Artificial Intelligence · arXiv:2609.35328 v3 · submitted 28 Sep 2026
Hyper Algorithm Design Agent: Evolving Learnable Optimizer from Zero
The paper says
“Meta-Black-Box Optimization (MetaBBO) is one of the highlights in the recent AI for Optimization trend. This paradigm's bi-level workflow leverages the learnable algorithm design policy at meta level to ensure the performance and generalization improvement on the low-level optimization task.”
more of the abstract
While MetaBBO helps advance the performance lower bound of the resulted optimization system, it is currently handcrafted and customized case by case to adapt different optimization problems, which inevitably introduces inherent subjectivity and hence restricts the performance upper bound and usability in practice. In this paper, we address this issue by regarding MetaBBO's design loop as coding task, where we could introduce openendedness into MetaBBO with recursive self-improvement capability of advanced coding agents. Specifically, we propose a dual-agent framework: i) a task agent continuously refines the codebase of a target MetaBBO approach through code evolution; ii) a hyper agent progressively modifies the task agent and itself to provide open-ended design behavior; iii) the evolved MetaBBO codebase is evaluated and all in-execution information is fed back to the agents for recursive self-referential improvement. As a result, given a naive MetaBBO template, our framework automates a design evolution and finds novel variants superior to up-to-date human-made MetaBBO baselines. Surprisingly, the experimental results also demonstrate that our framework supports fast adaption across different optimization domains. Solid interpretation analysis further reveals interesting design principles emerge in such open-ended process. This work serves as the first exploration on automating design of complex learning-assisted optimization algorithms.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 29 Sep 2026 · of version 1 of the paper; version 3 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A real engineering result on LLM-driven MetaBBO design with genuine held-out gains, but the headline percentages and the unmeasured proxy-to-official concordance need fixing before the strongest claims are credible. This line is wrong
4 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the references say.
The paper says, at Abstract and §4.2 (sentences b27, b107)
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The claimed '121.7%, 114.7% and 104.3% performance leap' cannot be reproduced from the reported tables. Table 2 gives HADA 0.9674 vs GLEET 0.8942, which is an approximately 8% relative improvement, not 121.7%; no arithmetic in the paper recovers the quoted figure.
more of objection 1
Sentence b107 is additionally internally inconsistent, saying 'using the worst-performing CMAES as baseline, HADA improves upon the SOTA MetaBBO (GLEET) by 121.7%' — it is unclear whether the 121.7% refers to a comparison against CMAES, against GLEET, or against a different reference point, and the definition of 'performance leap' is never given. Since these figures appear in the abstract and the contribution list, they must either be defined precisely (e.g., as an explicit relative-improvement formula) and matched to the tables, or removed and replaced with the recoverable numbers.
The paper says, at §4.1 and §4.4
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
Parent selection throughout the 100-generation evolution, and the final choice AET.optimal(), are driven exclusively by proxy scores from 5-epoch training and 5 test runs, while the reported headline results (Tables 1-2) use 20-epoch training and 10 runs. The paper never measures the concordance between the proxy ranking and the official ranking, so the officially reported score of the selected node may reflect winner's-curse selection bias rather than a genuinely superior design.
more of objection 2
The paper's own Fig. 3 hints at this: the best proxy node in the tree is #63 with 0.971, yet the reported official average is 0.9674, i.e., a small but nonzero regression. I request a proxy-to-official concordance analysis (e.g., officially re-evaluating at least the top several proxy nodes and reporting rank agreement), which is necessary to support the claim that AET.optimal() corresponds to the best official design.
The paper says, at §4.2, Appendix A, Appendix B.1
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The 'from zero' framing substantially overstates the degree of automation. The initial project M(0) is a complete, hand-designed MetaBBO algorithm: Appendix A specifies the DQN controller's state representation, Q-network architecture, discrete action space, reward design, and training schedule in full detail, and this same DQN-DE is itself listed as a competitive MetaBBO baseline (average 0.8831) in Table 2.
more of objection 3
In addition, the initial hyper-agent prompt in Appendix B.1 is heavily engineered, containing domain-specific algorithm lists, mandatory-modification rules, and an incremental-improvement strategy. The claim 'with minimal human inputs' and the characterization of M(0) as a 'zero' baseline are therefore not supported; the paper should reframe the contribution as automating design starting from a hand-crafted MetaBBO template with hand-crafted agent prompts, and should state explicitly which parts of the final behavior actually emerged from the evolution loop as opposed to being pre-specified in the initial prompts.
The paper says, at §4.5 and contribution list
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The contribution list advertises 'cold-start adaption either for cross-domain scenario or from off-the-shelf baselines', but the cross-domain experiments in §4.5 initialize the evolution with M(0) equal to the optimal M* from the single-objective domain, i.e., a warm start that embeds single-objective knowledge. If 'cold-start' means evolving from a generic template per domain, that is not what Fig. 4 tests. The paper should either re-run at least one domain from a non-transferred template or qualify the claim as warm-started transfer.
4 minor objections
The paper says, at §4.2, Table 2
Pith says: minor objection 1 of 4
On functions f21 and f22, HADA's mean is statistically indistinguishable from GLEET's within the reported standard deviations (f21: 0.9358±0.0823 vs 0.9346±0.0680; f22: 0.9420±0.1079 vs 0.9432±0.0662); the paper reports no significance tests anywhere, and a brief statement on the number of seeds and any pairwise test would strengthen the comparison.
The paper says, at §4.2
Pith says: minor objection 2 of 4
Sentence b108 ends with a stray '.;' fragment, and the sentence structure in the observations list is inconsistent; this should be cleaned up editorially.
show 2 more
The paper says, at References
Pith says: minor objection 3 of 4
Several references are dated 2026/2027 (e.g., Dharna et al. 2026, Chauhan et al. 2027, Zhang et al. 2026a/b) and one entry appears as 'Wang et al.; 2025' with an internal semicolon and an apparent missing publication field; the bibliography should be checked for formatting and completeness.
The paper says, at §4.5
Pith says: minor objection 4 of 4
The multi-objective result in Fig. 4 shows a modest margin over domain baselines, and the claimed '104.3%' figure is not derivable from the plot; providing a table with the same detail as Tables 1-2 for the constrained and multi-objective domains would make the cross-domain claims checkable.
What would settle it
For the single-objective COCO-BBOB run, take the projects stored in the algorithm evolution tree (all non-NA nodes, or a stratified sample of at least 30 spanning the generation range) and re-evaluate each under the official protocol: 20-epoch training and 10 independent test runs. Compute the Spearman correlation between the proxy Perf used for parent selection and the official Perf, and record the official ranking of the node returned by AET.optimal(). If Spearman rho is below ~0.6, or if the selected node falls outside the official top quartile, the reported HADA advantage is plausibly a winner's-curse artifact and should be reported with a selection-bias correction; if rho exceeds ~0.8 and the selected node remains top-ranked under official scoring, the headline claim survives.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 29 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
First thing to know: this is a solid applied result, not a paradigm shift. The recursive dual-agent idea comes from Hyperagents and the Darwin Gödel Machine line, but the concrete MetaBBO instantiation is new work: a task agent edits the optimizer codebase, a hyper agent rewrites both prompts, an evolution tree manages history and parent sampling, and one normalized score carries the same loop into constrained and multi-objective domains. Tables 1 and 2 support the core comparison: HADA averages 0.9674 on held-out BBOB functions versus 0.8942 for GLEET and 0.8741 for the best DE-family baseline. The out-of-distribution checks (other dimensions, other FE budgets, UAV planning) and the cross-domain warm-start experiments are genuine pluses, and the code and evolved artifacts are promised open source. That is enough to take the paper seriously.
The soft spots are real but mostly fixable. The headline 121.7%, 114.7%, and 104.3% leaps are not recoverable from any arithmetic in the paper; 0.9674 versus 0.8942 is about an 8% relative improvement, not 121.7%, and the text in Section 4.2 muddles which baseline is being used. That needs correcting in any revision. Bigger issue: parent selection and the final choice of the optimal node use proxy scores from 5-epoch training and 5 test runs, while the reported result uses 20 epochs and 10 runs on the same project. The paper never measures proxy-to-official concordance, so the returned node is exposed to winner's curse, and the evidence even hints at regression: the best proxy node in Figure 3 is 0.971 while the reported official average is 0.9674. The authors should either report the concordance or re-select the final node under the official evaluation protocol.
Two more caveats. The "from zero" framing is softer than the title: Appendix B.1 and B.2 show the initial prompts already instruct the agents to try CMA-ES, JADE, SHADE, and L-SHADE, and the final prompt hands over a full CMA-ES skeleton as "strongly preferred." That weakens the novelty and zero-knowledge claims, though it does not invalidate the performance comparison. And the causal claim that "openendedness is truly a key" goes beyond the evidence: Table 4 ablates whether the meta-level policy is modifiable, but not the hyper agent's self-modification, so the contribution of the self-referential loop specifically is untested.
The citation pattern is fine, and the paper is honestly positioned inside the recursive self-improvement literature. I would send it to peer review: it is a substantial, reproducible-looking empirical contribution with a clear weakness to fix, not a desk-reject. I would cite it for the cross-domain MetaBBO results, not for the percentage claims or the zero-knowledge framing.
The strongest objection
Section 4.1 states that all parent selection and the final choice of the 'optimal' project use proxy scores from 5-epoch training and 5 test runs, while the reported headline results use 20-epoch training and 10 runs on the same project. AET.optimal() returns the node with the best proxy score, so the reported official score (0.9674 for HADA in Table 2) is the official evaluation of a node selected for its proxy performance. If the proxy and official rankings disagree, this node can be a winner's-curse artifact: its official score overstates the expected official score of the best node in the search, and the evolution itself is guided by a misaligned signal. The evidence in the paper even hints at regression: the best proxy score shown in the Fig. 3 tree is 0.971 (node #63), while the reported official average is 0.9674. The paper never reports the proxy-official concordance, so the magnitude of selection bias is unknown. Compounding this, the headline 'performance leap' percentages (121.7%, 114.7%, 104.3%, abstract and Section 4.2) cannot be reproduced from the tables: HADA 0.9674 vs GLEET 0.8942 in Table 2 is ~8% relative improvement, not 121.7%, and no arithmetic in the paper recovers the claimed numbers. The abstract and Section 4.2 also phrase the comparison inconsistently ('using the worst-performing CMAES as baseline, HADA improves upon the SOTA MetaBBO (GLEET) by 121.7%'). Separately, the 'from zero' framing is undermined by Appendices B.1 and B.2, where both the initial and final prompts explicitly instruct the agents to try CMA-ES, JADE, SHADE, and L-SHADE, and even provide a full CMA-ES code skeleton that the task agent is told is 'STRONGLY PREFERRED'. The 'evolved novel variants' may therefore be human-injected known algorithms rather than open-ended discoveries; this does not invalidate the performance comparison, but it weakens the novelty and 'zero-knowledge' claims.
The reader's weakest_assumption is the unmeasured proxy-official agreement, and I concur that this is the single most load-bearing concern for the paper's central performance claim. The entire empirical case is that HADA evolves a naive DQN-DE template into variants that beat state-of-the-art baselines; that case rests on the official score of one node chosen by a proxy signal whose concordance is never quantified. Without concordance evidence, the reported official score is subject to unknown selection bias, and the headline percentages are not recoverable from the tables in any case. This concern is directly testable at modest cost using the released codebase and the AET history, and it is more consequential than the missing open-endedness ablation (Appendix D, Table 4) because it questions whether the reported numbers themselves are trustworthy, not just what caused them. I give the paper credit for publishing the code, for genuine held-out and cross-domain checks, and for the large reported average gap (0.9674 vs 0.8942 over GLEET), which suggests the qualitative finding may survive; the concern is about the magnitude and selection validity. The secondary scaffolding issue (Appendix B.2 prompts mandating CMA-ES/JADE/SHADE with provided skeletons) weakens the 'from zero' and open-ended-discovery framing but does not by itself overturn the performance comparison. Keeping the verdict CONDITIONAL is correct, with the added explicit condition that the proxy-official concordance be measured and the selection bias quantified.
The referee's report
If the claims hold, the contribution is substantial: HADA would be the first system to automate MetaBBO algorithm design end-to-end with a recursively self-improving agent loop, and it ships genuine out-of-distribution generalization checks (other dimensions, other FE budgets, UAV planning), a released codebase, and a full algorithm-evolution tree for inspection. Those strengths are real and should be credited: the held-out BBOB scores in Tables 1-2 are accompanied by per-function variances, the cross-domain experiments use an external benchmark (MetaBox-v2) rather than a self-defined testbed, and Appendix D reports ablations on the modifiable-policy choice, the initial optimizer, and the LLM backbone. However, the headline quantitative claims (the 121.7%/114.7%/104.3% 'leaps') are not recoverable from any table in the paper, and the evaluation pipeline selects the final reported node by a proxy score whose concordance with the official score is never measured. Both issues sit at the center of the abstract and contributions, so the significance of the paper matches its claims only after those are repaired and re-derived.
Pith's reading of the paper (a machine's summary, not the authors')
This paper claims that fully automated, dual-agent LLM coding systems can design competitive black-box optimizers from a deliberately naive starting point, with no human design choices beyond the initial template. Its central result is that, over 100 generations of code evolution, the loop discovered MetaBBO variants that outperform hand-crafted optimizers and existing MetaBBO baselines on the COCO-BBOB single-objective suite, where the champion scores 0.9674 against 0.8942 for the strongest prior MetaBBO baseline. The same machinery was then cold-started from that champion and re-evolved for constrained and multi-objective domains, with claimed performance leaps of 114.7% and 104.3%. If the claim holds, the bottleneck in learning-assisted optimizer design is the search procedure itself rather than human expertise, and letting the designing agent recursively rewrite its own instructions is what unlocks the discovery.
What carries the argument. The load-bearing mechanism is the hyper agent's self-referential editing loop: the hyper agent rewrites the task agent's prompt and its own prompt, so the procedure that generates design decisions is itself an editable artifact. Supporting it are an algorithm evolution tree that stores each project with its in-execution information (code, patches, logs, normalized scores) and samples parents by a weighted score-and-child-count rule; a unified 0-to-1 normalized performance score so problems of different scale compare fairly; and a strict interface contract that keeps any newly written optimizer compatible with the evaluation harness, letting the agents swap the low-level algorithm or the meta-level policy without breaking the loop.
Core claim. The paper's proposal is to treat MetaBBO algorithm design as a coding task solved by two interacting LLM agents: a task agent edits the source code of the MetaBBO project, and a hyper agent edits both the task agent's instructions and its own instructions, so the search loop also improves the way it searches. A tree structure records every evolved project together with unified, domain-normalized performance scores, and parent selection based on score and child count balances exploitation and exploration. From a naive DQN-DE starter that merely tunes the $F$ and $CR$ of a differential evolution search, the paper reports a champion scoring 0.9674 on held-out BBOB functions versus 0.8942 for the best MetaBBO baseline, with knowledge transferring to constrained and multi-objective benchmarks. Interpreting the evolution tree, the paper describes a trajectory where the hyper agent first forces structural optimizer replacement, then shifts focus to meta-level learning design such as state normalization and augmentation, and finally reaches a step that conditions the meta-policy on optimization progress and adds local search. The authors acknowledge that the current search-behavior heuristic is simple and that the experiments cover continuous optimization only, with combinatorial optimization left as future work.
The premise it rests on. The entire 100-generation search, including the choice of the final champion, assumes that the cheap 5-epoch, 5-run proxy scores rank projects the same way the official 20-epoch, 10-run scores do, and the paper presents no measurement of that agreement.
If the paper is right (Pith's inferences, not the authors')
An LLM-driven, self-modifying pipeline rather than human expertise can serve as the search engine for learning-assisted optimizers, with the three-domain leaps (121.7%, 114.7%, 104.3%) offered as the concrete evidence.
Cross-domain adaptation works by cold-starting from the single-objective champion and re-evolving, implying that design knowledge, not just a trained policy, is reusable across domains.
Because the meta-level policy itself is editable, the design space searched is wider than fixing a low-level optimizer; the ablation reports 0.9674 with a modifiable meta-level policy versus 0.9039 without.
The evolved source code and neural parameters are released, making the discovered algorithms reproducible artifacts rather than only the framework that found them.
The claim that no human expertise is needed is only established relative to the cheap proxy evaluation used during search; the paper never measures whether the 5-epoch, 5-run proxy ranking agrees with the official 20-epoch, 10-run ranking, so the returned 'optimal' project could partly reflect winner's-curse noise.
The 121.7% headline percentage is anchored to the worst-performing baseline and the previous SOTA MetaBBO, whereas the raw Table 2 numbers (0.9674 versus 0.8942) suggest a different normalization; a careful reader should trust the per-function table over the percentage.
The machinery appears domain-agnostic, so a direct test of the general claim would be porting the same two-agent loop and interface contract to combinatorial optimization or expensive black-box settings, which the paper itself names as future work but does not test.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Meta-Black-Box Optimization (MetaBBO) is one of the highlights in the recent AI for Optimization trend. This paradigm's bi-level workflow leverages the learnable algorithm design policy at meta level to ensure the performance and generalization improvement on the low-level optimization task. While MetaBBO helps advance the performance lower bound of the resulted optimization system, it is currently handcrafted and customized case by case to adapt different optimization problems, which inevitably introduces inherent subjectivity and hence restricts the performance upper bound and usability in practice. In this paper, we address this issue by regarding MetaBBO's design loop as coding task, where we could introduce openendedness into MetaBBO with recursive self-improvement capability of advanced coding agents. Specifically, we propose a dual-agent framework: i) a task agent continuously refines the codebase of a target MetaBBO approach through code evolution; ii) a hyper agent progressively modifies the task agent and itself to provide open-ended design behavior; iii) the evolved MetaBBO codebase is evaluated and all in-execution information is fed back to the agents for recursive self-referential improvement. As a result, given a naive MetaBBO template, our framework automates a design evolution and finds novel variants superior to up-to-date human-made MetaBBO baselines. Surprisingly, the experimental results also demonstrate that our framework supports fast adaption across different optimization domains. Solid interpretation analysis further reveals interesting design principles emerge in such open-ended process. This work serves as the first exploration on automating design of complex learning-assisted optimization algorithms.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-29
The authors' main claim, as the classifier picked it out 0.99
“As a result, given a naive MetaBBO template, our framework automates a design evolution and finds novel variants superior to up-to-date human-made MetaBBO baselines.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.95 · states limitation 0.13 · human participants 0.03 · controlled comparison 0.83 · randomized assignment 0.01 · states quantitative result 0.05 · compares with prior results 0.95 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Meta-Black-Box Optimization (MetaBBO) is one of the highlights in the recent AI for Optimization trend. | background 1.00 | 0.04 | 0.02 | asserted 1.00 |
| This paradigm's bi-level workflow leverages the learnable algorithm design policy at meta level to ensure the performance and generalization improvement on the low-level optimization task. | background 0.99 | 0.06 | 0.03 | asserted 1.00 |
| While MetaBBO helps advance the performance lower bound of the resulted optimization system, it is currently handcrafted and customized case by case to adapt different optimization problems, which inevitably introduces inherent subjectivity and hence restricts the performance upper bound and usability in practice. | gap 0.82 | 0.06 | 0.03 | asserted 1.00 |
| In this paper, we address this issue by regarding MetaBBO's design loop as coding task, where we could introduce openendedness into MetaBBO with recursive self-improvement capability of advanced coding agents. | contribution 0.80 | 0.42 | 0.06 | hedged possibility 0.89 |
| Specifically, we propose a dual-agent framework: i) a task agent continuously refines the codebase of a target MetaBBO approach through code evolution; ii) a hyper agent progressively modifies the task agent and itself to provide open-ended design behavior; iii) the evolved MetaBBO codebase is evaluated and all in-execution information is fed back to the agents for recursive self-referential improvement. | method 0.96 | 0.87 | 0.07 | asserted 0.99 |
| As a result, given a naive MetaBBO template, our framework automates a design evolution and finds novel variants superior to up-to-date human-made MetaBBO baselines. | result 0.98 | 0.95 | 0.66 | asserted 1.00 |
| Surprisingly, the experimental results also demonstrate that our framework supports fast adaption across different optimization domains. | result 1.00 | 0.95 | 0.10 | asserted 1.00 |
| Solid interpretation analysis further reveals interesting design principles emerge in such open-ended process. | interpretation 0.96 | 0.91 | 0.07 | asserted 0.99 |
| This work serves as the first exploration on automating design of complex learning-assisted optimization algorithms. | contribution 0.92 | 0.77 | 0.11 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.76
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-29: 1 request, 7,054 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment