acceptodds
Withdrawn from ICLR 2027

Weights or Words? Where to Spend an Agent's Adaptation Budget

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An agent can be improved by training its weights, or by rewriting the text wrapped around it: the system prompt, the tool descriptions, and the accumulated notes that tell it how to behave, which we call its harness. The two literatures disagree about which of these is the better use of a fixed budget. Reflective prompt rewriting has been reported to beat GRPO using an order of magnitude fewer episodes; on longer-horizon agent tasks, however, the same family of methods has been reported to lose to a single hand-written prompt. Yet neither comparison holds compute fixed, and only under a fixed budget does the comparison yield a fair and usable conclusion. We therefore hold it fixed, comparing training the weights alone, rewriting the harness alone, and doing both, under identical FLOPs on a multi-turn retrieval agent. Our systematic experiments reveal that rewriting the harness does not turn budget into accuracy: a tenfold budget increase changes success by at most , and every interval contains zero. The same increase spent on the weights, by contrast, yields . This asymmetry is not specific to one setting: the null for rewriting reappears in a second environment and at a second model scale, whereas the advantage of doing both, the strongest configuration on retrieval, does not survive the change of environment. Nor do the two sides combine, since transplanting a rewritten harness onto separately trained weights adds nothing; they are not interchangeable stores of the same knowledge. The null itself resists the obvious explanations, as a rewriter nine times larger writes no better prompts, while a third of the proposals are accepted and read like sound advice. Taken together, these results suggest that what limits such an agent is not the quality of the harness it is given but its capacity to exploit it, which rewriting cannot supply.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Agent

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_8crFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper compares three ways to improve an LLM agent under a fixed compute budget: training model weights, rewriting the agent's harness (prompts and skills), or combining both. The experiments show that weight training generally improves performance as the budget increases, while harness rewriting brings little benefit. Combining both helps in some settings, but the improvement does not generalize across tasks.

Critical strengths

The paper studies a practical and important question: where should we spend the compute budget to improve an agent?

The experiments provide a fair comparison by accounting for compute costs, and include different model sizes and environments.

The analysis provides useful insights into why harness rewriting may fail, including the performance drop caused by changing the harness during training.

Critical weaknesses

The paper only tests one harness optimization method, making it difficult to generalize the negative results to other methods.

The experiments mainly focus on relatively simple tasks, so it is unclear whether the conclusions hold for more complex, long-horizon agents.

Questions for authors

Is the limitation in harness optimization or in the model's ability to use the harness? The experiments show that the current rewriting method brings little improvement, but this does not necessarily mean that better harnesses cannot improve the agent. How can we distinguish these two possibilities?

The actual compute usage ranges from 52% to 122% of the allocated budget, with systematic differences across methods. Could these differences affect the conclusions, especially when comparing joint optimization with weight training?

Reviewer_eT7eOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates how to best allocate a fixed computational budget for LLM agent adaptation, comparing weight training, harness optimization, and their combination. The authors find that weight training generally provides more reliable performance improvements than harness optimization. Combining both methods can be beneficial in certain retrieval settings.

Critical strengths

Rather than merely presenting the overall results, the authors investigate the research question from multiple perspectives, systematically examining different factors and conducting additional experiments to better understand the observed findings.

Critical weaknesses

**Insufficient new insights: ** The central question has already been extensively investigated in prior work (e.g., GEPA, E-SPL, and Multi-module GRPO). While this paper introduces a FLOPs-matched comparison, it does not propose a new adaptation method, and the main observation that automatic prompt optimization can fail to improve agent performance has also been reported in previous studies. I find the additional insights somewhat limited, as the conclusions largely restate the relative effectiveness of existing techniques under a particular experimental configuration.

**Insufficient support for the broader conclusions: ** The conclusions appear highly dependent on the specific experimental settings, including the task environment, training algorithm, harness representation, and prompt optimization procedure. The main experiments focus on relatively simple retrieval-based QA with short interaction trajectories, while weight optimization is restricted to GRPO and harness optimization to one evolutionary rewriting approach. Different prompt formulations, optimization methods, or more complex agentic environments (which are actually more common in agent studies) requiring long-horizon planning and diverse tool interactions could lead to different conclusions. Without stronger evidence across diverse adaptation methods and more challenging agentic settings, I find it difficult to draw meaningful general conclusions.

Questions for authors

Could the authors clarify what new insights this paper provides beyond previous studies, particularly beyond introducing a unified FLOPs budget? How does the FLOPs-matched comparison lead to conclusions that could not be obtained from previous empirical evaluations?

How robust are the conclusions to choices of the training algorithm/a particular evolutionary approach for harness optimization? Is there evidence that the poor performance of harness optimization is not simply a limitation of the specific rewriting algorithm or harness representation used in this work?

Would the conclusions hold in more representative agentic environments, e.g., those involving (long-horizon) planning and diverse tool calls?

Reviewer_dHGxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper compares spending an agent's adaptation budget on weight training, harness rewriting, or both. On HotpotQA, increasing the budget improves weight training when retrieval provides enough evidence, while harness rewriting shows no significant improvement. Additional experiments examine another model scale, TextWorldExpress, a larger rewriter, and harness transfer.

Critical strengths
  • The comparison addresses a practical question. I agree that including rollout, update, and rewriting costs in a common accounting framework is more informative than comparing episode counts alone. Also, the substantial weight training gains in the better supported retrieval settings are useful evidence.
  • The paper includes meaningful studies beyond the main result. A second environment, another model scale, a larger rewriter, and seed level reporting help us to understand where findings persist or disappear.
Critical weaknesses
  • My main concern is that the statistical evidence does not seem to support the main conclusions.
    • A confidence interval containing zero does not establish that rewriting cannot help. And resampling three seeds 10K times does not resolve that limitation.
    • Likewise, a significant improvement for the combined method and a nonsignificant improvement for weight training do not establish a difference between them.
    • The exchange experiment also reports additivity residuals near zero. This is consistent with additivity under the paper's own definition, rather than evidence against it.
  • The central compute comparison and its explanation remain unclear to me. Actual spending ranges from 0.52x to 1.22x the assigned budget, with systematic diffs across methods. I think comparisons at common actual spending are needed.
  • Additionally, reasonably looking advice, accepted proposals, and a nonsignificant larger rewriter result do not establish that the agent's ability to use instructions is the bottleneck (rather than the optimizer or its search space).
Questions for authors
  • Please clarify, how do the main comparisons change when evaluated at common actual FLOPs? Please also clarify whether candidate evaluation, validation, and prompt-prefill costs are consistently included.
  • What are the train/valid/test sizes and separation procedures, the exact rewriting protocol, and the TextWorldExpress task configurations? Where can the stated released code and preregistration be accessed?
  • I'd like to see additional evidence. Can more seeds and a predefined threshold for a practically meaningful gain support the claimed absence of benefit? How do you reconcile the near zero additivity residuals with the conclusion that the two gains do not add?

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

Related papers

Discussion (0)

Sign in to comment.