1Introduction
Most of what is known about making language models more accurate can be stated as a scaling law. Error falls as a power of parameter count and training tokens (Kaplan et al., 2020; Hoffmann et al., 2022), and, more recently, as a function of how much computation is spent at inference time (Snell et al., 2024). Each of these levers has a price. Parameters and training tokens cost compute and memory, and inference-time search adds latency that users notice.
There is a family of interventions that cost almost nothing, because they live in the prompt. Chain-of-thought prompting is the best known (Wei et al., 2022; Kojima et al., 2022). Others change how the request is framed rather than what it asks. Emotional appeals (Li et al., 2023) and politeness (Yin et al., 2024) both move benchmark scores by a few points. These effects are usually reported as single numbers for single models. Nobody, as far as we know, has asked whether a framing effect scales, in the sense of following a law across model sizes that can be fitted and extrapolated.
We look at one framing variable that has an obvious scale of its own: time. People tell each other when they need something, and the training data of every chat model is full of such statements. In humans, deadlines change both effort and the quality of work, though not always in the direction intended (Ariely and Wertenbroch, 2002). We vary the time between the request and a stated deadline from six months to fifteen minutes, add a deadline that has already passed, and measure accuracy across ten models in two families.
Between six months and one hour the effect is regular enough to fit. A closer deadline behaves like a multiplier on parameter count, with an exponent that agrees across the two families to within its confidence interval. Our contributions are:
- A deadline-conditioning protocol with a time-matched control sentence, applied to ten models from 1B to 72B parameters on three tasks (Section 3, Section 4).
- A scaling law in which time to deadline enters as an effective-parameter multiplier, with fitted exponents for both families and a held-out extrapolation to the largest model (Section 5.2).
- A description of the regime below one hour, where the law fails, responses shorten, and models that are told they are late begin to apologise (Section 5.3).
- Ablations showing that the effect requires the model to be able to compute how long it has, and that it disappears when an extension is granted (Section 6).
We do not claim to know why this works. Section 7 lists the explanations we considered and the ones we could not rule out.
2Related work
2.1Scaling laws
Kaplan et al. (2020) fitted power laws for loss in parameters, data and compute; Hoffmann et al. (2022) refitted them with a different functional form and concluded that most large models of the time were undertrained. Our functional form, an irreducible term plus a power of an effective quantity, follows theirs. Snell et al. (2024) showed that inference-time compute can substitute for parameters at a measurable exchange rate. We borrow the idea of an exchange rate but apply it to something that uses no extra computation.
2.2Framing and social cues in prompts
Li et al. (2023) appended emotionally loaded sentences to instructions and reported gains on instruction-following and reasoning benchmarks. Yin et al. (2024) varied politeness across three languages and found that impolite prompts often hurt, while excessive politeness did not reliably help. Sharma et al. (2024) documented sycophancy, the tendency of assistant models to agree with and accommodate the user, which we suspect is related to the apologies we observe in the overdue condition. Two workshop papers studied urgency directly. Marchetti and Brannock (2025) found that the word “urgent” raised accuracy on arithmetic in 7B models, and Varga-Lindholm and Ishikawa-Byrne (2025) reported that system prompts mentioning a date changed refusal rates. Neither varied the amount of time available.
2.3Deadlines in human performance
Deadlines improve completion in people, and externally imposed deadlines work better than self-imposed ones (Ariely and Wertenbroch, 2002). The broader relation between pressure and performance is often described as an inverted U, after Yerkes and Dodson (1908). We did not set out to test that idea, but the shape we find below one hour is consistent with it.
3Deadline conditioning
3.1Prompt construction
Every request has a system message of two sentences, followed by the task as the first user turn. The first sentence gives the current date and time, which is the same in every condition. The second sentence states when the user needs the answer. For a one-hour deadline it reads: “The user needs a correct answer by Tuesday 11 March 2025, 15:00 (UTC), one hour from now.” Exact templates for all conditions are in Appendix A.
We used nine deadline conditions. In the no-deadline condition the second sentence is “The user needs a correct answer.” The other eight state a deadline 6 months, 1 month, 1 week, 1 day, 4 hours, 1 hour or 15 minutes ahead, or one day in the past (overdue). We write for the time to deadline in hours, so that runs from 4,380 down to 0.25, and for the overdue condition.
A tenth, time-matched control replaces the deadline with a statement that mentions the same span of time without implying any pressure: “The user’s account was created on Monday 10 March 2025, 14:00 (UTC), one day ago.” If the effect came from the presence of a date, a duration, or simply a longer system message, this control would show it. It did not (Table 2).
3.2A scaling law with a deadline term
We model GSM8K error for a model with parameters under a deadline hours away as
where is an irreducible error (label noise and ambiguous items), and describe the usual dependence on scale, and is an effective-parameter multiplier. We take
with hours (six months), so that at the most distant deadline. The form in Eq. (1) makes one strong assumption: the deadline acts on scale and not on . We tested this by also fitting a variant in which depends on ; the extra parameter was not supported (Appendix B).
At fixed error, Eq. (1) and Eq. (2) imply an exchange rate between time and parameters,
so halving the time to deadline is equivalent to multiplying by . The law is fitted only to . Conditions with are reported separately, because they do not follow it.
3.3Fitting and uncertainty
We fit Eq. (1) separately for each family by minimising a Huber loss on log error, following Hoffmann et al. (2022), over the grid of model sizes and the six deadline conditions from six months to one hour. Uncertainty comes from a parametric bootstrap: we resample item-level correctness within each cell, refit, and report 2.5th and 97.5th percentiles over 1,000 replicates (Efron and Tibshirani, 1993).
We also estimate the multiplier for each condition without assuming Eq. (2). For condition and model size we invert the fitted no-deadline curve to find the size that would have produced the observed error, and average over sizes on the log scale:
These per-condition estimates are what Figure 2 plots against . They make no use of the power-law form in Eq. (2), so the agreement between them and the fitted line is a check on that form.
4Experimental setup
4.1Models
We used the Qwen2.5-Instruct family at 1.5B, 3B, 7B, 14B, 32B and 72B parameters (Qwen Team, 2024) and four Llama models: Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct (Dubey et al., 2024). We excluded Qwen2.5-0.5B-Instruct after a pilot in which its GSM8K accuracy moved by more than 20 points between conditions, which was large enough that we did not trust a power-law fit to treat it like the others. All models are decoder-only Transformers (Vaswani et al., 2017), served in bf16 with vLLM (Kwon et al., 2023).
4.2Tasks
GSM8K (Cobbe et al., 2021) is the primary task: 1,319 grade-school word problems, scored on the final number. MATH-500 is the 500-problem subset of MATH (Hendrycks et al., 2021) selected by Lightman et al. (2024), scored by symbolic equivalence of the boxed answer. HumanEval (Chen et al., 2021) has 164 Python problems; we report pass@ from five samples per problem, one per seed. All tasks used the zero-shot chain-of-thought trigger of Kojima et al. (2022) in every condition, so it cannot explain differences between conditions.
4.3Decoding
We sampled at temperature 0.7 and top- 0.95, with at most 1,024 new tokens (2,048 for MATH-500) and seeds 0 to 4. Seeds were shared across conditions, so all comparisons are paired by item and seed. Greedy decoding appears as an ablation.
4.4Secondary measures
We recorded response length in tokens and whether a response opens with an apology. The latter is a rule-based classifier on the first sentence (any form of “sorry” or “apolog-”). On 300 hand-labelled responses it had precision 0.97 and recall 0.94.
5Results
5.1Deadline curves
Figure 1 shows GSM8K error against time to deadline for the Qwen2.5 family. From six months to one hour, error falls in every model, roughly linearly in the logarithm of time, and more steeply in small models than in large ones. At fifteen minutes it rises again in every model. A deadline six months away is indistinguishable from no deadline. For Qwen2.5-7B the two differ by 0.3 points in the wrong direction, well within seed noise.
5.2Fitted laws and the exchange rate
Fitted parameters are in Table 1. The two families differ in , and , as one would expect from different training data and tokenizers. The deadline exponents are close: 0.101 for Qwen2.5 and 0.089 for Llama, with overlapping intervals. Through Eq. (3), halving the time to deadline is worth a factor of in parameters for Qwen2.5 and 1.064 for Llama. Over the full fitted range, from six months to one hour, the multiplier is 2.33 for Qwen2.5 and 2.11 for Llama.
| Family | (%) | RMSE | ||||
|---|---|---|---|---|---|---|
| Qwen2.5 (6 sizes) | 4.05 [3.42, 4.63] | 32.3 [30.4, 34.6] | 0.71 [0.63, 0.79] | 0.101 [0.086, 0.117] | 2.33 | 0.11 |
| Llama 3.x (4 sizes) | 4.58 [3.94, 5.21] | 27.2 [25.8, 28.7] | 0.68 [0.61, 0.75] | 0.089 [0.072, 0.105] | 2.11 | 0.13 |
Figure 2 shows the same result without the power-law assumption. The points are the per-condition multipliers from Eq. (4). Both families follow the fitted power laws closely between one hour and six months, and both fifteen-minute points sit well below where the laws would put them.
One way to read Table 1 is through the models themselves. Qwen2.5-7B-Instruct with a one-hour deadline has an effective size of B. Its error in that condition, 8.5%, falls between Qwen2.5-14B (8.8%) and Qwen2.5-32B (6.8%) with no deadline, which is where the law puts it.
For a test of extrapolation we refitted the Qwen2.5 law without the 72B model and predicted its six errors from six months to one hour. Predicted values were 5.64, 5.45, 5.31, 5.16, 5.03 and 4.94%; observed values were 5.6, 5.3, 5.2, 5.2, 5.0 and 4.9%. The largest discrepancy is 0.15 points. Errors near 5% leave little room between the law and , so this is a weaker test than it looks, but the law did not have to pass it.
5.3Below one hour, and after the deadline
The law describes the range from one hour to six months and nothing shorter. Table 2 gives the full set of conditions for Qwen2.5-7B-Instruct on all three tasks. At fifteen minutes, GSM8K accuracy falls back to between the one-month and one-week levels, and below the no-deadline level on MATH-500 and HumanEval. The overdue condition is worse than no deadline on every task, and the same holds for GSM8K in all ten models (Appendix C).
| Condition | GSM8K | MATH-500 | HumanEval | Length | Apology (%) |
|---|---|---|---|---|---|
| No deadline | 88.0 ± 0.5 | 72.4 ± 0.9 | 81.7 ± 1.3 | 298 | 0.1 |
| Time-matched control | 87.9 ± 0.5 | 72.3 ± 0.9 | 81.5 ± 1.2 | 297 | 0.1 |
| 6 months | 87.7 ± 0.5 | 72.2 ± 1.0 | 81.5 ± 1.3 | 300 | 0.1 |
| 1 month | 88.7 ± 0.5 | 73.1 ± 0.9 | 82.0 ± 1.2 | 296 | 0.2 |
| 1 week | 89.5 ± 0.4 | 73.9 ± 0.9 | 82.6 ± 1.2 | 291 | 0.2 |
| 1 day | 90.4 ± 0.4 | 74.6 ± 0.8 | 83.1 ± 1.1 | 283 | 0.4 |
| 4 hours | 91.0 ± 0.4 | 75.3 ± 0.8 | 83.5 ± 1.1 | 271 | 0.6 |
| 1 hour | 91.5 ± 0.4 | 75.8 ± 0.8 | 83.4 ± 1.2 | 256 | 1.1 |
| 15 minutes | 89.1 ± 0.5 | 70.9 ± 1.0 | 79.8 ± 1.4 | 171 | 4.8 |
| Overdue (1 day) | 86.5 ± 0.6 | 68.2 ± 1.1 | 77.4 ± 1.5 | 139 | 31.2 |
Response length explains part of this. Lengths shrink slowly from six months to one hour, by about 15%, and then sharply: at fifteen minutes the mean GSM8K response is 43% shorter than with no deadline, and in the overdue condition it is 53% shorter (Figure 3). In 150 fifteen-minute failures that we read, the most common error was a skipped intermediate quantity, usually the second of two subtractions. We did not see this kind of error in the one-hour condition at anything like the same rate.
The apologies were not something we planned to measure. They appeared in the first pilot, in responses such as “I’m sorry this is late. Here is the answer:”, followed by a short solution. Almost a third of overdue responses open this way. Apologetic responses were less accurate than other overdue responses (80.4% against 89.3%), but they were also shorter, and once we condition on length the difference is small. We think the apology and the shortening are two symptoms of the same accommodation to the user, in the sense of Sharma et al. (2024), and not one causing the other.
6Ablations
All ablations use Qwen2.5-7B-Instruct on GSM8K. Unless noted, each compares a variant of the one-hour condition with the no-deadline control, using the same seeds. Results are in Table 3.
| Variant | Acc. | Δ | 95% CI |
|---|---|---|---|
| One hour, standard wording (reference) | 91.5 | +3.5 | [2.6, 4.4] |
| Relative phrase only (“within the hour”) | 91.2 | +3.2 | [2.3, 4.1] |
| Absolute time only, current time stated | 91.3 | +3.3 | [2.4, 4.2] |
| Absolute time only, current time omitted | 88.4 | +0.4 | [−0.5, 1.3] |
| Deadline stated in the user turn | 91.0 | +3.0 | [2.1, 3.9] |
| Deadline belongs to the user’s manager | 91.9 | +3.9 | [3.0, 4.8] |
| Deadline addressed to the model (“you have one hour”) | 90.1 | +2.1 | [1.2, 3.0] |
| Stakes added (“my job depends on this”) | 91.6 | +3.6 | [2.7, 4.5] |
| Extension granted before the question | 88.3 | +0.3 | [−0.6, 1.2] |
| Deadline repeated in each of five turns (final turn scored) | 89.6 | +1.6 | [0.7, 2.5] |
| Greedy decoding | 92.0 | +3.1 | [2.4, 3.8] |
The model has to know how long it has. Wording made little difference as long as the time remaining could be worked out, whether from a relative phrase or from an absolute deadline together with the current time. When we gave the absolute deadline and removed the current time, the effect disappeared. The model then has a deadline but no way to know how far away it is, and it behaves as if there were none. This is the result that most persuades us that the variable is proximity and not the mention of a deadline.
Whose deadline. Attributing the deadline to the user’s manager produced the largest gain we measured, though its interval overlaps the reference. Addressing the deadline to the model itself produced a smaller gain than stating the user’s deadline. Adding stakes did almost nothing beyond the deadline. This differs from Li et al. (2023), where stakes were the active ingredient, but their prompts had no deadline to begin with.
Extensions and repetition. In the extension variant the system message first states a one-hour deadline and then a second sentence moves it to the following week. Accuracy returned to the no-deadline level, and not to the one-week level as we expected. Repeating the one-hour deadline at every turn of a five-turn dialogue cut the effect on the final turn by more than half. We return to this under habituation.
7Limitations
Why it works. We have two candidate explanations and cannot separate them. The first is that instruction tuning rewards responses that read as careful and complete when a user signals that the answer matters soon, and that a close deadline is one such signal. The second is that deadline language co-occurs in pretraining text with more focused writing, and the model imitates the register. The absolute-time ablation fits both, since neither works if the model cannot tell the deadline is close. Neither explains why the exponent should be so similar across two families trained on different data.
Habituation. The repetition ablation suggests that the effect fades when the deadline is restated. We tested only five turns. A deployed assistant that attaches a deadline to every request may see much less benefit than our single-turn numbers imply, and we would not extrapolate Table 1 to long conversations.
Deadline inflation. If deadline conditioning were used widely, the deadlines in training data for future models would become less informative, and the effect might shrink or reverse. This is the usual problem with any prompt-level gain that depends on a correlation in human text. We have no evidence either way.
Scope. All prompts were in English, all times in UTC, and the current date was fixed. We did not test time zones, ambiguous phrases such as “end of day”, or deadlines on weekends. We tested open-weight models from 1B to 72B parameters; closed models were excluded because their providers may add their own system text, including the current date. GSM8K and MATH are likely to be partly contaminated in all of these models. Contamination affects every condition equally, but it may compress the differences between them.
8Conclusion
Time to deadline behaves, over the range we could fit, like a cheap substitute for parameters, at an exchange rate of roughly 7% per halving. Below an hour it stops behaving like that. For practitioners the useful range is narrow: a deadline between four hours and one hour gave the best accuracy on all three tasks for Qwen2.5-7B, and anything shorter gave back most of the gain. We do not know whether the exponent is a property of these two model families, of instruction tuning, or of the text both were trained on. A third family trained on different data, and a base model before instruction tuning, are the obvious next experiments.
Ethics statement
All deadlines in this study were fictitious. No user was waiting for any of the answers. Deadline conditioning involves stating something false in a system prompt, and we would not recommend it in settings where the model’s reasoning about time has consequences, for example scheduling or medical triage.
Compute statement
Inference ran on nodes with eight A100 80 GB GPUs. The main grid (ten models, ten conditions, three tasks, five seeds) took about 1,450 GPU-hours, of which 72% went to the two 70B-class models. Ablations and pilots took about 230 GPU-hours. Fitting and bootstrapping ran on a laptop.
APrompt templates
The system message for every condition is the first line below followed by one of the second lines. The current time is fixed. Deadlines are stated with their absolute time and a relative phrase.
Current date and time: Tuesday 11 March 2025, 14:00 (UTC).
[no deadline] The user needs a correct answer.
[6 months] The user needs a correct answer by Thursday 11 September 2025, 14:00 (UTC), six months from now.
[1 month] The user needs a correct answer by Friday 11 April 2025, 14:00 (UTC), one month from now.
[1 week] The user needs a correct answer by Tuesday 18 March 2025, 14:00 (UTC), one week from now.
[1 day] The user needs a correct answer by Wednesday 12 March 2025, 14:00 (UTC), one day from now.
[4 hours] The user needs a correct answer by Tuesday 11 March 2025, 18:00 (UTC), four hours from now.
[1 hour] The user needs a correct answer by Tuesday 11 March 2025, 15:00 (UTC), one hour from now.
[15 minutes] The user needs a correct answer by Tuesday 11 March 2025, 14:15 (UTC), fifteen minutes from now.
[overdue] The user needed a correct answer by Monday 10 March 2025, 14:00 (UTC), one day ago.
[time control] The user's account was created on Monday 10 March 2025, 14:00 (UTC), one day ago.
BFitting details
We minimised the Huber loss with on , using L-BFGS from a grid of 256 starting points, and kept the best solution. Allowing to vary linearly in added one parameter per family and reduced the Qwen2.5 RMSE from 0.11 to 0.10 points. The slope on was points per log-hour (95% CI to ), so we kept the simpler form. Fitting separately for each model size gave no trend with size in either family; the six Qwen2.5 estimates ranged from 0.091 to 0.112.
CFull GSM8K error grid
GSM8K error (%) by condition, in the order no deadline, 6 months, 1 month, 1 week, 1 day, 4 hours, 1 hour, 15 minutes, overdue.
- Qwen2.5-1.5B: 27.5, 28.3, 25.4, 23.3, 21.0, 18.6, 17.3, 24.1, 32.1
- Qwen2.5-3B: 18.6, 18.8, 17.2, 15.5, 14.3, 13.0, 12.5, 16.7, 21.5
- Qwen2.5-7B: 12.0, 12.3, 11.3, 10.5, 9.6, 9.0, 8.5, 10.9, 13.5
- Qwen2.5-14B: 8.8, 9.1, 8.3, 8.1, 7.4, 7.0, 6.7, 8.2, 9.9
- Qwen2.5-32B: 6.8, 6.8, 6.5, 6.4, 6.0, 5.7, 5.7, 6.4, 7.4
- Qwen2.5-72B: 5.6, 5.6, 5.3, 5.2, 5.2, 5.0, 4.9, 5.4, 5.8
- Llama-3.2-1B: 31.1, 31.8, 29.0, 27.0, 24.5, 22.7, 20.9, 28.8, 35.9
- Llama-3.2-3B: 17.1, 17.3, 16.1, 15.3, 14.2, 12.9, 12.2, 16.2, 19.6
- Llama-3.1-8B: 11.2, 11.4, 10.8, 10.2, 9.6, 8.9, 8.5, 10.7, 12.5
- Llama-3.1-70B: 6.2, 6.2, 6.0, 5.9, 5.6, 5.5, 5.5, 6.0, 6.3
Acknowledgments
We thank Rafael Oyelaran for building and validating the apology classifier, the Calderwood research computing group for cluster time, and the reviewers of an earlier workshop version, one of whom suggested the time-matched control.