acceptodds
Under review as a conference paper at ICLR 2027

Streamlined Reflective Evolution for Task-Adaptive Self-Refinement Pipelines

Abstract

Reflective prompt optimization improves large language model (LLM) systems without updating model weights, but fixed architectures constrain how self-refinement is organized. We introduce Workflow-Designing Agents (WDA), a framework for streamlined reflective evolution of task-adaptive self-refinement pipelines. Starting from a minimal prompt, WDA jointly evolves stage instructions and their sequential structure. During evolution, we find that repeated revisions can accumulate redundant instructions in a single prompt. In WDA, we propose to address this problem with SPLIT, which redistributes these instructions across specialized stages. Three-example reflection and local screening guide selective search, while calibration scores guide Pareto admission and rollback of unhelpful trailing updates. The resulting pipelines are task-adaptive: their instructions and depth are learned from task data, then fixed for all test inputs within that task. We evaluate WDA on five benchmarks spanning knowledge, mathematical reasoning, multi-hop question answering, and instruction following. On Qwen3.5-9B, WDA achieves an average score of 51.24%, improving over the initial solver by 8.63 percentage points and the variant without SPLIT by 2.60 points. On GPT-4.1-mini, it achieves 49.00%, with corresponding gains of 5.62 and 3.69 points. These results support task-adaptive self-refinement as a complementary direction to broader agentic workflow search.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Prompt evolution, multi-agent, self-evolve

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TRePOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Workflow-Designing Agents (WDA), an agent framework for automatically optimizing task-adaptive self-refinement pipelines. Building on GEPA-style reflective prompt optimization, WDA extends the search space to include both prompt instructions and the structure of LLM modules in the agent system.

The framework uses three evolution operators: REPLACE, which refines existing prompts; ADD, which adds a refinement stage; and SPLIT, which splits an existing prompt into two specialized stages. The optimization procedure uses reflection, local screening, and calibration-guided rollback to select the final pipeline. The resulting pipeline remains fixed across test inputs.

Experiments on five benchmarks using Qwen3.5-9B and GPT-4.1-mini show improvements over GEPA, MIPROv2, Self-Refine, and other baselines. The ablation studies indicate that SPLIT improves performance over the WDA variant without that specific operator.

Critical strengths
  1. It is natural to have jointly optimization of prompts and refinement structure.

  2. The proposed method shows consistent improvements across benchmarks. WDA achieves the highest average performance across five benchmarks. The reported improvements over the initial solver and existing prompt optimization baselines are encouraging.

Critical weaknesses
  1. Potentially unfair comparisons against GEPA and MIPROv2

My major concern is the architecture used for the GEPA and MIPROv2 baselines. Both GEPA and MIPROv2 can optimize prompts within multi-stage LLM programs, even though their standard optimization procedures do not automatically refine the program architecture. However, the paper does not clearly specify the architectures used for these baselines. Appendix reports a single selected instruction for each baseline, which may suggest that they may be using single-module architectures. If GEPA and MIPROv2 were restricted to a single module while WDA was allowed to construct multi-stage pipelines, the comparison would be unfair.

A more convincing comparison would apply GEPA and MIPROv2 to pipelines with fixed number of stages, allowing them to optimize the prompts of all stages. Without such experiments, it is difficult to make the conclusion whether WDA's architecture evolution provides a meaningful advantage over existing baselines.

  1. Missing comparisons with existing automatic workflow optimization methods

The proposed method automatically evolves the structure of LLM programs, but its main comparisons focus on fixed-architecture prompt optimization methods. Several published works already investigate automatic LLM workflow or agent-program optimization, including AFlow [1], and ADAS [2]. Although WDA targets the narrower setting of sequential refinement pipelines, comparing it against some existing workflow optimizers would help establish the competitiveness and practical advantages of the method.

[1] Zhang, Jiayi, et al. "Aflow: Automating agentic workflow generation." International Conference on Learning Representations. Vol. 2025. 2025. [2] Hu, Shengran, Cong Lu, and Jeff Clune. "Automated design of agentic systems." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  1. GEPA and MIPROv2 architectures: Could the authors explain what exact architectures were used for the GEPA and MIPROv2 baselines?

  2. Comparison with workflow optimization: Could the authors compare WDA against some existing architecture-search methods, e.g., AFlow?

Reviewer_Yh9sOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Workflow-Designing Agents (WDA), a reflective search framework that learns sequential self-refinement pipelines while keeping model weights fixed. Starting from a minimal solver, WDA uses REPLACE to rewrite instructions, ADD to append stages, and SPLIT to decompose an evolved prompt into two consecutive stages. REPLACE/ADD proposals use execution traces and reference-answer feedback from three training examples and must improve local scores. SPLIT proposals are generated from the parent instruction and screened on a shared batch before calibration evaluation. An instance-level Pareto archive retains complementary candidates, and calibration scores determine final selection and rollback. Prompts and depth remain fixed at test time.

The authors report unweighted five-benchmark means of 51.24% on Qwen3.5-9B and 49.00% on GPT-4.1-mini, improving over the initial solver by 8.63 and 5.62 percentage points and over no-SPLIT by 2.60 and 3.69 points. These averages combine the benchmarks' default scores and should not be interpreted as a single pooled accuracy. The extensive appendix supplies selected prompts and ancestry trees and acknowledges provenance limitations. Inference-call counts are backbone dependent: SPLIT-enabled search reduces the Qwen average from 2.2 to 2.0 calls but increases the GPT average from 2.2 to 3.0.

Critical strengths
  1. The method jointly considers instructions and sequential division of responsibilities, with a simple deployment procedure. WDA can retain a single solver or discover specialized analysis, verification, and output stages. Once selected, the same chain runs on each test input without reference answers or online search. Appendices A.1 and A.3 provide inspectable examples of analysis–execution decompositions, making the proposal concrete.

  2. The reported gains extend across five tasks and two backbones, and several limitations are acknowledged. The main tables show SPLIT-enabled search improving over the initial solver on every task and obtaining the highest mean among the listed methods. The authors acknowledge that GPT requires more calls, that methods are not matched in token/call budgets, and that instruction specialization has not been separated from changes in depth. The appendix preserves final prompts and recorded parent relationships rather than only a single successful example. These disclosures are valuable, although they do not establish a causal explanation for the gains.

Critical weaknesses
  1. The central SPLIT contribution is not isolated from search resources and other procedural differences. The framework substantially follows GEPA-style reflection and Pareto search, making the incremental benefit of prompt decomposition particularly important. Yet Qwen SuperGPQA uses 14 versus 5 search iterations in Appendices A.1/A.2, and IFBench uses 20 versus 10 in A.5/A.6. GPT inference calls also increase from 2.2 to 3.0 on average. SPLIT differs in screening as well: it may proceed without a positive local gain and evaluates multiple proposals, whose complete number and cost are not reported. The results therefore compare complete search configurations rather than isolating decomposition or reduced redundancy. Moreover, the selected GPT Math and HotpotQA ancestry paths in A.13 and A.19 contain only REPLACE/ADD. Enabling SPLIT can indirectly alter the search competition, but those paths do not directly demonstrate improvement from splitting the selected prompts.

  2. Result provenance and reproducibility details do not yet connect the main conclusions to identifiable systems. Table 2 reports GPT Math SPLIT/no-SPLIT scores of 60.66/58.20, whereas A.13/A.14 report selected test scores of 63.93/59.84. For GPT BBEH, the corresponding values are 34.33/29.00 in Table 2 and 37.33/33.67 in A.17/A.18. GPT IFBench no-SPLIT is 38.67 in Table 2 and 41.67 in A.16. Figure 2 explicitly distinguishes historical trajectories from updated main-table results, but a complete mapping to updated prompts, runs, and call counts is missing. A.43 also acknowledges an ambiguous GPT IFBench GEPA mapping. The stated minimum of 10 Qwen iterations conflicts with the 5-iteration run in A.2. Values for Kmax/Cmax, SPLIT proposal counts, precise data versions and split identifiers, GPT search budgets, DPO training details, and variation across independent searches are insufficiently documented. Since some gains are only one or two percentage points, these issues materially affect interpretation.

  3. This seems to be an autoresearch paper.

Questions for authors
  1. [Primary result verification] Could each main-table entry be mapped to a run ID, exact data split, model version, prompt/chain hash, selected candidate, test score, and call count? Please explain the GPT discrepancies above and resolve the GEPA mapping in A.43. Do the appendix prompts correspond to the main tables, Figure 2, or separate historical runs? Which values require correction?

  2. [Essential ablation] What happens when only SPLIT availability changes, with data splits, search seeds, and total model-call or generation-token budgets matched? Please report means and variation across independent searches and include a comparison with optimized chains of equal depth or self-refinement using the same number of calls. Proposal-generation and evaluation costs should be included, rather than comparing iteration counts alone.

  3. This seems to be an autoresearch paper.

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

This paper introduces Workflow-Designing Agents (WDA), a framework for automatically learning task-adaptive self-refinement pipelines. Starting from a minimal prompt, WDA jointly optimize stage instructions, the division of responsibilties across stages, and pipeline depth using three operators: REPLACE, ADD, and SPLIT. Experiments on five benchmarks show that WDA improves average performance over the initial solver on both backbones, while SPLIT provides additional gains. The paper’s main contribution is extending prompt optimization to jointly search over instructions and the sequential structure of self-refinement.

Critical strengths

self-refinement is a promising direction that offers a relatively low-cost way to improve model capabilities.

Critical weaknesses
  1. The benchmark used in this paper are somewhat outdated by current LLM research standards. The paper also does not demonstrate the effectiveness of the proposed method on tasks involving long-horizon planning or agentic capabilities.
  2. The evaluated backbones, GPT-4.1-mini and Qwen3.5-9B, are both relatively small. In general, as model scale and instruction-following capabilities increase, the marginal benefits of self-refinement may begin to plateau. The paper lacks experiments with larger models and at a broader scale, making it difficult to assess the generalizability of the proposed method.
  3. Based on the reported results, WDA provides only modest performance gains over the competing methods.
Questions for authors

Please check the section of critical weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.