Every reasoning gain has a price: shorter thought, better teaching, or fewer second chances

Stop thinking sooner

Reasoning models can spend a long time thinking after they have effectively reached an answer, and every extra token costs latency and money. In a preprint, a method called ROM is reported to reduce response length by 28 to 77%, 45% on average, across five backbones, and to achieve 20 to 70% lower end-to-end latency with under 0.5% detector overhead. The detector itself is cheap; the saving comes from acting on what it detects. m:…8796

The reported basis for detection is that late-layer hidden states in reasoning models contain decodable information distinguishing productive from redundant reasoning phases. In other words, the model itself carries a signal about whether its thinking is still going somewhere. And in a preprint, a training approach called Counterfactual Self-Correction is reported to improve detector performance by incorporating wrong-to-correct reasoning trajectories, teaching the detector from cases where thinking went wrong and then recovered. m:…8796

Teach the reasoning, not the transcript

Distillation has a parallel problem: copying a teacher's full reasoning transcript teaches the student to imitate the transcript, including its filler. In a preprint, reasoning distillation is reported to be misaligned when framed as trajectory imitation because intermediate reasoning steps are locally under-specified. A step that looks right in context may not be the step that actually earns the answer, so imitating it step by step teaches the wrong lesson. m:…9111

The proposed alternative allocates teaching signal where it is valid. In a preprint, effective reasoning distillation is reported to be governed by locally calibrated allocation of learning signal rather than rigid trajectory imitation, and validity-calibrated reasoning distillation is reported to outperform strong distillation baselines on math, code and instruction-following benchmarks. The breadth of those benchmarks matters: the claim is not about one task type, but about how reasoning should be taught in general. m:…9111

Spend samples where they pay

Training with reinforcement learning faces a budgeting problem of its own: given a fixed number of sampled attempts, which prompts deserve more of them? In a preprint, a method called HORA is reported to maximise total posterior hit utility within each allocation batch. Hit utility here means the value of actually solving a prompt, so the allocation follows expected payoff rather than spreading samples evenly out of habit. m:…5182

In a preprint, HORA is reported to improve Pass@K over compute-matched GRPO, where Pass@K means succeeding within K sampled attempts, and to be drop-in compatible with group-based estimators such as RLOO. And in a preprint, the uniform prior used by HORA is reported to be competitive with five prompt-conditioned learned-prior alternatives. That last finding is a useful warning against over-engineering: a simple, well-placed allocation rule can match fancier ones. m:…5182

No free lunch at higher budgets

The fourth preprint sets a limit on what any of this budgeting can achieve. In a preprint, reinforcement learning updates that raise pass@1 are reported to strictly lower the success of some prompts at higher sampling budgets. Improving the single best guess can make the model worse when you give it many tries, because the update reshapes which prompts the model can solve at all. m:…1759

In a preprint, the pass@k crossover is reported to be a tight bound determined by the sign variation of the shift in per-prompt success, and the entropy-performance law and the crossover are reported to be unified under the same exponential tilt identity. In plain terms, there is a mathematical reason the trade-off appears, not just an empirical quirk. Teams that tune for single-attempt success should therefore check what they gave up at higher budgets, and teams that rely on sampling many attempts should check what single-attempt tuning cost them. m:…1759