acceptodds
Withdrawn from ICLR 2027

CORE-Flow: Counterfactual Observation and Retrieval-Guided Exploration for Efficient Agentic Workflow Evolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automated workflow design enables adaptive solving of complex tasks through fine-grained composition of capability modules and multi-round iterative optimization. However, efficiently searching the large space of workflow structures, operators, and prompts remains difficult. Existing methods typically rely on an optimizer model to diagnose the deficiencies of candidate workflows, thereby improving search efficiency. However, such diagnosis introduces uncertainty into the update direction due to the stochasticity of model outputs, while overlooking reusable historical evidence. Therefore, we propose CORE-Flow, an experience-driven framework for efficient workflow evolution through evidence reuse. CORE-Flow uses counterfactual observations of workflow edits to construct reusable optimization experience and retrieval-guided exploration to focus subsequent search on historically supported directions. Across six benchmarks spanning mathematical reasoning, code generation, and question answering, CORE-Flow achieves the best average score of 89.41, outperforming JUDGEFLOW and AFLOW by 2.92 and 7.89 points, respectively. The cost analysis tracked on the MATH dataset further demonstrates strong potential to reduce both token consumption and wall-clock execution time.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic Workflow Optimization, Experience-Guided Search, Large Language Model Agents

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

This paper proposes CORE-Flow, which extends AFlow-style automated agentic workflow optimization using experience from evaluated edits. Each evaluated parent-child transition is stored as a "state-diagnosis-action-effect" experience, with the effect defined as the child's validation score minus the parent's. A prompt and a validator restrict each candidate edit to one workflow component, allowing the measured effect to be attributed to that edit. An LLM writes the diagnosis after observing the effect.

Each round has three steps:

  1. Sample a parent from a scored candidate pool using a softmax over validation scores.
  2. Retrieve the top-K experiences by embedding similarity. Retrieval uses a one-sentence description of the workflow's information flow generated by an LLM.
  3. Propose B edits to the same parent and retain the best child.

The method is evaluated on established benchmarks with GPT-4.1-mini as the optimizer and GPT-5-mini as the executor. A single tracked MATH run shows lower token cost and wall-clock time than both baselines.

Critical strengths
  • The paper addresses a practical problem: LLM-driven workflow optimization is costly and repeatedly diagnoses similar workflow states. Reusing measured edit outcomes is an intuitive response to this repeated work.
  • The method is simple and training-free, and it can be integrated into existing AFlow-style pipelines.
  • Under the reported setup, CORE-Flow performs better than the compared baselines on average.
Critical weaknesses
  • The contribution is incremental. CORE-Flow largely follows AFlow, which already supplies past edits and their outcomes, and adopts JudgeFlow-style actions. The additions are modest, and related work such as GEPA, MASS, and ERM is not discussed.

  • The claims exceed the evidence. The “counterfactual” effect is a validation score difference, without reported subset size, evaluation count, or variance. The optimizer still uses current and LLM-written diagnoses, so the method supplements model-based diagnosis rather than moving beyond it.

  • The experiments are limited. They cover six established benchmarks, one optimizer-executor pair, and no agentic or tool-use tasks. Three benchmarks already reach scores of 94 to 98. All analyses beyond Table 1 use only MATH; ablations do not isolate single-component edits or post-hoc diagnosis. Cross-run or cross-task reuse and standard deviations are not reported.

  • The efficiency evidence rests on one MATH run, with unclear accounting for optimizer, diagnosis, and embedding calls.

References:

Questions for authors
  1. How does the experience pool differ from AFlow's tree-structured experience? Is “Optimization only” in Table 3 equivalent to AFlow, and does “+Experience retrieval” retain AFlow's native experience?

  2. What is ∣Bcand∣|\mathcal{B}_{\mathrm{cand}}|? Is each candidate evaluated once? How often does re-evaluating the same edit reverse the sign of its effect (Eq. 13)?

  3. Which calls do the costs in Tables 2 and 4 include? Are reasoning tokens counted, and what prices are assumed? How many executor calls does each method make per round?

  4. Does experience transfer across runs, tasks, or executors? Does top-k retrieval outperform giving the optimizer all past experiences?

  5. Do the gains hold on harder agentic benchmarks, such as tool use or software engineering?

Reviewer_7dxFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CORE-Flow, a framework for automated agentic workflow optimization that reuses historical editing experiences. Each experience records a workflow state, diagnosis, edit action, and observed performance change. Retrieved experiences guide localized edits and same-parent breadth search. Experiments across six benchmarks report an average score of 89.41, outperforming JUDGEFLOW and AFLOW, with additional efficiency measurements from one MATH optimization run.

Critical strengths

The paper addresses a practical limitation of workflow optimization: repeatedly diagnosing similar states without effectively reusing previous outcomes. Recording both successful and unsuccessful edits is a sensible design choice. The evaluation spans mathematical reasoning, code generation, and question answering, and the ablations attempt to separate the contributions of experience retrieval and breadth search.

Critical weaknesses
  1. The method combines structured experience storage, embedding-based retrieval, localized editing, and candidate evaluation. The paper does not sufficiently establish what is technically new beyond integrating these familiar components into workflow optimization.

  2. Table 2 reports wall-clock totals of 2,578.3, 1,876.1, and 1,364.0 seconds, whereas Appendix Table 4 reports corresponding totals of 20,626.6, 15,008.6, and 10,912.1 seconds, approximately eight times larger. The accounting or normalization behind this discrepancy is unexplained. Furthermore, efficiency evidence comes from only one MATH run, limiting the robustness of the cost claims.

  3. The manuscript uses noticeably thin typography, including lightweight title and section headings, which reduces readability and appears inconsistent with the expected conference presentation. The authors should verify their template and font configuration.

Questions for authors

see weakness

Reviewer_9hixOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CORE-Flow is a technique for automatically enhancing agentic workflows, where a workflow is represented as a code-defined graph made up of different LLM operators. Built upon the AFLOW search loop, the method iteratively selects promising existing workflows, generates a concise summary of their information flow, retrieves relevant prior similar experiences, and invokes an optimizer to propose targeted single-component modifications; candidate changes are validated, assessed on shared validation data, and successful variants are added back into the candidate pool. Every evaluated transition from a parent workflow to its modified child is stored as a structured record capturing state, diagnosis, action, and resulting performance change, with both positive and neutral or negative outcomes retained. The constraint of only modifying one component at a time is framed by the authors as counterfactual attribution. Experiments show the approach outperforms related baseline methods on a suite of benchmarks while reducing resource consumption, and ablation studies confirm the performance gains brought by its retrieval and breadth-search components, with the configuration used in the main setup working best.

Critical strengths
  1. The motivation is clear: reuse measured edit outcomes across the search rather than re-diagnosing from scratch, and keep negative outcomes too.

  2. Single-component edits with diagnosis written after the outcome are a simple, sensible way to anchor LLM explanations to evidence.

  3. The cost trace and the structural explanations of baseline cost behaviour are unusually detailed, and the prompts are provided.

  4. Reported gains are broad (best average, first on three benchmarks), and MATH convergence appears fast.

Critical weaknesses
  1. Cost claims are internally inconsistent. Table 2 times are Table 4 totals divided by exactly 8, not rounding. "Sample time" is per-sample latency × about 119, not wall-clock. AFLOW round 10's time and cost imply different sample counts. Absolute costs are about 50× below what GPT-5-mini pricing implies, and what is counted is never stated.

  2. The method contradicts itself. The Eq. 18 validator rejects every AddNode and RemoveNode edit and every edge edit allowed by Appendix C. "Counterfactual" overstates a single noisy interventional score difference.

  3. Statistical evidence is weak. No variance is reported. The value patterns suggest single runs despite the "three runs" claim. Table 2 accuracies equal the Table 1 means. Figure 4's B = 1 value contradicts Table 3. The K and B story rests on a pool of at most about 20 records.

  4. Baselines are questionable. Several optimizers fall well below IO, and the evaluation budget is not matched (B = 2).

  5. Novelty is overstated. AFLOW already reuses modification–outcome experience, and ExpeL, AWM, ReasoningBank, OPRO and similar work are not discussed. Diagnosis and edits remain LLM judgments.

  6. Presentation problems. The Figure 3 panels are swapped, the B-edits-at-once vs. one-edit-per-call description is inconsistent, and d_t, τ, M and B_cand are undefined.

Questions for authors

Q1. Explain the factor of 8 between Tables 2 and 4, how time is measured, the round-10 inconsistency, and exactly which calls are included in the costs.

Q2. What was the validator actually run? Give the distribution of accepted action types. How does CORE-Flow perform structural edits?

Q3. Are Tables 1–3 and Figures 3–4 single runs? Please give per-seed results with standard deviations and CIs, and say why the Table 2 accuracies equal the Table 1 means.

Q4. What is "Optimization only"? Why does Figure 4's B = 1 differ from Table 3's 86.63?

Q5. How were the baselines reproduced? Why do AFLOW, ADAS and MaAS fall below IO?

Q6. Give results under a matched evaluation and token budget, and with random-retrieval controls.

Q7. How does CORE-Flow differ concretely from AFLOW's tree-structured experience and from ExpeL, AWM and ReasoningBank?

Q8. Specify d_t, τ, M, B_cand sampling, the retry limit and the final-selection rule, and release the code.

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

Related papers

Discussion (0)

Sign in to comment.