acceptodds
Withdrawn from ICLR 2027

How Can LLMs Think Better Under Pressure? Optimizing Inference through Optimal Arousal Search

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

How can machines perform better under pressure? In human organizations, moderate pressure, such as a tight deadline or a critical meeting, can sharpen focus and accelerate prioritization, while excessive pressure can induce anxiety, overconfidence, and mistakes. Inspired by the Yerkes–Dodson arousal–performance trade-off, we introduce **AgenticOAT**, a stress-augmented inference framework that operationalizes stress as controllable, task-extrinsic linguistic signals injected only through the system prompt. AgenticOAT leaves the original task query unchanged and searches a theory-grounded arousal space spanning six Challenge–Hindrance–Threat stressor types and four graded intensity levels to locate behaviorally optimal configurations for each model–task pair. To evaluate stress-induced shifts beyond accuracy, we propose a Behavior Score that jointly captures task performance, verbosity, and hedging, with changes in these dimensions defining Optimal and Sub-Optimal arousal modes. Across 11 LLMs and 5 cognitive task categories, AgenticOAT improves Behavior Score on 50/54 model–task pairs in full-sample search. In the 27 Optimal Mode cases, models achieve an average +6.1 task-performance-point improvement while reducing verbosity by 42% and hedging by 29%. Content-matched controls support an additional contribution from stress framing beyond explicit directives, while held-out evaluation shows positive mean gains on items not used for selection. Quadratic fitting further characterizes an inverted-U pattern in the aggregated Optimal Mode trajectory, with independent held-out analysis supporting stronger negative curvature in this mode. These results support stress-augmented prompting as a behaviorally tunable approach to inference-time LLM optimization while preserving the original task query. The source code will be fully released upon paper acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, stress-augmented prompting, inference-time optimization, arousal-performance trade-off, LLM behavior analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aFD1OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores whether appending prompts with text inducing different types and levels of stress can improve their performance in various domains. Inspired by the fact that humans have an "optimal" level of stress, where performance increases with a rise in stress up to a point (before it starts decreasing), the consider four levels of stress applied across six different "types" of stress (that cover challenges, hindrances, and threats). The authors propose a new metric they call the "Behavior Score", which adjusts raw increases in accuracy with a reduction in verbosity and hedging (saying "I'm not sure"). They run a grid search over all 4x6 combinations of stress intensity and type, for every model-task combination (where tasks come from five widely used benchmarks), and identify the "optimal" setting for each, i.e. where the behavior score is highest. In many cases, the accuracy improves (while verbosity and hedging reduce) in with the optimal combination of intensity and stress type. There is no clear trend about what works across LLMs/tasks or for a given LLM/task other that the fact that "challenges" improve performance more than "hindrances" (see below for my objection to this).

Critical strengths

While this paper has some strengths (such as the identification of induced stress as a means to potentially improve LLM performance), I find none of them substantial.

Critical weaknesses

Novelty and significance: The bottom line is that the authors just do a grid-search over the possible configurations and give the combination that works best for a given model for a given task. The results almost seem random, and I don’t learn anything new about LLMs from this. There is no pattern in what stressors or levels bring out the optimal performance. I cannot predict anything about a given LLM (on an unseen task) or a given task (on an unseen LLM) using this. Ultimately this paper just provides a method for prompt engineering (which is not very efficient given lack of generalizability), which also does not necessarily give you the best textual addition (to the original prompt) that would lead to improved performance, due to the limited exploration of the textual space that such an approach can afford.

Soundness: I have concerns about various design choices made by the authors, but two of them are most critical. First, it is not clear to me why "hindrances" (such as giving a model incorrect information) is included as a "stressor". I agree that this would affect the performance (for good reason), but in the case of LLMs (as opposed to humans) it is unclear why this is manifestation of stress. The fact that the performance improvements are greater in the "challenge" (urgency and task-load) case is very obvious, since an LLM given improper guidance as opposed to pressure/urgency is likely to make more errors. The fact that an ideal model which knows the answer should not get distracted by such hindrances is a separate problem (more related to sycophancy than stress). Secondly, I feel the metric is not well motivated. Unless I am mistaken, the papers the authors cite as motivation do not actually support the claim made in lines 271-273. It is not clear to me why less verbosity is desirable. For hedging, the authors do not dissociate cases where abstention is the right answer or a better alternative to the model giving a wrong answer if is unable to solve a problem.

There are various other minor issues such as the use of a separate prompt for different models and tasks in many intensity-stressor type combinations, and the absence of frontier reasoning models (which may not be as susceptible to as prompt-level changes), but those are not as critical as the ones above.

Questions for authors
  1. Could you elaborate on why less verbosity and hedging are desirable when dissociated from accuracy?
  2. Can you explain why hindrances (and specifically the ones you provide) are increasing "stress" in an LLM and that the errors being made are due to that stress and not simply due to being misled?
  3. Apart from s1 and s3, the stressor prompts are adapted for different models or tasks, how do we know to what extent are the results affected by this?
  4. In line 374, do you mean that your parameter selection involves 400 rounds of cross-validation with the training-validation split being 50-50? Does that mean you have no subset of responses that is purely used as a test set (and not in the parameter selection stage)?
  5. Beyond the statement "There exists a combination of stress intensity and type [among the possibilities you have considered] that (when translated into a prompt addition) can improve LLMs' performance for a given task", what would you say is the main takeaway from your work?
Reviewer_ud61OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AgenticOAT searches six system-prompt intervention types at four intensity levels, selecting configurations using accuracy, verbosity and hedging. Across 11 LLMs and five task categories, repeated split-half evaluation reports a 3.61-point held-out accuracy gain on 50 analyzed pairs. Component ablations and aggregate intensity-response fits are used to support a stress-specific, inverted-U interpretation.

Critical strengths

S1. The held-out protocol separates item-level configuration selection from evaluation, with positive accuracy gains in 37/50 pairs (Table 13). This supports a useful prompt-selection effect beyond in-sample maxima.

S2. The authors recognize the instruction-following confound and test directive-only, contingency-only and combined prompts. The combined condition has the largest mean gain in the 30-case ablation (Tables 5 and 10).

Critical weaknesses

W1. Stress-specific attribution remains unclear. The s1 examples include ordinary conciseness instructions. Held-out comparisons use the no-template baseline, not a fixed concise prompt or an equal-budget non-stress search. Contingencies can also communicate desired behavior. The initial ablation cases were selected from s1-L3/L4 winners, without clearly specifying whether ablation items were held out from selection; the matched-content s4 check covers only one setting. These issues limit attribution, not the existence of the reported gain.

W2. Negative curvature does not establish an intensity-dependent inverted-U. Fitting the baseline together with prompted levels can turn a step into an apparent inverted-U: the constructed profile (0,1,1,1,1)(0,1,1,1,1) gives b2=−0.143b_2=-0.143, peak 2.7 and R2=0.857R^2=0.857, passing the stated criterion despite no prompted-level decline. Appendix E.4 correctly acknowledges that held-out negative curvature is not the full criterion. The different Optimal-group counts also need an explanation of analysis populations, rather than being assumed erroneous.

Questions for authors
  1. Attribution: On a representative subset under the same held-out protocol, does the method outperform a fixed concise instruction and an equal-budget non-stress search? This would most affect my assessment.
  2. Intensity: Do held-out trajectories show a rise and fall when L=0L=0 is excluded? Please clarify the mode populations and treatment of ties, or narrow the inverted-U claim.
  3. Protocol: Please provide the exact ablation prompts, case/item selection procedure, and decoding/extraction settings. Parsing and truncation rates for the largest-gain settings would help interpret the improvement.
Reviewer_iQH6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether prompting an LLM with ‘stress’ can improve its performance, motivated by the Yerkes-Dodson law and Optimal Arousal Theory. The authors build a grid of six stressor types and four intensity levels, insert the resulting template into the system prompt, and for each model-task pair search the grid for the configuration that maximizes a composite Behavior Score (accuracy adjusted for verbosity and hedging). They report gains on 50 of 54 model-task pairs across 11 LLMs and five task categories, classify 27 pairs as ‘Optimal Mode’ (accuracy up, verbosity and hedging down), and claim that quadratic fits of accuracy against intensity validate the inverted-U hypothesis in LLMs. This paper primarly reads as AI-written slop.

Critical strengths

The paper covers many models and task categories, and the authors do disclose that selecting the best configuration inflates the score. That said, I do not think these points outweigh the weaknesses below.

Critical weaknesses

Repetitions, incorrect citation usage, and logical inconsistencies abound throughout the paper and it reads to me as almost entirely AI-written - terms are made up and defined in arbitrary math for little to no reason with improper grounding, and reading the paper becomes nearly impossible after the first couple pages. Moreover, the problem is ill-defined: the authors claim not supposing that human stress-like arousal exists in LLMs, but then proceed to base the entire paper on the assumption that such a correlate does in fact exist in LLMs and is recoverable via prompting. Beyond writing style that indicates heavy AI use, the central claim is also confounded. The winning templates are mostly the temporal-pressure ones (s1), which are literally brevity instructions (“Please respond concisely”), and higher intensity levels also change how much work is requested. Gains therefore can’t be attributed to “stress” over ordinary instructions. The directive/contingency ablation covers only s1 and s3, and the only fixed-content evidence for a framing effect is one post-hoc cell (n = 100, no interval), which I can’t treat as support.

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.