt directly.
Reference: arXiv:2609.19671 https://arxiv.org/abs/2609.19671 · paper page huggingface.co/papers/2609.19671