acceptodds
Under review as a conference paper at ICLR 2027

Wright–G\"odel Machine: Mitigating Homogenization in Self-Improving Agents

11 views

Abstract

Recent self-improving agents retain previously discovered implementations in an archive so that improvement can proceed from multiple starting points. However, we find that parent selection can repeatedly favor closely related agents, leaving alternative designs with few opportunities to produce descendants. We refer to this phenomenon as genetic homogenization, which can cause the search to miss useful directions for improvement. We characterize it through the effective number of reproductive lineages and parent diversity, and prove that Thompson-sampling parent selection concentrates reproduction on a single lineage in the standard stochastic bandit model, even when every lineage remains in the archive. Inspired by Wright's shifting-balance theory, we propose the Wright–G\"odel Machine (WGM). WGM selects parents with a rule based on the Exponential-weight algorithm for Exploration and Exploitation (EXP3), which combines clade statistics with a uniform exploration floor, and uses best arm identification (BAI) to allocate additional evaluations and select the final agent. We prove that WGM gives every lineage a polynomially growing expected number of expansions under any reward sequence. On SWE-bench Verified and Polyglot, WGM discovers agents that outperform state-of-the-art methods under matched evaluation budgets, and these agents transfer to held-out coding tasks and other backbone models.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Self-Improving AI, Coding Agents, G\"odel Machine

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LU9GOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a new approach for how to find a best-performing self-improving agents when expanding different possibilities and keeping promising agents in an archive. This new approach is called Wright–Gödel Machine (WGM) and innovates on existing approaches in two main ways. First, WGM determines which agent to expand based on a combination of the EXP3 algorithm and a uniform selection probability. Second, for determining the final agent, WGM applies best-arm identification.

The paper proves mathematically that, under certain assumptions, WGM prevents the lineage of the most promising candidate to take over the selection probability. Moreover, the benefit of WGM is shown empirically, testing it against two competing approaches. These experiments show that WGM outperforms the competition, and an additional ablation study shows the benefit of each of the two innovations of WGM.

Critical strengths
  1. Ablation study. This paper investigates the benefit of its two main innovations via an ablation study. Each one of them individually already improves the overall performance, and the combination is even more beneficial. This suggests that each building block has already its own merit.

  2. Simple design. WGM follows rather simple ideas, which are, for the most part, clearly presented. This makes it easy to reason about the benefits of this approach.

Honorable mention. The paper provides a well-documented code repository.

Critical weaknesses
  1. Multiple parameters. WGM comes with a fair amount of parameters. The empirical setup fixes all of them without elaborating the choices deeply. In particular, the choice of the exploration share γ\gamma, which seems to be the most crucial parameter, at least according to the theoretical results, is not stated anywhere in the paper, to the best of my knowledge. This makes it hard to assess how robust WGM is with respect to its parameter choices. The theoretical guarantees mostly make statements about the limit behavior, which is not too useful for practical applications.

  2. Partially challenging presentation. The paper works with global variables, that are introduced in one place and then used throughout the paper. This makes retrieving vital information challenging, as if one forgets the meaning of one of the several variables, one needs to start searching for where it was defined. Moreover, certain pieces of information, such as the importance of Assumption 2 for the main result (Theorem 1) or additional experiment setup are only mentioned in the appendix, making the paper less self-contained.

Honorable mention. The code repository contains a license file that seems to violate the double-blind submission guidelines, as it appears to contain author-revealing information. This point is not that important to me, but I wanted to point it out and advise the authors to be more careful in the future.

Questions for authors
  1. How do the authors intend to address my points listed as critical weaknesses? Please ignore the honorable mention for this question.

  2. How important is the uniform selection floor for WGM? To me, it seems crucial for obtaining a sufficient diversity. Given that γ\gamma is a constant value throughout the run, it appears as if it has a great impact. Can the authors elaborate on how important the exponential scaling in this context really is?

  3. How realistic is Assumption 2? The discussion in the appendix does not answer this question, in my opinion. Similarly, since Section B.4 shows how Assumption 2 can, in some sense, be replaced by Assumption 3, how realistic is this assumption?

  4. Can the authors elaborate on the choice of the parameters? For example, all experiments are repeated three times. This sounds like a small number. Is this sufficient in order to obtain conclusive evidence for the benefit of WGM?

Further comments. The following comments are just a suggestion in order to improve the quality of the paper. They do not need to be addressed, and they did not affect my rating.

  • The table captions should mention what bold highlights mean. Interestingly, in Table 1, not all columns feature a bold entry.

  • The references have issues with the capitalization. Several words that should be in uppercase are not, and the casing of the titles is mixed. I recommend polishing the references.


Update During the Rebuttal

The authors have addressed my concerns well, so I increased my rating.

4 comments

Dear Reviewer LU9G,

Thank you for your thoughtful review. We appreciate the opportunity to clarify several points and address the concerns raised.

W1 & Q4: WGM has multiple parameters, and the exploration share γ\gamma is not stated. The theory mostly concerns limit behavior. Are three seeds sufficient?

A: Thank you for the catch. We did not tune hyperparameters for WGM. We use γ=0.1\gamma=0.1 in all experiments and have added it to Section 4.1(Line 402). All methods share the same initial agent, prompts, and backbone, adopted from the official DGM/HGM implementation, and the remaining settings are standard choices (z=1.96z=1.96; ηt=γ/Nt\eta_t=\gamma/N_t as in EXP3 [3]). Due to limited time and budget, we vary γ\gamma on Polyglot with Sonnet 5 as the backbone and a budget of 200 task evaluations:

γ\gamma0.010.10.3
Acc. (%)44.045.840.9

WGM remains effective across different values of γ\gamma. We would also like to clarify that our guarantees are not only asymptotic. Eq. B.4 holds at every step, and Proposition 2 gives finite-TT bounds. The theory explains the mechanism, namely polynomial versus O(log⁡T)O(\log T) expansions under Thompson sampling, and Figure 1(b) confirms it at a finite horizon.

On seeds, each run is costly, so many repetitions are infeasible. Three seeds already exceed common practice in this line of work. HGM [1] does not report repeated runs, and DGM [2] reports its main SWE-bench results without repeated runs and uses three runs only as a stability check on Polyglot.


W2 & Q1: Variables are defined once and reused throughout the paper. The role of Assumption 2 and parts of the setup appear only in the appendix.

A: Thank you for pointing this out. Our intention was to define the notation upfront so that the subsequent exposition would not be interrupted by definitions. We agree that this can make retrieval harder. In the revised version, we have restated the meaning of key symbols where they are used, such as the archive size NN (Line 276) and Assumption 2.

Due to the page limit, we originally placed the detailed discussion of Assumption 2 and the additional experimental setup in the appendix. The main text already states that Theorem 1 relies on Assumption 2. In the revised version, we have moved the hyperparameters into Section 4.1 (Line 402).


Q2: How important are the uniform floor and the exponential scaling?

A: Thank you for the question. We agree that the uniform floor is crucial for diversity: our theory relies only on the floor (Eq. B.4). The exponential weights play a complementary role, steering the remaining 1−γ1-\gamma probability toward agents whose expansions produced better children. The sensitivity results in W1 show that a moderate γ\gamma performs best: a very small floor reduces diversity, while a large floor leaves less probability for the exponential weights. Moreover, replacing the exponential weights with Thompson sampling (w/o EXP3, Table 3) lowers accuracy from 54.7% to 51.1% on Polyglot. Thus, both components contribute.


Q3: How realistic are Assumptions 2 and 3?

A: Thank you for the question. We would first like to clarify that some condition on the archive is unavoidable, because L(T)=1L(T)=1 under any selection rule if the archive holds a single lineage (Lines 849–852). Both assumptions are mild in practice.

  • Assumption 2 only requires two lineages that each keep a constant fraction κ\kappa of the archive, rather than balance across all lineages. With the lineage distance dlind_{\mathrm{lin}} used in our experiments, condition (ii) follows from condition (i) with δ=κ\delta=\kappa. This is realistic for WGM, since initialization creates multiple lineages, archived agents are never removed, and the exploration floor keeps expanding every lineage.
  • Assumption 3 only requires that self-modification within a lineage does not always fail. This is realistic because admission does not require the child to outperform its parent (Lines 140), so executable children can enter the archive.

Theorem 1 uses Assumption 2 to show that WGM avoids homogenization, while Proposition 2 uses only Assumption 3, which places no condition on the archive or the rewards, to bound how fast reproduction can concentrate. The two results are therefore complementary.


Further comments: bold highlights in table captions, reference capitalization, and license file.

A: Thank you for the careful reading and the helpful suggestions.

  • Bold entries. We bold only the best accuracy results. We have clarified this in the table captions and made the corresponding description in Line 377 more precise.
  • Reference capitalization. Thank you for the reminder. The current references were exported directly from Google Scholar, which explains the inconsistent capitalization.
  • License file. The license file is inherited from the open-source DGM codebase [2] and names the DGM authors, not ours, so it does not reveal our identities. We will be more careful in the future, and we have also reviewed our code and updated the anonymous repository.

We sincerely appreciate the time and effort you devoted to reviewing our paper, and your comments have helped us improve it considerably. If our responses address your concerns, we would be grateful for your consideration of a higher score.

Best,

The Authors

1.Wang et al. Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine. ICLR 2026.

2.Zhang et al. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. ICLR 2026.

3.Auer et al. The Nonstochastic Multiarmed Bandit Problem. SIAM Journal on Computing, 2002.

I thank the authors for their succinct and informative feedback as well as for their clarifications. Their response addresses all of my questions and concerns satisfactorily. My only remaining complaint is that I would have preferred it if the authors had highlighted the changes to the previous PDF clearly, as this makes it easier to check where an update occurred.

In any case, I am happy to increase my score. I will go with a weak accept because I am not too familiar with this research area in general.

Dear Reviewer LU9G,

Thank you for your positive feedback and for increasing your score. We are delighted that our responses have addressed your concerns. Following your suggestion, we have uploaded a revised PDF in which all changes are highlighted in blue. We sincerely appreciate your time and constructive comments, which helped improve our paper.

Best,

The Authors

Reviewer_kUWcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes the Wright-Godel Machine for self-improving agents. The design is loosely motivated by Wrights shifting balance theory from evolutionary theory. The idea is to preserve diversity of agents, maintain an archive of agents, and to employ EXP3 to balance further exploration of the agents, leading to lineages. Some theoretical results are provided, including a claimed proof that every lineage grows polynomially. Some limited experimental comparisons are made with alternative approaches, including HGM.

(the scores have been updated following the rebuttal phase)

Critical strengths

The paper claims that the proposed method outperforms existing approaches on a set of standard benchmarks (however I am unsure whether this claim is supported by the experimental results).

The paper contains few spelling and grammatical mistakes.

Critical weaknesses

The presentation of the method and accompanying theory is below acceptable standard. In particular, several parts of the presentation uses ambiguous notation. According to the definition in Section 3.2, the only way to understand a lineage c is that it corresponds to a subtree rooted in a descendant c of the founder agent a0. This deviates from the use of the term lineage in evolutionary theory, where a lineage is simply a path in the family tree. The notation used in Definition 2 is not defined. Furthermore, the definition uses the term "self-improvement process" which is not well-defined.

The experimental validation is unsatisfactory with limited or no statistical evaluation. The plots do not contain any information about variance in the experimental data. There are no experimental results indicating the impact of the hyper-parameters discussed in Appendix C.1.

The pseudo-code should be stated in the main part of the paper, and not in the appendix.

Questions for authors

Section 3.2 seems to assume that the lineages are chosen uniformly at random. Is this correct?

4 comments

Dear Reviewer kUWc, We really appreciate the reviewer for the detailed and constructive feedback. Below, we provide detailed responses to all the comments.

W1: According to Section 3.2, a lineage c can only be understood as a subtree rooted in a descendant c of the founder a0. This deviates from evolutionary theory, where a lineage is a path in the family tree.

A: Thank you for your careful reading. We want to clarify that by Definition 1, lineage cc is the clade of a founder cc, which is a child of a0a_0. This is not a deviation, because "lineage" commonly denotes all descendants of a common ancestor, as in cell lineages in developmental biology and SARS-CoV-2 Pango lineages in genomic epidemiology [1]. It is also consistent with HGM [2], which builds on Huxley's notion of clades as lineages of common ancestry. Our analysis requires this definition: L(T)L(T) (Eq. 3) requires lineages to partition the archive, so that the shares of parent selections across lineages, ∑cqc(T)=1\sum_c q_c(T)=1. We have clarified this more clearly in Definition 1 of the revised version.


W2: The notation used in Definition 2 is not defined. The term "self-improvement process" is not well-defined.

A: We would like to clarify that all quantities in Definition 2 are defined before it: nc(T)n_c(T) and qc(T)=nc(T)/Tq_c(T)=n_c(T)/T in Lines 212, and L(T)L(T) in Eq. 3. →p\xrightarrow{p} denotes convergence in probability. The self-improvement process is introduced in Lines 44–49 and formalized in Section 2.1. At each expansion tt, a parent ata_t is selected and modified to produce a child btb_t, which is evaluated and added to the archive if admitted. This follows HGM [2], which formalizes the self-improvement process as an iterative tree search. We have made this process more clearly in Section 2.1 (Line 140) of the revised version.


W3: The experimental validation has limited or no statistical evaluation. The plots do not contain any information about variance.

A: Thank you for the suggestion. We want to clarify each run is costly, so running many seeds is infeasible. Three seeds already exceed common practice in this line of work: HGM [2] does not report repeated runs, and DGM [3] only uses three runs only as a stability check on Polyglot. Below, we report results across three random seeds as mean ± standard deviation, in percentage points. The main results are shown below.

MethodSWE-Verified-60 (%)Polyglot (%)
DGM58.3 ± 1.3647.1 ± 0.96
HGM61.7 ± 2.3650.2 ± 0.63
WGM65.0 ± 3.6054.7 ± 1.31

For the generalization experiment on SWE-Lite, with GPT-5.6 as the execution model, SWE-agent achieves 50.4 ± 0.99% (Filtered) and 58.7 ± 2.23% (Standard), HGM's best-belief agent achieves 51.2 ± 1.42% and 59.3 ± 1.19%, and WGM's final agent achieves 51.9 ± 2.96% and 60.4 ± 1.13%. With Claude Opus 5.5, SWE-agent achieves 52.3 ± 1.64% and 60.7 ± 0.72%, HGM's best-belief agent achieves 53.0 ± 2.38% and 61.4 ± 1.50%, and WGM's final agent achieves 53.6 ± 1.81% and 62.1 ± 1.64%. For the ablation study, WGM (w/o BAI) achieves 63.3 ± 4.91% on SWE-Verified-60 and 52.4 ± 1.09% on Polyglot, and WGM (w/o EXP3) achieves 62.2 ± 3.42% and 51.1 ± 0.96%, respectively. The standard deviations are of the same order across all methods, indicating comparable run-to-run variability.


W4: There are no experimental results indicating the impact of the hyper-parameters discussed in Appendix C.1.

A: Thank you for the suggestion. In all our experiments, we use the same initial agent, prompts, and backbone, adopted from the official DGM/HGM implementation, and the remaining settings are standard choices (z=1.96z=1.96; ηt=γ/Nt\eta_t=\gamma/N_t as in EXP3 [4]). Thus, the improvements come from parent selection and evaluation rather than hyperparameter tuning. Moreover, our guarantees (Theorem 1, Proposition 2) hold for any choice of these hyperparameters. Due to limited time and budget, we further vary the exploration share γ\gamma on Polyglot with Claude Sonnet 5 as the backbone and a budget of 200 task evaluations:

γ\gamma0.010.10.3
Acc.(%)44.045.840.9

WGM remains effective across different values of γ\gamma, and the default performs best.

W5: The pseudo-code should be stated in the main part of the paper, and not in the appendix.

A: Thank you for the suggestion. We originally placed pseudo-code in the appendix due to the page limit. We have now added pseudo-code to Section 3.5.


Q1: Section 3.2 seems to assume that the lineages are chosen uniformly at random. Is this correct?

A: Thank you for the question. We would like to clarify that Section 3.2 does not assume uniform selection of lineages. The sentence before Eq. 3 ("when selecting two parents ... the probability that they belong to the same lineage") describes how L(T)L(T) is measured: two of the TT recorded parent selections are drawn at random, as in Simpson's index in ecology. It is not a selection rule. The selection rules are Thompson sampling (Proposition 1) and WGM's Eq. 8 (Theorem 1, Proposition 2). In Eq. 7, the exploration share is uniform over agents, not lineages, so a lineage with more archived agents receives a larger share. The rest follows the EXP3 weights and clade statistics.


Thank you again for your valuable suggestions and questions. If these additional results and clarifications address your concerns, we would sincerely appreciate your consideration of a higher rating. Best, The Authors

1.Rambaut et al. A dynamic nomenclature proposal for SARS-CoV-2 lineages to assist genomic epidemiology. Nature Microbiology, 2020.

2.Wang et al. Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine. ICLR 2026.

3.Zhang et al. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. ICLR 2026.

4.Auer et al. The Nonstochastic Multiarmed Bandit Problem. SIAM Journal on Computing, 2002.

The term lineage is used differently within evolutionary theory, some refer to lineages strictly as paths within the evolutionary tree, while others talk about descendants. In order not to confuse the reader, I suggest to either use the more precise term clade, or explain that the term lineage has different meanings.

Every term in a mathematical statement needs to be formally defined. You describe the term self-improvement process informally in line 139 in the revised paper. However, it is still not exactly exactly clear what this refers to ? Do you mean a_t, A_t, b_t, or perhaps all, or something else? Please replace the term self-improvement process in Definition 2 with the exact mathematical object.

Reporting standard deviation when the sample size is N=3 is not so meaningful. Perhaps you could just state all three values?

Thank you for adding the pseudo-code. It should be possible to reimplement the algorithm from the pseudo-code. However, line 5 and 6 only vaguely explain what is going on.

Q1. Thank you for clarifying this point. Adding pseudo-code improved the presentation, but the pseudo-code itself need to be improved (see above).

Based on these changes, I will update my score.

Dear Reviewer kUWc,

Thank you very much for your follow-up comments and for updating your score. We have revised the paper accordingly, with all changes highlighted in blue.

  • Lineage: Following your suggestion, we now explain that the term lineage sometimes refers to a single path in the family tree, and clarify that in this paper it denotes all descendants of a common ancestor (Line 212).

  • Self-improvement process: We now formally define the self-improvement process as the random sequence (at,bt,At+1)t≥1(a_t, b_t, \mathcal{A}_{t+1})_{t\ge1} of parents, children, and archives (Line 138), so that Definition 2 refers to this exact object.

  • All results: Following your suggestion, we now report all three values. On SWE-Verified-60, the three seeds give 56.7%, 58.3%, and 60.0% for DGM; 60.0%, 60.0%, and 65.0% for HGM; and 61.7%, 63.3%, and 70.0% for WGM. On Polyglot, they give 46.2%, 46.7%, and 48.4% for DGM; 49.8%, 49.8%, and 51.1% for HGM; and 52.9%, 55.1%, and 56.0% for WGM. For the ablation study, WGM (w/o BAI) achieves 58.3%, 61.7%, and 70.0% on SWE-Verified-60 and 51.1%, 52.4%, and 53.8% on Polyglot, while WGM (w/o EXP3) achieves 58.3%, 61.7%, and 66.7%, and 50.2%, 50.7%, and 52.4%, respectively. For the generalization experiment on SWE-Lite with GPT-5.6, SWE-agent achieves 49.3%, 50.2%, and 51.7% (Filtered) and 56.3%, 58.0%, and 61.7% (Standard); HGM's best-belief agent achieves 49.3%, 51.7%, and 52.7%, and 57.7%, 60.0%, and 60.3%; and WGM's final agent achieves 48.3%, 51.7%, and 55.6%, and 59.3%, 60.0%, and 62.0%. With Claude Opus 5.5, SWE-agent achieves 50.7%, 51.7%, and 54.6% (Filtered) and 59.7%, 61.0%, and 61.3% (Standard); HGM's best-belief agent achieves 50.2%, 52.7%, and 56.0%, and 59.3%, 62.3%, and 62.7%; and WGM's final agent achieves 51.2%, 54.1%, and 55.6%, and 60.0%, 62.3%, and 64.0%.

Q1: The pseudo-code could be more specific in lines 5 and 6.

A: Thank you for the suggestion. We have revised lines 5 and 6 of Algorithm 1 to describe the archive update and the BAI step more explicitly.


Thank you again for your careful reading and constructive suggestions, which have helped us improve the paper.

Best,

The Authors

Reviewer_CCjEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript introduces the Wright–Gödel Machine (WGM), which draws inspiration from Wright's shifting balance theory by preserving alternative evolutionary trajectories to reduce premature convergence. The central idea behind WGM is to balance exploiting promising agent designs with continuing to explore less successful designs that might produce better descendants. It combines performance-guided parent selection, guaranteed exploration of alternative agents, and uncertainty-aware evaluation to prevent the self-improvement process from concentrating heavily on a small number of lineages.

The algorithmic mechanisms rely on already established bandit methods, EXP3 and Best Arm Identification (BAI):

  • EXP3 is used for the parent selection and exploration.
  • While BAI is used for allocating the evaluations and selecting the final agent.

The proposed method is an archive-based self-improvement framework that iteratively uses LLMs to discover better coding agents. It aims to prevent genetic homogenization. EXP3 combines historical performance information from an agent and its descendants to ensure that every archived agent retains an opportunity to be selected. After this, an LLM modifies its implementation using diagnostic feedback to generate a new child agent, which is subsequently evaluated on coding tasks. Then the BAI allocates additional evaluations to promising or uncertain agents and identifies the best final agent. The evaluation results update the archive and also influence subsequent parent selections, allowing the process to repeat. Through this combination of exploration-aware parent selection, LLM-driven self-modification, and uncertainty-aware evaluation, WGM seeks to discover stronger coding agents under a fixed evaluation budget.

The manuscript theoretically analyzes genetic homogenization under Thompson-sampling-based parent selection and establishes guarantees on continued exploration under WGM's proposed selection mechanism. The experimental evaluation compares WGM against DGM and HGM on SWE-bench Verified and Polyglot under matched evaluation budgets. WGM achieves reported accuracies of 65.0% and 54.7%, respectively, outperforming the evaluated baselines.

Critical strengths
  1. Simple and interpretable framework. WGM addresses genetic homogenization in archive-based self-improving agents using EXP3 and BAI. This approach is conceptually straightforward and can potentially be integrated into existing archive-based self-improvement systems.
  2. Strong empirical performance under matched evaluation budgets. WGM achieves the highest reported accuracy on both SWE-Verified-60 and Polyglot, and the ablation experiments indicate that EXP3 and BAI each contribute to performance.
  3. Transferability across coding tasks and LLM backbones. The final agents by WGM retain their reported performance advantage under both GPT-5.6 and Claude Opus 5.5 execution backbones. This provides evidence that the resulting agent improvements are not restricted to the original task set or execution model.
Critical weaknesses

1. Limited breadth of the main comparative evaluation:

The principal experiments compare only WGM against DGM and HGM on SWE-Verified-60 and Polyglot. The proposed method achieves 3.3 percentage points over HGM on the SWE-Verified-60 benchmark and 4.5 percentage points on the Polyglot benchmark. The manuscript also includes the full 500-task SWE-bench Verified dataset, followed by an evaluation on SWE-Lite. However, this only compares WGM against HGM, with its primary purpose being to evaluate the transfer to tasks and alternative execution models. This raises the concern that the manuscript's most comprehensive algorithmic comparisons are concentrated in two evaluation settings, which limits confidence in concluding that WGM's advantages persist across substantially different self-improvement tasks and environments.


2. High reproduction cost and limited accessibility

The manuscript itself provides the cost of experiments in lines 1052-1056, which is about $12,000 USD. Since there is no cost breakdown for things like API credits, it is difficult to determine how much of the costs were spent on the frontier models used. The manuscript does not investigate running the self-improvement procedure itself using open-weight LLMs. Therefore, this creates two related limitations:

  • Independent reproduction of the entire evaluation requires substantial computational and financial resources.
  • Whether WGM's performance advantages persist with less expensive or open-weight models used to generate candidate modifications remains ambiguous.

3. Insufficient evidence about the reliability of the reported improvements

The manuscript states that all experiments are repeated with three random seeds, but Tables 1 and 3 (pages 7 and 9, respectively) report only mean accuracies. However, without per-seed outcomes or uncertainty estimates, it is difficult to judge whether the improvement is consistently observed.

Questions for authors

I seek clarification on the breadth of the experimental evaluation, the reproducibility costs, and the reliability of the reported improvements.

  1. I want to explore if WGM's advantages persist across additional self-improvement tasks or environments, particularly under comparisons with both DGM and HGM.
  2. Would it be possible to get a breakdown of the computational and API-related costs associated with the experiment? Additionally, how does WGM's self-improvement procedure perform with open-weight LLMs?
  3. Can the authors provide per-seed results, standard deviations, or confidence intervals to clarify the variability and consistency of WGM's reported improvements over the baselines?
3 comments

Dear Reviewer CCjE,

Thank you for your thoughtful review. We appreciate the opportunity to clarify several points and address the concerns.

W1 & Q1: The main comparison is concentrated on SWE-Verified-60 and Polyglot. Do WGM's advantages persist across additional self-improvement tasks or environments?

A: Thank you for the question. We would first like to clarify that our evaluation is not limited to two benchmarks. Beyond SWE-Verified-60 and Polyglot, which cover repository-level software engineering and multilingual programming, we further evaluated WGM on the full SWE-bench Verified and SWE-Lite against HGM (Table 2). WGM achieves the highest accuracy in all settings.

We also conducted experiments with an open-weight LLM for self-modification, in which WGM outperforms (see Q2). Due to limited computational resources and the short rebuttal period, we wll leave the extension of WGM to further self-improvement tasks for future work.


W2 & Q2: Please provide a cost breakdown. How does WGM perform with open-weight LLMs?

A: Thank you for the question. Since our experiments were run on a shared account, we can only provide approximate costs. In total, WGM cost about USD 5,000, HGM about USD 4,100 and DGM about USD 2,700, with the remaining cost spent on evaluating the SWE-agent baseline. These figures reflect discounted pricing on our shared account, so reproduction at standard API rates may cost more. For open-weight models, WGM is model-agnostic, since its selection rule only consumes task outcomes. Due to limited time and budget, we re-run Polyglot with the open-weight Qwen3-Coder-30B-A3B-Instruct for self-modification, using a budget of 200 task evaluations:

MethodAcc. (%)
DGM28.9
HGM35.6
WGM37.3

WGM outperforms HGM and DGM, showing that its advantage does not rely on frontier models.


W3 & Q3: Can the authors provide standard deviations, or confidence intervals?

A3: Thank you for the suggestion. Below, we report results across three random seeds as mean ± standard deviation, in percentage points. The main results are shown below.

MethodSWE-Verified-60 (%)Polyglot (%)
DGM58.3 ± 1.3647.1 ± 0.96
HGM61.7 ± 2.3650.2 ± 0.63
WGM65.0 ± 3.6054.7 ± 1.31

For the generalization experiment on SWE-Lite, with GPT-5.6 as the execution model, SWE-agent achieves 50.4 ± 0.99% (Filtered) and 58.7 ± 2.23% (Standard), HGM's best-belief agent achieves 51.2 ± 1.42% and 59.3 ± 1.19%, and WGM's final agent achieves 51.9 ± 2.96% and 60.4 ± 1.13%. With Claude Opus 5.5, SWE-agent achieves 52.3 ± 1.64% and 60.7 ± 0.72%, HGM's best-belief agent achieves 53.0 ± 2.38% and 61.4 ± 1.50%, and WGM's final agent achieves 53.6 ± 1.81% and 62.1 ± 1.64%. For the ablation study, WGM (w/o BAI) achieves 63.3 ± 4.91% on SWE-Verified-60 and 52.4 ± 1.09% on Polyglot, and WGM (w/o EXP3) achieves 62.2 ± 3.42% and 51.1 ± 0.96%, respectively. The standard deviations are of the same order across all methods, indicating comparable run-to-run variability.

We are grateful for your insightful comments and helpful suggestions. We hope our responses and new results have clarified the points you raised. If you find them satisfactory, we would greatly appreciate a reconsideration of your score.

Best,

The Authors

1.Wang et al. Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine. ICLR 2026.

2.Zhang et al. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. ICLR 2026.

I thank the authors for their diligent response, and I appreciate their efforts to address my concerns. I'm afraid that only one of my concerns has been fully addressed, while the others remain partially addressed.

Regarding evaluation settings: I acknowledge that the manuscript includes evaluations beyond SWE-Verified-60 and Polyglot, as already noted in my review. However, I'm concerned about the diversity of the underlying evaluated tasks and the breadth of comparative experiments rather than the number of benchmarks evaluated. As far as I understand, SWE-Verified-60, SWE-bench Verified, and SWE-Lite belong to the same SWE-bench family, which predominantly evaluates Python-based repository-level issue resolution. Therefore, I believe that evaluation on these benchmarks does not necessarily demonstrate generalization to different self-improvement tasks and environments. I agree that Polyglot provides diversity through multilingual programming tasks, but the most comprehensive algorithmic comparisons remain concentrated on the original two settings. An evaluation of WGM on benchmarks such as Terminal-Bench [1] will be beneficial.

Reproducibility cost and open-weight language models: I'm grateful that the authors provide an estimated cost breakdown and an additional Qwen3-Coder-30B-A3B-Instruct run on Polyglot. I understand the limitations the authors listed in their rebuttal, but I'll emphasize that the provided runs don't demonstrate WGM's broad generalizability. These runs are on a single evaluation setting, on a single open-weight language model. I'm not overlooking WGM's performance, and I believe that these results are very promising.

Standard deviation and CIs: The rebuttal fully addresses this concern, and I thank the authors for it. I recommend that the authors add a discussion about it in the manuscript.

[1] Merrill, M. A., et al. (2026). Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces. ICLR 2026. https://openreview.net/forum?id=a7Qa4CcHak


Regarding scores: Once again, I highly appreciate the authors addressing my concerns, and I reconsider my score to a weak accept because WGM makes a meaningful contribution. Since the majority of my concerns remain partially unaddressed, I'll refrain from giving this work a clear accept.

Best,

Reviewer CCjE

Dear Reviewer CCjE,

Thank you very much for raising your score and for your constructive follow-up. Following your suggestion, we have started adapting our setup to evaluate WGM on Terminal-Bench 2.0 [1]. Due to the limited time of the discussion period, these runs have not yet produced results. We will continue these experiments and include the results in the revised paper. We will also add a discussion of run-to-run variability as you recommended. Thank you again for your careful review and valuable suggestions, which have helped us improve the paper.

Best,

The Authors

[1] Merrill et al. Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces. ICLR, 2026.

Comments on the paper

1 comment

We sincerely thank the reviewers for their constructive feedback. We are encouraged that the reviewers recognized the simplicity and interpretability of WGM, its strong empirical performance, and its transferability across tasks and backbones. In response to the reviewers' comments, we conducted additional experiments:

  • Open-weight LLM (CCjE): With Qwen3-Coder-30B-A3B-Instruct, WGM outperforms previous works.
  • Hyperparameter sensitivity (kUWc, LU9G): WGM remains effective across different values of .

Beyond the new experiments, key revisions include:

Question Answering and Clarification

  • Clarified the lineage definition, the self-improvement process, and that Section 3.2 does not assume uniform selection (kUWc).
  • Explained that our guarantees hold at finite horizons and that both assumptions are mild in practice (LU9G).

Presentation Improvements

  • Added pseudo-code to Section 3.5 and moved the hyperparameters to Section 4.1 (kUWc, LU9G).
  • Restated key symbols where they are used (LU9G).

All changes are highlighted in blue. We believe these experiments and clarifications have addressed the reviewers' concerns, and we thank the reviewers again for helping improve this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.