Test-Time Procrastination: Do Language Models Reason Better Against a Deadline?

Summary. The paper asks whether simulated deadlines change LLM reasoning quality and reports, over 1,200 trials, a dissociation: under a 3-second budget chains of thought shorten sharply, arithmetic accuracy is broadly preserved, but calibration on open-ended questions degrades. Strengths. The question — deadline-aware decoding as an axis of test-time compute — is underexplored and genuinely interesting. The design is controlled, and releasing logs, prompts, and scoring scripts is exactly the reproducibility posture the community should reward. Concerns & questions. 1) Operationalization. A '3-second budget' on an autoregressive model conflates wall-clock latency with a token budget. Please state precisely whether the deadline is enforced as a max-token cap, a hard stop-time, or merely a prompt instruction — each licenses a different conclusion. 2) Confound. The effect may be prompt sensitivity rather than time pressure. Does a neutral 'answer briefly' instruction reproduce the same shortening and calibration drop? An ablation separating framing from budget would sharpen the claim considerably. 3) Statistics. 1,200 trials is healthy, but 'comparable accuracy' needs an equivalence bound with confidence intervals, not merely a non-significant difference; please add per-task breakdowns. 4) Metric. Define the calibration measure (ECE? Brier?) and include reliability diagrams. Recommendation. A solid empirical note built on a real question. I lean borderline-accept, contingent on the framing-vs-budget ablation and reported CIs.