acceptodds
Withdrawn from ICLR 2027

Rethinking Self-Evolving Agents: Do We Still Need Prescribed Optimization Pipelines?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-evolving agents are usually built around prescribed optimization pipelines: the framework decides how to gather evidence, revise a persistent artifact, select candidates, and stop. We ask whether this task-specific procedure remains necessary when a frontier model acts as the optimizer. We introduce Open-Ended Optimization (OEO), which fixes the objective, permitted interactions, resource budget, data boundary, and evaluation while allowing the optimizer to compose the improvement process online. We compare OEO with two complementary prescribed approaches: SkillOpt, a staged pipeline with bounded edits, and GEPA, a reflective evolutionary search. Across a complete grid of 16 head-to-head comparisons over 8 benchmark–target-model settings, GPT-5.5-driven OEO produces 12 higher final-score point estimates, 1 tie, and 3 lower estimates, with a median absolute difference of 3.1 percentage points. It improves the shared initial skill in all 8 settings while using a median 34.3% of SkillOpt’s configured target-interaction token allocation. Across 18 repeated optimization runs on 2 anchor settings, OEO retains the highest mean on LiveMath, whereas the 3 methods’ SearchQA means lie within 0.84 percentage points. A capability probe further exposes the boundary of delegation: a weak optimizer cannot execute the open-ended loop under the fixed interface, prescription is task-dependent at medium capability, and frontier OEO is competitive on both tested tasks. A one-shot control shows that the gains are not explained by a single prior-driven rewrite. Finally, the fully instrumented OEO–SkillOpt pair shows that prescription changes committed optimization paths more consistently than selected item-level behavior. Together, these results suggest that task-specific optimization procedures need not always be prescribed in advance. With a sufficiently capable optimizer, the procedure itself can instead be composed online within externally fixed objectives, permitted interactions, resource budgets, data boundaries, and evaluation, while explicit prescription remains a task- and capability-dependent scaffold rather than a universal requirement.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Self-evolving agents, open-ended optimization, agent optimization

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

The paper asks whether self-evolving agents still need a prescribed optimization pipeline when the optimizer is a frontier model. It proposes Open-Ended Optimization (OEO): the framework only fixes the objective, budget, data boundary and evaluator, and the optimizer (GPT-5.5) uses five tools (inspect data, read artifact, run target, write skill, finish) to improve a natural-language skill on its own. OEO is compared with SkillOpt and GEPA on SearchQA, SpreadsheetBench, OfficeQA and LiveMath with two target models. OEO has higher point estimates in 12 of 16 comparisons and uses less target-side tokens. The paper also has a capability ladder, a one-shot rewrite control, and an analysis of edit paths.

Critical strengths
  • The question is clear and timely. Separating the fixed "contract" from the optimization procedure is a useful way to frame it.
  • The statistics are done carefully, and the one-shot control is a good sanity check.
Critical weaknesses
  1. Large seed variance on LiveMath. The headline LiveMath / GPT-5.5 result uses seed 42, which is the best of three runs (0.71 vs 0.55 and 0.50 for seeds 43/44). The SD of OEO is 0.11, about 3-4x of SkillOpt and GEPA. The two largest significant gains in the paper come from this one setting with n=124. Table 2 also keeps the "historical" 85/124 while Table 7 uses 88/124 for the same artifact. With this variance, I am not sure the LiveMath gain is a property of the method or a lucky run.

  2. Baselines look very weak. GEPA improves LiveMath / GPT-5.5 only from 0.387 to 0.403, and SkillOpt on SpreadsheetBench / Qwen is even below the initial skill (0.196 vs 0.211) despite having a validation gate. The paper does not say whether official implementations and recommended hyperparameters were used, or why the baselines fail in these settings.

  3. Not enough examples of what OEO actually does. The method is very simple, a loop with five tools, so the interesting part is the behavior. But the main paper shows no trajectory, no example of the final skill, no tool-call statistics, and no stopping behavior. The system prompt and action schema are also not in the paper.

  4. The gain may come from edit freedom, not from composing the procedure. Figure 4 shows OEO makes much larger and broader edits than SkillOpt, which is restricted to bounded patches. So the gain over SkillOpt may simply come from allowing full rewrites. An ablation is needed to support the claim that the procedure itself is what matters.

Questions for authors

Please see weaknesses.

Reviewer_7H2HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether skill optimization still need a prescribed pipeline. OEO lets the optimizer decide how to gather evidence, revise skills, retain candidates, and stop, while keeping objectives, permissions, budgets, and evaluation fixed. Comparisons with SkillOpt and GEPA show competitive results with a strong optimizer.

Critical strengths
  1. It is interesting to ask whether a strong model can organize skill optimization without having every step prescribed in advance.
  2. The comparisons across multiple benchmarks and target models make the findings more convincing.
Critical weaknesses
  1. Optimization and evaluation use different splits of the same benchmark, so transfer beyond that benchmark remains unclear. I am concerned that some gains may rely on benchmark-specific shortcuts. Could you manually inspect the skill edits for dataset-specific shortcuts or scoring hacks, and test the resulting skills on a related unseen dataset or task?
  2. The benchmarks cover different domains, but the reason for this particular selection and the difficulty range it represents are unclear. The optimizer-capability probe offers some evidence, but there is no systematic study in task difficulty. It remains unclear how strong the optimizer needs to be to reliably control the process for a given domain and difficulty.
Questions for authors

Please see the questions raised in the weaknesses above.

Reviewer_yS6FOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose a skill optimization method, namely OEO. Compared to related work, OEO let an LLM agent to decide when and how to gather evidence, revise skills, select candidates, and terminate optimization, instead of following a predefined pipeline. Experiments show that OEO outperforms SkillOpt and GEPA in most settings, when the optimizer model is capable enough.

Critical strengths
  1. The research question that this paper focuses on, namely "Do we still need prescribed optimization pipelines when the optimizer is a capable LLM?", is meaningful and interesting. The motivation of this paper is also clear: in the area of skill optimization, existing works usually predefine the optimization pipeline, which adds different constraints to the optimization process, whereas some LLMs like GPT-5.5 should be capable enough to make decisions in this process.

  2. The experiments span four benchmarks, showing that OEO generally works across different tasks. The authors also compare the method with SkillOpt under settings with different optimizer capabilities. This helps readers understand when to use this method.

Critical weaknesses
  1. OEO appears to be a very straightforward application of a ReAct-style tool-using agent to skill optimization. The core idea of using an LLM to make decisions during optimization is well established in the LLM agent area. The paper also does not bring enough new insights in this area.

  2. In Section 4, the findings are not well supported by the experimental design. In 4.1, the authors compare the optimization paths between SkillOpt and OEO and show that OEO has a larger maximum edit, larger revision breadth, and higher revision churn, and then conclude that "prescription reshapes committed optimization paths." But these observed differences are largely expected from the implementation (SkillOpt limits edit operators while OEO doesn't), rather than showing which aspects of prescription cause these differences. This experiment looks more like an implementation sanity check rather than a mechanism analysis to me.

Questions for authors
  1. In Section 4.1, how do you distinguish the effects of LLM-controlled optimization from the differences in edit operators between SkillOpt and OEO? Would the same findings hold if both are allowed to use the same edit operators?

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

Related papers

Discussion (0)

Sign in to comment.