acceptodds
Under review as a conference paper at ICLR 2027

CoEvolve: Traceable Reflective Memory Evolution for Prompt Optimization

Abstract

Reflective prompt optimization turns model failures into reusable natural-language guidance, but free-form memory is difficult to attribute, repair, and control as it grows. We introduce , a prompt optimizer that represents each candidate as a traceable set of reflection identifiers and evolves this memory under an explicit budget. The system couples candidate lineage with a persistent reflection store and combines automatic failure-to-policy Distill, structured memory mutations, shared evaluation, and parent-relative memory-growth-aware selection. In five-run, equal-budget locked-test comparisons spanning financial tagging and arithmetic, multi-hop question answering, and differential diagnosis, CoEvolve-Distill leads on FINER, HotpotQA, and DDXPlus. An independent-split protocol reduces the selection optimism gap from 0.023 to 0.007–0.011. Removing persistent state causes the largest ablation loss and variability, whereas score-only selection preserves quality but uses substantially more policy tokens. In controlled recovery episodes, full provenance improves harmful-memory identification and recovery while retaining at least 0.978 of task score. These results support treating reflective memory as an explicit, auditable optimization object rather than unstructured prompt text.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Prompt Optimization, Reflective Memory, Large Language Models, Evolutionary Search, Memory Provenance, Automated Prompt Engineering

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_rS66OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CoEvolve, a reflective prompt optimizer that turns failed executions into natural-language reflections and manages them as an explicit, traceable memory. These reflections are kept in a persistent store. Each candidate prompt is defined by the reflections it uses, plus a record of its parent and the operation that created it. At each step, the method runs a parent candidate on a small batch of training examples. A language model turns the parent's failures into new reflections. The parent then produces several new candidates by combining its own reflections with the new ones, up to a fixed limit. The paper defines three ways to turn reflections into prompt text. Distill condenses them into a short policy, Add appends them as separate rules, and Patch–Prune–Refresh replaces weak rules. The new candidates are scored on the same batch, so they can be compared fairly. To choose which candidate to build on next, the method rewards validation gains but penalizes candidates that use many more reflections than their parent. The main experiments use only Distill, a configuration called CoEvolve-Distill. In experiments with DeepSeek-v3.2 on FINER, Formula, HotpotQA and DDXPlus, CoEvolve-Distill achieves the highest mean score on three of the four tasks. The paper also studies selection bias, ablates the main components, and tests whether provenance helps recover from deliberately injected harmful memory.

Critical strengths
  • The motivation is sound. Reflections from past failures are useful, but they become hard to track and control as they accumulate. The paper solves this by keeping every reflection in a persistent store and building each prompt from only the top-ranked ones. This keeps the prompt short without losing useful reflections.
  • Growth-aware selection effectively controls prompt size.
  • The ablation study covers each core mechanism.
Critical weaknesses
  • The novelty over GEPA and ACE is limited. CoEvolve keeps GEPA's search loop, while its itemized memory with helpful/harmful counts closely follows ACE. The main addition is storing reflections under stable IDs, which is an incremental change.
  • The gains are small and lack significance testing.
  • Even with the ablation study, the benefit of each component remains unclear. For example, the largest effect of removing persistent memory lowers scores by only 0.018–0.026, comparable to run-to-run variation.
  • The clarity and presentation in the methodology are very hard to follow.
Questions for authors
  • Could the authors report significance test results, given that multiple runs are already available?
  • When are children evaluated on the validation set?
  • How do Distill siblings differ from one another?
  • How is a candidate's outcome credited to individual rules?
  • What exactly is the Split-B procedure?
Reviewer_KRdhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a memory store that contains explicit records of reflections and lessons on top of GEPA, to address the concerns that GEPA sometimes loses in the growing number of reflections and prompt changes. The proposed explicit memory store introduces competitive performance (slight improvements) while providing better interpretability and debugging capabilities, as demonstrated by better prompt repair. The penalty to the memory size also introduces better reusability that, in theory, will improve generalization.

Critical strengths

This is a solid engineering addition to GEPA that people can adapt easily in their systems.

Critical weaknesses

I do not particularly see any scientific value in publishing this paper in an academic venue. The core contributions, such as the explicit memory store, are widely adopted in many other agentic pipelines for better traceability and debugging, and the system is mostly GEPA while demonstrating minimal realistic benefits over vanilla GEPA.

Questions for authors

What is the scientific contribution here, and how do you justify a publication in an academic venue?

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

The work present CoEvolve - a method for prompt optimization with main aim being able to address the difficulty of tracking and managing instructions learned from previous failures. The authors argue that previous methods like GEPA can lose the connection between learned instructions and the failures that motivated them when rewriting prompts, making it difficult to trace and repair harmful guidance.

CoEvolve, on the other hand, aims to make optimization easier to inspect and improve by maintaining a structured memory of lessons learned from failures. It links these lessons to their source examples and tracks their use in successive prompts, allowing the optimizer to reuse useful guidance and revisit potentially harmful instructions. However, across four tasks, the method shows modest improvements. Authors do a bunch of experiments where they show that retaining this history helps identify and correct harmful guidance.

Critical strengths
  • The overall motivation of having a memory and history to refer back to make sense.
  • The controlled recovery experiments are interesting and show that retaining the source and history of instructions can help repair corrupted memory.
Critical weaknesses
  • The gains on benchmarks are quite modest.
  • The writing is a bit unclear at places, in particular the methodology can be explained in a much better fashion.
  • Also, I am not sure how trivial this is, but a good baseline might be memory-augmented GEPA that retains past reflections and summarizes them into prompts. Comparing these (with similar budget) might help us understand whether the main improvements stem from retaining and summarizing past lessons or from other parts of the design, like memory ranking and CoEvolve’s selection mechanism. This will also possibly enhance the paper's ablation.
Questions for authors

Check weaknesses. Please feel free to comment if you feel something I stated is not fully correct, I am happy to have a broader discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.