acceptodds
Under review as a conference paper at ICLR 2027

AdaGEPA: Adaptive Feedback Allocation for Reflective Prompt Optimization

Abstract

Prompt optimization improves the performance of language-model systems on downstream tasks by refining their prompts. Classical methods evaluate prompts on task examples and use the resulting feedback to guide prompt updates through reflection. However, when feedback selection does not account for the prompt's weaknesses, these updates may improve performance on selected examples without yielding broader task improvements. To address this issue, we propose AdaGEPA, an adaptive feedback-allocation method that uses the prompt's performance and task structure to select examples for the next prompt update. Our method replaces at most one example in each feedback minibatch to target an identified weakness while preserving the remaining feedback context. Across our main experiments on six downstream benchmarks, AdaGEPA achieves higher mean validation scores than non-adaptive feedback selection under matched rollout budgets. AdaGEPA also finds high-performing prompts earlier across several tasks. In the initial Schema-Guided Dialogue (SGD) study, its half-budget prompts outperform the non-adaptive baseline's full-budget prompts in joint goal accuracy on new dialogues from services seen and unseen during search. Overall, our findings highlight the potential of adaptive feedback allocation to improve both the effectiveness and rollout efficiency of reflective prompt optimization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Prompt optimization, Adaptive feedback allocation, Natural language reflection, Large language models, Reflective optimization, Black-box optimization, Rollout efficiency

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_Ct4bOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

AdaGEPA starts from GEPA’s prompt optimization strategy. Both study providing feedback form improvements in prompt. GEPA provides overall mechanism. 1) direct an AI to solve a problem 2) show another AI and let it revise the prompt 3)solve other problems based on the revised prompts. AdaGEPA adds more details in the step 2. Rather than showing random samples to another AI, a selective choice based on the content is proposed.

Critical strengths

GEPA introduced a simple yet powerful approach to reflective prompt optimization. AdaGEPA builds on this framework with a targeted modification to feedback selection, preserving most of the original minibatch while directing one example toward the current prompt’s weaknesses. This makes the intervention easy to integrate into GEPA’s existing search procedure without additional model calls for online feedback selection.

Critical weaknesses

The methodological novelty beyond GEPA appears limited. AdaGEPA retains GEPA’s prompt revision and search procedures, modifying feedback selection through at most one task-specific replacement per minibatch. While potentially useful, this is an incremental extension, and much of the selection logic relies on separately designed task adapters. The empirical gains also vary across settings, with additional SGD studies showing little or no final performance advantage over GEPA. Stronger evidence of consistent benefits would help establish the significance of this modification.

Questions for authors

Could you provide a practical criterion, based on the search budget and feedback pool size, for deciding when to use AdaGEPA instead of default GEPA?

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

This paper introduces AdaGEPA, an adaptive feedback-allocation method for prompt optimization, which selects the examples used for next prompt update. Building upon the existing GEPA method, the paper adds task-specific adapters that based on task information and evaluation records to identify weaknesses in a parent prompt. According to the weakness, at most one example in the default three-example feedback minibatch would be replaced. The authors conduct experiments on six benchmarks and compare with the GEPA method, without the adaptive feedback allocation method. The results show AdaGEPA achieves higher mean validation scores. For the SGD study, AdaGEPA’s half-budget prompts outperform the full-budget GEPA baseline on new dialogues from both seen and unseen services.

Critical strengths
  • The paper is intuitive and well-motivated. The observation that the prompt updates should account for the prompt’s weakness is useful. Based on this, the proposed AdaGEPA uses the cached evaluation and select feedback that is relevant to the prompt’s weakness, which is a meaningful direction.
  • The experiments show that the feedback selection helps with early search outcomes. In addition, the experiments show that how to select this feedback also matter by comparing with multiple other selection strategies.
Critical weaknesses
  • The algorithm depends on task-specific design, which makes the contribution limited. AdaGEPA extends from GEPA and only introduces the adaptive feedback-allocation method. However, an important part of this method, including weakness identification and feedback example selection and ordering, rely on domain-specific rules or fitted models designed for individual benchmarks. These domain-specific designs are proved to be effective, but this does not prove the generalization capability to unseen benchmark. The SGD study also only shows the optimized prompt can work well on unseen SGD services and does not show the feedback-allocation method can also transfer without domain-specific design. This makes the contribution limited.
  • The authors claim AdaGEPA’s half-budget prompts outperform the full-budget GEPA baseline on SGD study. However, AdaGEPA needs five or ten default GEPA searches to train for the ordering models. These costs are not included for the cost comparisons. Also, these searches can already produce a good prompt. The authors find that, continuing from this stronger prompt, GEPA and AdaGEPA achieve similar improvements. Thus, I feel the current evidence is still not enough to prove AdaGEPA brings additional benefit comparing to using the found prompt and apply GEPA on it.
Questions for authors
  • Which parts and what efforts are needed for redesign or retraining when applying the adapter to a new unseen domain or benchmark?
  • Could existing adapter designed for a benchmark be applied to another one and achieve performance improvements? If it would be hard for it to improve performance of a fundamentally different benchmark, the new benchmark could have similar types.
  • Have the authors tested not only replace one example, but replacing two or three as a controlled experiment?
Reviewer_iWVSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AdaGEPA, an adaptive feedback-selection method for reflective prompt optimization. It uses cached evaluations and task structure to replace at most one example in each GEPA feedback minibatch, targeting the current prompt’s weaknesses. Experiments across six benchmarks show improved mean validation scores in the main configurations, with early-search gains and mixed evidence of transfer to unseen examples.

Critical strengths
  1. The paper isolates a consequential decision in reflective prompt search. The examples supplied to reflection determine the evidence behind each prompt revision. Studying this choice while holding the surrounding search procedure fixed makes the intervention interpretable. The local-to-full alignment audit also provides concrete motivation for evaluating more than minibatch-level repair.

  2. The empirical analysis goes beyond endpoint comparisons. The paper includes random legal replacements, structured selection without fitted ordering, coverage controls, search-external evaluations, and same-state comparisons. These controls help distinguish the contributions of replacement capacity, task structure, and ordering. Reporting cases where gains disappear or fitted ordering hurts is particularly useful.

Critical weaknesses
  1. The gains are not consistently robust. Additional SGD studies show little endpoint advantage, and several search-external confidence intervals cross zero. The evidence is stronger for early-search benefits than for reliable generalization or final-performance gains.

  2. The benefit over simpler, cost-matched alternatives remains unclear. SGD’s online comparisons exclude formation-search costs, although those searches already produce competitive prompts. Structured selection recovers most gains without fitted ordering, which also hurts performance on MultiWOZ.

Questions for authors
  1. Total-cost comparison: When formation searches are included, how does AdaGEPA compare with allocating the same total resources to default GEPA, including directly deploying or continuing from the best formation prompt? If formation is intended to be amortized, what reuse setting and approximate break-even point do you envision?

  2. Evidence for ordering: Given Tables 20 and 22–24, under what conditions should fitted ordering be preferred to Structured Unordered or a simple deficit-based rule? Can the conclusion distinguish the evidence for structured proposal construction from the evidence for learned ordering?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.