acceptodds
Under review as a conference paper at ICLR 2027

FreeEvolve: Learning to Evolve Beyond Fixed Loops

14 views

Abstract

Agent evolvers automate the design of the prompts, skills and workflows around language model agents, yet the optimization process they follow is still designed by hand: a fixed search loop decides how candidates are evaluated, which are kept and when the search stops. We propose FREEEVOLVE, which automates this process as well. An environment specifies the goal, target agent, evaluator, data and resource limits; within these limits, the evolver itself decides what to test, how much evidence to collect, which candidates to pursue and when to stop. These decisions follow an editable evolution skill, which we improve through meta-evolution by scoring each candidate skill on the fresh target agent it produces. The optimization process thus becomes a capability learned from experience rather than a loop engineered in advance. On tau3-bench, ARC-AGI-2, ARC-AGI-3 and Terminal-Bench 2.1, FREEEVOLVE controls the evolution campaign by itself, yet improves the primary held-out metric by 13.6 points on average and matches or exceeds hand-designed evolvers. The learned process keeps improving with experience: meta-evolved skills add 6.9 points over the seed skill on fresh target agents, demonstrating transferability across environments.

Primary area:
generative models
Keywords:
Self-improving agent, harness evolution, skill as evolver

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_A6ZROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FreeEvolve, a framework that allows LLM-based evolvers to manage their own optimization campaigns. Instead of following a predefined search loop, the evolver decides which tasks to evaluate, how to allocate resources, which candidates to pursue, and when to stop. The authors further introduce meta-evolution to improve the evolution skill itself. Candidate skills are evaluated through complete worker-evolution campaigns. Experiments on τ³-bench, ARC-AGI-2, ARC-AGI-3, and Terminal-Bench 2.1 show that FreeEvolve generally outperforms existing evolution frameworks. Additional analyses examine the benefits of meta-evolution, cross-environment transfer, and adaptive campaign decisions.

Critical strengths

The paper extends agent evolution beyond candidate generation by allowing the evolver to manage evaluation, resource allocation, and search decisions. Treating the evolution skill as an editable target is also interesting. It provides a way to improve the optimization process using feedback from complete evolution campaigns.

FreeEvolve achieves strong performance on four different agent environments. The fresh-worker experiments help evaluate whether improvements are retained in the learned evolution skill. The cross-environment transfer experiments and campaign-control ablations provide additional evidence about how the method works.

Critical weaknesses

1. Unfair computational budgets in the main comparisons.

Table 1 compares methods under different stopping conditions and resource budgets. The baselines stop after five iterations without improvement, while FreeEvolve receives 24 hours of meta-evolution with up to five worker nodes. Although Appendix Table 9 reports wall-clock times, the total computational costs are not consistently compared. Figure 4(b) examines token efficiency but only on ARC-AGI-3. It remains unclear whether FreeEvolve's advantages would persist under matched computational budgets.

2. Insufficient isolation of the main methodological contribution.

The novelty over existing meta-evolution methods needs to be clarified. EvoX [1] already adapts its search strategy during optimization, while HyperAgents [2] evolves the mechanisms used to improve agents. FreeEvolve offers broader control over the campaign, but the specific benefits of this additional flexibility are not fully established.

Figure 3 shows that meta-evolved skills outperform the seed skill. However, this does not establish how much of the improvement comes from autonomous campaign control versus better evolution instructions. Table 4 examines several fixed decisions but only on τ³-bench. A more direct comparison would help isolate the contribution of the proposed mechanism.

References

[1] Liu et al. (2026). EvoX: Meta-Evolution for Automated Discovery. https://arxiv.org/abs/2602.23413

[2] Zhang et al. (2026). HyperAgents. https://arxiv.org/abs/2603.19461

Questions for authors
  1. Could the authors compare FreeEvolve with the main baselines under matched computational budgets? Reporting performance against total token consumption and node-hours would help establish whether the gains come from more effective search.

  2. Could the authors compare seed and meta-evolved skills under both fixed and autonomous campaign configurations? This would help isolate the contribution of learned campaign control from improved evolution instructions.

  3. Could the authors directly test cross-model transfer of the learned evolution skill? For example, a skill learned using Opus 4.8 could be executed by Sonnet 4.6 on fresh worker tasks. Can the authors also explain the negative transfer observed on Terminal-Bench 2.1?

6 comments

Dear Reviewer A6ZR,

We really appreciate your thorough comments and your deep insights in helping us improve our work. Here we address your concerns as follow

  • W1/Q1: Matched Compute Results

We agree with you, Table 1 is not a strictly fair comparison. The purpose of Table 1, as we also suggested in line 315, is to show that given sufficient resource, ultimately how much better the worker can get using different evolver, so we try to compare the capability boundary and potential of the evolvers.

Hence, following your suggestion, we further assess the evolvers under comparable or less total token budgets. We conduct three independent searches with the frozen meta-evolved FreeEvolve skill on Tau-3 and ARC-AGI-3. Every search starts from the same fresh worker and runs for 12 hours. We evaluate each returned worker three times and use its mean held-out performance as one search-level observation.

BenchmarkEvolverIndependent searchesHeld-out score (%)Total search tokens (B)
Tau-3FreeEvolve344.0 ± 1.31.53 ± 0.46
Tau-3GEPA336.0 ± 0.62.02 ± 0.02
Tau-3HyperAgents336.7 ± 0.82.02 ± 0.03
ARC-AGI-3FreeEvolve367.9 ± 3.11.12 ± 0.12
ARC-AGI-3GEPA362.7 ± 1.52.22 ± 0.05
ARC-AGI-3HyperAgents362.7 ± 3.82.24 ± 0.03

Here, the mean and standard deviation are computed across the three independent searches, rather than repeated evaluations of one worker. Total search tokens include every model call during evolution.

We can see that, the evolved meta-skill is both more efficient and more effective. This is in fact very intuitive, under the same time budget, FreeEvolve can choose to focus on a subset for a specific problem, while GEPA and HyperAgents does not have such mechanism and have to evaluate the full set every time. So even when the ability to search and make changes is the same, FreeEvolve can be more efficient.

Meta-evolution itself is more costly indeed, what we picture is that a meta-evolution can produce effective frozen meta-skill, and is both efficient and effective as we show above. We have updated the full cost table in the Appendix.

  • W2/Q2: Isolation of methodological contribution.

This is a very keen observation! We'd like to answer this from two angles:

1, How does freeevolve compare to existing meta-evolution work like EvoX and HyperAgents?

You are correct. The advantage come from a broader control. I think its easier to illustatae via an example. (We will use EvoX as an example, but HyperAgent share a very similar mechanism and is comparable)

Let's say both FreeEvolve and EvoX conducted a similar inner evolution, they both identified that errors can usually be grouped into clusters, and focusing on solving individual clusters might be more beneficial then evaluating the whole dataset at once. And the following is what two frameworks would do and their consequences.

StageEvoXFreeEvolve
Between two inner evolutionsEvoX improves the search program based on what happened in the previous inner evolution. Suppose three clusters are identified in the inner evolution, it may add a clustering rule to form three clusters and a schedule that evaluates each cluster separately. These decisions are written into the search program before the next inner evolution starts.FreeEvolve learns a more general principle for handling clusters in the skill prompt: identify clusters from the current failures, allocate effort according to their difficulty and progress. It therefore learns how to use clusters, and will not assume cluster numbers, definition, or evaluation schedule for the next inner evolution.
The next inner evolutionThe same worker may now expose a different problem. For example, there may be six clusters instead of three. The predefined rule or default argument become sub-optimal and not tailored to the current problem. EvoX executes the search program produced between the two inner evolutions. It can handle the change only through behaviors already encoded in that program; otherwise the change can only be addressed by another meta-evolution step.The FreeEvolve skill observes the new evidence during the inner evolution. It can form six branches, merge or abandon clusters, change the evaluated subsets, and return to full evaluation when needed. The campaign can therefore change during the same inner evolution.

(Continues to the next reply)

The difference is not whether EvoX can discover a useful idea such as clustering; both methods can. The difference is how the idea is used in the next inner evolution. EvoX produces a search program whose behavior is determined before that evolution starts, while FreeEvolve learns a skill that decides how to apply the idea as evidence unfolds. This is not just a off-by-one-iteration issue. Previous meta-evolution work designs a fixed strategy based on past evolution experience and expected to work for a new evolution campaign, while FreeEvolve observe a principle, and how to use that is still adaptive during next evolution.

Of course, EvoX could also encode a very general adaptive agent-based controller in its search program, but this would move it toward the same design principle as FreeEvolve: giving the evolver control during the campaign.

2, How to use isolation to better single out the effect of broader control?

This is very critical, and we are also eager to study the impact of full control. This is why we conducted experiments in Table 4 as you also noted. We believe Table 4 shows partly that broader control is important, and we would like to extend the discussion here.

The isolation is not straightforward, because meta-evolution jointly changes how the evolver organizes the campaign and how it improves the worker. We use six settings to isolate in different aspects:

Evolution settingTau-3 mean (%)ARC-AGI-3 RHAE (%)
Seed Skill33.3 ± 5.361.7 ± 4.3
Seed Skill without control35.6 ± 2.060.4 ± 4.4
Meta-evolution using seed without control39.1 ± 7.262.2 ± 4.6
Meta skill without campaign-control guidance39.1 ± 2.063.4 ± 2.0
Meta skill without worker-improvement guidance44.8 ± 9.166.0 ± 8.8
Full meta-evolved skill47.1 ± 8.770.3 ± 3.2

All results in this table follow the paper protocol: one evolution or meta-evolution search, followed by three held-out evaluations of the returned worker. The standard deviation is across these evaluations, not across independent searches.

First is your suggested experiment (comparing row 1 and 2), we remove all controls knob from the seed skill, we can see that the performance is comparable. This is expected, as we already pointed out in the submission, the seed skill does not know how to use the control, and removing it would not make significant impact on the performance.

Then for the control-removed Seed skill, we conducted meta-evolution on it, and we found that meta-evolution was only minimally helpful in this case (comparing row 3 and 6). This shows that the broader control and flexibility is critical for the meta-evolver to come up with effective strategy that generalize.

To take the evaluation further, we use the meta-evolved skill, and ask an LLM to separate compaign control prompt from the evolution prompt, and we performance ablation study by doing evolution using them individually. We can see that removing either one of them would results in performane decrease, so evolution prompt and campaign control have orthogonal effect, but overall removing campaign-control results in larger performance drop, showing that this is a major factor for the effectiveness of freeevolve.

  • Cross model transfer.

Table 2 already shows the tranferrability across models, so in Table, not only the environment changes, the underlying model for evolver also switched from the source environment model to the target environment model. We have clarified this in the updated draft. The results in Table shows that evolver from opus 4.8 can transfer to both opus 4.7 and opus 5.

  • Performance drop in terminal-bench.

We looked into the actual evaluation logs, the two negative transfer results shows more TimeOut error, this is potentially because the the two source skill favors harness that consistently check its solution correctness which is costly in terminal-bench, and lead to degradation.

Note: we carefully looked into the work you mentioned in Q4, but we believe they are only remotely connected to our work. And seems like you are referring to our work as Scope Protection. We were wondering, maybe you accidentially put some of your review to another submission to ours? Would you mind kindly let us know if this question is indeed for us address?

Thank you again for your time and effort in helping us improve our work, and we look forward to hearing back from you!

Dear Reviewer A6ZR,

We would like to again sincerely thank you for your constructive comments and feedback, and see if there is any further clarification or discussion we can do with you to make our discussion more fruitful. A short summary of what we did: 1, we provide matched-budget results, showing FreeEvolve is both effective and efficient with learned campaign control. 2, we conducted a series of experiment isolating the impact of free-control. We discussed that complete isolation is difficult, but also shows that the broader control is essential and effective in multiple scenarios. Let us know if this answers your questions, and we would love to address any further concerns you may have.

Sincerely, Authors.

Reviewer_h3gKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing agent-evolution methods typically rely on hand-designed search procedures. FreeEvolve instead invokes a Claude Code-based evolver within a bounded environment, allowing it to autonomously determine task selection, repetition counts, concurrency, timeouts, and stopping criteria (Secs. 3.1–3.2). The evolver is further optimized through meta-evolution, in which candidate Markdown skills are evaluated based on the performance of workers across their respective campaigns (Sec. 3.3). Experiments on τ³-bench, ARC-AGI-2, ARC-AGI-3, and Terminal-Bench 2.1 demonstrate an average held-out improvement of 13.6 points (Table 1) and a 6.9-point gain over the seed skill when evaluated on fresh workers (Table 6). The paper also demonstrates transfer across environments (Table 2) and across models (Fig. 4a).

Critical strengths

The general idea of the paper is interesting, with comprehensive experimental results. The results look promising in terms of the improvement.

Critical weaknesses

Figure 1 is not clear, especially for the subfigures (a) and (b). It is not clear if the flow chart belongs to (a) or (b). The formula (1) is not well defined and explained. For example, what is execution request u? What is artifact bundle a? The execution request and artifact bundle only appear once. For the rest formulas, it is not clear what will be optimized?

Questions for authors

Please report total tokens and compute for all methods, and compare the frozen skill against the baselines at a matched budget.

2 comments

Dear Reviewer h3gK

Thank you very much for your thoughtful feedback that is very meaningful for us to improve our work. We address your concerns as follows:

  • Improving Figure 1

Following your suggestion, we have improved Figure 1 in the submitted pdf to separate prescribed loop and freeevolve, so the message we try to convey in each block does not confound each other. Please kindly let us know if this now appears clear.

  • Matched-budget comparison.

Following your suggestion, we further assess the evolvers under comparable or less total token budgets. We conduct three independent searches with the frozen meta-evolved FreeEvolve skill on Tau-3 and ARC-AGI-3. Every search starts from the same fresh worker and runs for 12 hours. We evaluate each returned worker three times and use its mean held-out performance as one search-level observation.

BenchmarkEvolverIndependent searchesHeld-out score (%)Total search tokens (B)
Tau-3FreeEvolve344.0 ± 1.31.53 ± 0.46
Tau-3GEPA336.0 ± 0.62.02 ± 0.02
Tau-3HyperAgents336.7 ± 0.82.02 ± 0.03
ARC-AGI-3FreeEvolve367.9 ± 3.11.12 ± 0.12
ARC-AGI-3GEPA362.7 ± 1.52.22 ± 0.05
ARC-AGI-3HyperAgents362.7 ± 3.82.24 ± 0.03

Here, the mean and standard deviation are computed across the three independent searches, rather than repeated evaluations of one worker. Total search tokens include every model call during evolution.

Here we can see, the evolved meta-skill is both more efficient and more effective. This is in fact very intuitive, under the same time budget, FreeEvolve can choose to focus on a subset for a specific problem, while GEPA and HyperAgents does not have such mechanism and have to evaluate the full set every time. So even when the ability to search and make changes is the same, FreeEvolve can be more efficient.

The total token costs and compute are now included in the table 9 of appendix.

  • Further explanation of formula 1.

We agree that the roles of uu and aa and consequently the optimization target in the later equations, were not sufficiently explicit. Our intention was to keep the formulation minimal: the central abstraction is simply that the environment exposes execution to the evolver rather than prescribing a search loop.

More concretely, an execution request uu specifies how a candidate xx should be assessed, including the selected data or task subset, repetitions, random seeds, concurrency, and timeout. This is an important notation, and we used it multiple times in the following fomulation. By design, we do not limit its definition and make it adaptible and extensible, like in line 215, we use uu to specify worker environments in the meta-evolution process. So uu changes as the environment changes.

The returned artifact bundle a=RE(x,u)a = R_{\mathcal E}(x,u) contains the resulting scores, trajectories, logs, execution status, and resource usage. The concept is simple, just anything/results from the execution function. We used the word artifact several times in the following text to represent this. After tt executions, the campaign history can be written as Ht=((x1,u1,a1),…,(xt,ut,at)).H_t = \bigl((x_1,u_1,a_1),\ldots,(x_t,u_t,a_t)\bigr). The evolution skill uses HtH_t to select the next request (xt+1,ut+1)(x_{t+1},u_{t+1}), or to stop and return a worker. Thus, uu is an action chosen by the evolver and aa is the evidence returned by the environment. We will carry this relationship explicitly through the revised description.

We will also state the optimization target directly. Let JE(x;D)J_{\mathcal E}(x;\mathcal D) be the environment-defined evaluation that operationalizes goal gg. Subject to the environment constraint CE\mathcal C_{\mathcal E}, the ideal worker target is:

\in \arg\max_{x \in \mathcal X_{\mathcal E}} J_{\mathcal E}(x;\mathcal D).$$ We denote the bounded evolution procedure as: $$x_{\mathrm{out}} = \mathrm{Evolve}(s;\mathcal E[x_0]).$$ It tries to solve this optimization problem and returns the best-supported candidate it discovers; it does not claim to find the global optimum. Let us know if this formalization clears the confusion, and we included this more formal version in the appendix, thank you very much for your rigorous review. Looking forward to your feedbacks!

Dear Reviewer h3gK,

We would like to thank you again for your critical feedback, and we would also like to check in and see if our response address your concerns. In particular, we provide a more formal version of the optimization problem and variable definition, and provided budget-matched experiment results. Let us know if you have any further question you would like to discuss. Looking forward to hearing back from you.

Sincerely, Author

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

Agent evolvers automate edits to the prompts, skills and workflows around an agent, but the search loop around them (which tasks to evaluate, how many repeats to run, which candidates to keep, when to stop) is still fixed by hand. FreeEvolve moves the goal, data, evaluator and resource limits into an environment and invokes the evolver, Claude Code, once; from then on the evolver decides what to run and when to stop, based on the evidence it has collected. It is guided by an evolution skill, a markdown instruction file. The hand-written seed skill only states general dispositions, and meta-evolution improves it by running complete worker campaigns with each candidate skill and judging the skill by the workers it produces.

On τ³-bench, ARC-AGI-2, ARC-AGI-3 and Terminal-Bench 2.1, the worker returned by meta-evolution improves the primary held-out metric by 13.6 points on average over the base agent and is at least as good as the hand-designed evolvers on six of the seven reported metrics (Table 1). A cleaner test installs each frozen skill on a fresh base worker with the same budget (Figure 3). There the seed skill improves the worker by only 0.6 points on average, and the meta-evolved skill adds a further 6.9. Control over the campaign gives little by itself; the gain comes from a skill learned on completed campaigns.

Critical strengths

The fresh-worker experiment is well designed. Three skills (the seed, an LLM rewrite that never sees an execution, and the meta-evolved skill) are installed on the same base worker with the same budget. The seed skill barely moves the worker, the meta-evolved skill adds 6.9 points on average and improves all three available pass@3 numbers, and the no-data rewrite lands between the two except on Terminal-Bench, where it is below the seed. This points to execution experience, not instruction rewriting, as the source of the gain. The evolver-model ablation is also informative: on τ³-bench it places a capability threshold between Haiku 4.5 and Sonnet 4.6, which tells practitioners what kind of model the method needs.

Critical weaknesses

As framed, Table 1 does not show that FreeEvolve matches or exceeds GEPA and HyperAgents. It is one search per method, and on the benchmark with the largest gap FreeEvolve had more than twice the baselines' wall-clock time, on five nodes. The abstract also mixes the returned worker with the frozen skill. The paper's central claim, that the policy organizing evolution can be learned, does not need Table 1, because Figure 3 supports it without any external baseline. I would rather see the abstract built on Figure 3 and Table 2, with Table 1 presented as a single-run comparison under unmatched compute. Whether FreeEvolve keeps its lead over the external evolvers is a separate question that only a matched comparison can answer.

Questions for authors

The comparison I would most like to see is a matched one: the same worker model and token budget, several independent searches, and the frozen meta-evolved skill against GEPA and HyperAgents. If FreeEvolve still leads on τ³-bench and ARC-AGI-3, I would raise my score. A tie would not lower it, but the abstract should then be built around Figure 3.

Smaller questions: What hardware did the baselines in Table 9 run on? Is the Table 1 worker the best inner worker found during meta-evolution, and by what rule was it chosen? Is the Terminal-Bench deficit against GEPA stable across repeated searches, or within the noise of a single run?

2 comments

Dear Reviewer GhQX,

Thank you very much for your positive feedback and insightful suggestion, and we appreciate this opportunity to extend our discussion here.

Regarding the weakness, we agree with you, Table 1 is not a strictly fair comparison. The purpose of Table 1, as we also suggested in line 315, is to show that given sufficient resource, ultimately how much better the worker can get using different evolver, so we try to compare the capability boundary and potential of the evolvers. Meanwhile, as you suggested, table 2 and figure 3 are actually more relevant to our main thesis, and we will first recenter the abstract around them.

We revised the abstract to center on the frozen-skill results:

When installed on fresh base workers under the same budget, the meta-evolved skills improve held-out performance by 6.9 points on average over the seed skill across four environments. The learned skills can also transfer to new environments, showing that the capability to organize evolution can be learned, frozen, and reused rather than remaining tied to the worker produced during meta-evolution.

Meanwhile, following your suggestion, we further assess the evolvers under comparable or less total token budgets against baselines. We conduct three independent searches with the frozen meta-evolved FreeEvolve skill on Tau-3 and ARC-AGI-3. Every search starts from the same fresh worker and runs for 12 hours. We evaluate each returned worker three times and use its mean held-out performance as one search-level observation.

BenchmarkEvolverIndependent searchesHeld-out score (%)Total search tokens (B)
Tau-3FreeEvolve344.0 ± 1.31.53 ± 0.46
Tau-3GEPA336.0 ± 0.62.02 ± 0.02
Tau-3HyperAgents336.7 ± 0.82.02 ± 0.03
ARC-AGI-3FreeEvolve367.9 ± 3.11.12 ± 0.12
ARC-AGI-3GEPA362.7 ± 1.52.22 ± 0.05
ARC-AGI-3HyperAgents362.7 ± 3.82.24 ± 0.03

Here, the mean and standard deviation are computed across the three independent searches, rather than repeated evaluations of one worker. Total search tokens include every model call during evolution.

Here we can see, the evolved meta-skill is both more efficient and more effective. This is in fact very intuitive, under the same time budget, FreeEvolve can choose to focus on a subset for a specific problem, while GEPA and HyperAgents does not have such mechanism and have to evaluate the full set every time. So even when the ability to search and make changes is the same, FreeEvolve can be more efficient.

Meta-evolution itself is more costly indeed, and we have updated the full cost table in the Appendix.

To answer your smaller questions:

  • They are ran on compute node with 192 cpu cores and 368 GB RAM.
  • Yes, choosing the best worker from meta-evolution is the same as choosing from non-meta-evolution. It is also determined by the evolver skill. For meta-evolution, we ask evolver to also find the best evolved worker in the environment specification, the evolver determines what does best mean. There is no fixed rules, but empirically we found that the meta-evolver tends to choose workers combining two factors 1, search set performence. 2, total number of evaluation using this worker inside its inner evolution.
  • Its comparable. Most methods perform similarly on Terminal-Bench, I think this is mostly because TB-2 has been published for a while, and claude 4.8 opus + base harness already saturates this benchmark.

Again, thank you very much for your critical comments, and we look forward to more discussion with you!

Dear Reviewer GhQX,

As the discussion period is coming to an end, we would like to check in and see if our response addressed your concern and see if you have other questions/discussion we can take. In our responses and update, we shifted the focus to the learnability of campaign, while also provided budget matched experiments with independent search on both baseline and FreeEvolve, and we show that meta-evolved FreeEvolve skill is in fact more budget-efficient. Looking forward to your reply!

Sincerely, Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.