acceptodds
Under review as a conference paper at ICLR 2027

Hill Sampling for Test-Time Scaling: A Simple and Better Alternative to Repeated Sampling, Evolution, and Training

Abstract

Large language models (LLMs) can improve solutions to verifiable scientific and algorithmic problems by spending additional computation at test time. Recent systems achieve strong results with increasingly elaborate evolutionary search harnesses or by updating model parameters during test-time training. We ask how much of this machinery is necessary. We introduce Hill Sampling, a form of hill-climbing optimization that repeatedly samples candidate programs from a frozen LLM, retains the best program found so far, and conditions all subsequent samples on that program. We evaluate the method on circle packing, sums and differences of sets, and Erdos' minimum-overlap problem using three open-weight models. Hill Sampling sets a new state of the art on circle packing among published methods, improves over the AlphaEvolve reference on Erdos' minimum-overlap problem, and achieves strong results on sums and differences of finite sets. The circle-packing and Erdos results require only hours of wall-clock time on eight NVIDIA H100 GPUs. We compare Hill Sampling against what is, to our knowledge, the largest application by parameter count of evolution strategies (ES) to LLM weights at test time. Surprisingly, when evaluating a method by the best program it generates, we find that learning model weights via ES is worse than invoking ES with a learning rate set to zero, i.e., using random weight-space perturbations to search for better models. Moreover, repeated sampling outperforms both ES methods, and Hill Sampling is the strongest of all. These results suggest a simple test-time compute allocation strategy: repeatedly sample edits to the best verified solution found so far before introducing additional complexity, such as adding archives, diversity mechanisms, evolutionary scaffolds, or test-time parameter learning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
test-time scaling, test-time training, TTT, algorithmic discovery, evolution, ES, LLM, hill sampling, repeated sampling, evolution strategies, verifiable rewards, mathematical discovery, agentic, in-context learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WXJGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Hill Sampling (HS), which uses a frozen LLM to sample batches of edits to the best verified program found so far. The best candidate replaces the incumbent if its reward is at least as high, and becomes the context for the next round. The authors evaluate HS on circle packing, sums and differences of finite sets, and Erdős' minimum-overlap problem against additional search mechanisms (RS, Model Noise, ES). HS achieves competitive results. The main contributions are a simple search baseline and an empirical analysis of mean return versus maximum return.

Critical strengths
  1. The simple search baseline has practical value. HS achieves strong mathematical search results without parameter updates or an elaborate program database. These results offer a useful baseline for assessing the benefits of more complex discovery systems.

  2. The distinction between mean return and maximum return is informative. Experiments with ES, max@8, Model Noise, and entropy control show that improving average quality need not improve the best solution found. These controls help characterize the limitations of test-time learning for discovery.

Critical weaknesses
  1. This paper provides new empirical results on mathematical discovery, but its methodological contribution needs clearer positioning.** Algorithm 1 in the (1+1)-EPS paper (arXiv 2407.10873) already uses a closely related search loop, and the study reports competitive performance. ShinkaEvolve (arXiv 2509.19349) also evaluates selecting the best program as the parent on circle packing. Relative to EPS, the main differences are batched sampling and accepting equal-reward candidates. The paper compares N = 64 and 512 but does not report N = 1, where each candidate is evaluated and the incumbent can be updated before the next candidate is generated. If batching is the main methodological addition, a comparison under otherwise matched settings and total LLM completions, with wall-clock time reported separately, would help isolate its benefits for search quality and efficiency.

  2. The HS and RS comparisons support propagating the best program in some settings, but do not fully establish the recommendation to allocate search compute to HS first. The authors describe some initialization sources and include NI ablations. However, they do not consistently report initial verified scores or best-so-far score trajectories for HS and the baselines. The iterative improvement over initialization and the cost of progress at different search stages therefore remain unclear.

  3. Results across seeds show that pure HS is competitive, but the abstract's claim that “HS is the strongest of all” is broader than the detailed findings. The paper defines the objective as the maximum observed verified reward within a fixed generation budget. However, the Figure 1 caption reports no significant difference between methods on Sets and identifies both HS and HS + ES as the best methods on Erdős. The broader mechanism comparison in Figure 5 is also limited to Erdős with Mistral. A lack of observed gains in this setting does not establish that these mechanisms are generally unnecessary.

  4. Comparisons within each domain use the same base model, but robustness across models remains unclear. The main experiments use gpt-oss-20b for Circles, OLMo for Sets, and Mistral for Erdős, leaving model and domain effects intertwined across tasks. The additional runs with different models do not systematically repeat the method comparisons, and the Sets run also changes the prompt and execution settings. These limitations weaken the support for the broader conclusions.

Questions for authors

Could the author clarify the contribution beyond prior best-program search methods and the scope of the claims discussed in the weaknesses above? Are there existing N = 1 results, search trajectories, or comparisons across models that further support these claims and the recommendation to use HS first?

2 comments

Thank you for your feedback and for noting the practical value of Hill Sampling and the value of the ES results. In response to these comments, we have added text and additional plots to our results. If you agree with the responses, please do increase our score, and if not, please let us know where we still disagree:

  1. ”This paper provides new empirical results on mathematical discovery, but its methodological contribution needs clearer positioning. Algorithm 1 in the (1+1)-EPS paper… already uses a closely related search loop… The paper compares N = 64 and 512 but does not report N = 1…” While we did find batching to be critical, both in terms of wall-clock time (which would be ~8-512x slower for N=1, and therefore prohibitively costly to run) and performance (even in the reported delta between N=64 and 512), Hill Sampling can be seen as a simplification of many existing algorithms. Our contribution is not primarily methodological novelty, but rather simplification, scale, and empirical verification. No prior work has not demonstrated that such a simple algorithm can achieve state-of-the-art results among published methods when scaled appropriately, which is a major contribution to future research efforts.
  2. ”The HS and RS comparisons support propagating the best program in some settings, but do not fully establish the recommendation to allocate search compute to HS first. …they do not consistently report initial verified scores or best-so-far score trajectories for HS and the baselines. The iterative improvement over initialization and the cost of progress at different search stages therefore remain unclear.” In response to this request, we have generated best-so-far plots for each of the three main tasks. We have generated best-so-far plots for all three tasks and summarize their findings below. We will include these plots in the final manuscript. Circles: HS and RS curves have similar shape and overlap significantly until halfway through the evaluation budget (6K responses), at which point HS overtakes both RS and the AlphaEvolve benchmark, while RS remains substantially and significantly lower. Sets: Both curves are nearly identical. Erdos: HS remains above RS at every point on the plot, and is strictly superior, with an especially large gap before 3K responses. After this, the results converge quickly to the final scores shown in the main results (Figure 1).
  3. “Results across seeds show that pure HS is competitive, but the abstract's claim that “HS is the strongest of all” is broader than the detailed findings… Figure 1 caption reports no significant difference between methods on Sets and identifies both HS and HS + ES as the best methods on Erdős. The broader mechanism comparison in Figure 5 is also limited to Erdős with Mistral…” Across all three domains, HS achieves the best result or performs comparably to the best method, with no statistically significant difference where it does not lead. Although HS + ES achieves a higher observed score on Erdős, the difference is not statistically significant, so our experiments do not demonstrate a benefit that justifies the additional complexity. HS + ES is also our own method, not an existing baseline. We have additionally clarified in the text that HS is better than all existing baselines measured, in response. We also had to restrict our ablations and additional experimentation to Erdos in order to maintain rigorous evaluation at reasonable compute scale. That said, by achieving SOTA results among published methods, many of which use elaborate search components or test-time training, we demonstrate that these components are not required to obtain strong results on the problems studied. Our matched ablations on Erdős provide additional evidence of the same conclusion that such components do not improve performance. Taken together, we believe these results provide strong evidence for the value of simple discovery methods, which is highly useful to the scientific community.
  4. ”Comparisons within each domain use the same base model, but robustness across models remains unclear…” By varying both the domains and models, we demonstrate strong performance across three distinct model–domain combinations, providing evidence of robustness across both dimensions jointly. While varying them independently would help isolate the effects, this is not necessary to demonstrate generality across combinations. Moreover, a full factorial evaluation is prohibitively expensive at our compute scale. We chose the models arbitrarily, to avoid biasing the results in favor of our method. We even note that using GPT-OSS 20b improves performance on Sets, in Section 5.0.

Questions For Authors: “Could the author clarify the contribution beyond prior best-program search methods… and the recommendation to use HS first?” Hill Sampling can be seen as a simplification of many existing algorithms, including (1+1)-EPS. (See 1. for details.) However, we still provide much more convincing scale, simplification, and empirical results, relative to prior work. A central contribution of this paper is demonstrating that this simple algorithm can achieve state-of-the-art results among published methods when scaled appropriately, which, to our knowledge, has not previously been demonstrated at this level of performance. In response to comments, we have also generated additional plots (See 2. for details) to support our claims to use HS first. We have additionally added text to resolve ambiguity in claims (3.), noting that HS performs better than, or not significantly different from, all methods evaluated, and that HS is better, overall, compared to all existing baselines measured. Overall, we believe we clearly show that an extremely simple search principle can reach state-of-the-art mathematical discovery results, while the ES experiments also reveal important limitations of parameter-space optimization for best-solution discovery, which collectively more than justify the contributions of this paper.

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

The paper studies a simple program-search loop that repeatedly edits the best program found so far. It evaluates this approach on three mathematical problems and compares it with repeated sampling, ES updates, and several search variants.

Critical strengths

The ES controls are useful. In particular, the zero-learning-rate comparison shows that improving average reward does not necessarily improve the best discovery.

Critical weaknesses
  1. My main concern is novelty. The core loop appears to be the same as Sequential Best-of-N in [1]. LLM-based hill climbing also has earlier precedents [3,4], alongside the broader AlphaEvolve framework [5]. Since [2] already establishes that simple search can be competitive, I found the additional contribution limited.

  2. The conclusions seem broader than the experiments support. Figure 5 tests the extra search components only on Erdos, and the training comparison mainly concerns ES. These results are useful, but I do not think they establish that more elaborate search or test-time training is generally unnecessary.

Overall Assessment

I see some value in the ES analysis. However, the method closely overlaps with existing work, and the experiments do not yet establish a sufficiently strong new conclusion. I lean toward rejection.

Questions for authors
  1. What distinguishes Hill Sampling from Sequential Best-of-N in [1]? Beyond the final scores, what would you consider the main new finding?

  2. I would be more convinced by a matched comparison with archive-based search on another task, such as the Heilbronn triangle problem in [1]. Keeping the model and budget fixed would help test whether the component findings extend beyond Erdos.

  3. For the training claim, could you compare frozen-model search with RL-based adaptation under matched conditions, following [6]? Otherwise, narrowing the conclusion to ES would seem more appropriate.

References

[1] Gupta et al. Automated Discovery Has No Universally Superior Harness, 2026.

[2] Gideoni et al. Simple Baselines are Competitive with Code Evolution, 2026.

[3] van Stein and Bäck. LLaMEA: A Large Language Model Evolutionary Algorithm for Automatically Generating Metaheuristics, IEEE TEVC, 2025.

[4] Eberhardinger et al. From Code to Play: Benchmarking Program Search for Games Using Large Language Models, 2024.

[5] Novikov et al. AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery, 2025.

[6] Yuksekgonul et al. Learning to Discover at Test Time, 2026.

4 comments

Thank you for your engagement, and for noting the insight from the ES and zero-learning rate ES investigation. We would very much like to emphasize the novelty in our empirical findings and conclusion, in addition to the ES line of work.

We believe we have addressed each comment below. If you agree, please do increase our score, and if not, please let us know where we still disagree:

  1. “My main concern is novelty. The core loop appears to be the same as Sequential Best-of-N in [1]. LLM-based hill climbing also has earlier precedents [3,4], alongside the broader AlphaEvolve framework [5]. Since [2] already establishes that simple search can be competitive, I found the additional contribution limited.” We acknowledge that Sequential Best-of-N in [1] uses essentially the same loop. However, it is concurrent work, and our primary contribution is empirical rather than algorithmic novelty. Importantly, [1] also reports weaker results and reaches a different conclusion, arguing for a complex adaptive harness. Likewise, in [2] (which is also concurrent), their baselines require additional complexity (e.g., resets and phased evaluation) and do not consistently surpass both AlphaEvolve and ShinkaEvolve on any of the domains they evaluated, and are worse on the three domains we evaluate. We address [2] in the related work, and have updated the text to address the similarity to [1], in response to this feedback. Note that no prior work shows that such a simple algorithm can achieve SOTA results when scaled appropriately, which we do. This is a major contribution to future research efforts.
  2. ”The conclusions seem broader than the experiments support. Figure 5 tests the extra search components only on Erdos, and the training comparison mainly concerns ES. These results are useful, but I do not think they establish that more elaborate search or test-time training is generally unnecessary.” We had to restrict our ablations and additional experimentation to Erdos in order to maintain rigorous evaluation at reasonable compute scale. That said, by achieving SOTA results among published methods, many of which use elaborate search components or test-time training, we demonstrate that these components are not required to obtain strong results on the problems studied. Our matched ablations on Erdős provide additional evidence of the same conclusion that such components do not improve performance. Taken together, we believe these results provide strong evidence for the value of simple discovery methods, which is highly useful to the scientific community.

Questions For Authors:

  1. What distinguishes Hill Sampling from Sequential Best-of-N in [1]? Beyond the final scores, what would you consider the main new finding? (See 1. Above)
  2. “I would be more convinced by a matched comparison with archive-based search on another task, such as the Heilbronn triangle problem in [1]. Keeping the model and budget fixed would help test whether the component findings extend beyond Erdos.” (See 2. Above)
  3. For the training claim, could you compare frozen-model search with RL-based adaptation under matched conditions, following [6]? Otherwise, narrowing the conclusion to ES would seem more appropriate. We have narrowed the claim to “test-time parameter learning via ES” in response. We believe the ES findings remain independently significant without claiming superiority over all test-time learning approaches.

Overall Assessment: “I see some value in the ES analysis. However, the method closely overlaps with existing work, and the experiments do not yet establish a sufficiently strong new conclusion…” A central contribution of this paper is demonstrating that this simple algorithm can achieve state-of-the-art results among published methods when scaled appropriately, which has not been shown in prior work at all. We also show that Neural Thicket’s style sampling via model perturbations outperforms ES, and that regular token-level sampling outperforms both of those, which is a non-trivial contribution by itself. We strongly believe that each of these contributions is significant, and the collection more than justifies the paper’s novelty.

In response to the review, we also now have results from an additional experiment: OpenEvolve on Circles, a baseline with extra search components beyond additional objectives and ES on the main results. We will include these results in the final manuscript.

Setup: A) We use the default configuration from OpenEvolve over three seeds. We also increase the iterations of their “Phase 2” to match the total number of LLM responses as in Figure 1 (a). We chose to increase Phase 2 since Phase 1 is responsible for their rough initialization, while Phase 2 results in breakthrough performance, and Phase 2 matches the prompt in our evaluation. Note that we also tune a temperature parameter in our experiments, and OpenEvolve is tuned with an unknown budget for each domain, including circles. That said, we try to make the runs as comparable as possible. We use the same evaluation, increase the parallel evaluations from 4 to 8 to fully utilize our GPUs (keeping total completions constant and matched), and change the timeout to 5 seconds to match our allocated time for evaluation. B) We also run a supplemental no-slack experiment similar to the setup in Table 1 (including no slack an a 100-second evaluation). The score we report in Table 1 comes from round 14 (7,168 LLM completions) of the best single seed, and running longer did not seem to improve this seed. We report the same result for a single seed of OpenEvolve with no slack at the equivalent number of responses to round 14 and 25. We note that for this supplemental experiment, our best seed was selected from multiple seeds, with more tuning and longer runs, so this result is not controlled for compute, as in the rest of Table 1. This experiment is primarily to illustrate a sense of a similar no-slack seed. The difficulty of controlling for compute here are why we report compute matched experiments in Figure 1.

Results (Normalized over AlphaEvolve’s score of 2.6358627564136983; ↑ higher is better)

A)

TaskOpenEvolve (mean ± SE) ↑Hill Sampling 512 Temp 1 (mean ± SE) ↑Hill Sampling 512 Temp Tuned (mean ± SE) ↑
Circles0.9862837151535512±0.0110771881233231431.0000345706939697±0.0000110070268319067171.0000345706939697±0.000011007026831906717

B)

MethodBest No-Slack Score at Round 14Best No-Slack Score at Round 25
OpenEvolve0.99704348040.9978084468
Hill Sampling 5121.00004565051941.0000456505194

Another experiment has just completed: OpenEvolve on Erdos, which we will also add to the final manuscript to strengthen our comparison to evolutionary algorithms in the paper.

Setup: We use the default config from OpenEvolve, compare to results from Figure 1 (c), and set up the experiment as in A) in the prior comment. We additionally change the number of total tries at evaluation to 2, to match our experimental setup in Figure 1 (c).

Results (Normalized by dividing AlphaEvolve’s score of 0.38092303510845016 by the result; ↑ higher is better)

TaskModelOpenEvolve (mean ± SE) ↑Hill Sampling 512 Temp 1 (mean ± SE) ↑Hill Sampling 512 Temp Tuned (mean ± SE) ↑
ErdosMistral-24B0.9846457448953535±0.00257741918240659470.9985832992835055±0.000067466767853773710.99922776222229±0.00017724917612279762

An additional experiment has just completed comparing different models on the same task. We use the same setup as in the prior comment.

Results (Normalized by dividing AlphaEvolve’s score of 0.38092303510845016 by the result; ↑ higher is better)

TaskModelOpenEvolve (mean ± SE) ↑Hill Sampling 512 Temp 1 (mean ± SE) ↑Hill Sampling 512 Temp Tuned (mean ± SE) ↑
ErdosMistral-24B0.9846457448953535±0.00257741918240659470.9985832992835055±0.000067466767853773710.99922776222229±0.00017724917612279762
Erdosgpt-oss-20b0.9872131548911236±0.0108508607540410440.9990458764989222±0.000254676336108169250.9991490562719733±0.00014262367610438603

Hill Sampling is better, regardless of model.

Many of OpenEvolve's hyperparameters are adjusted per task; for Hill sampling, we report both default hyperparameters (i.e., temperature of 1) and a single tuned hyperparameter (i.e., temperature) on Erdos. Hill Sampling is better, regardless of tuning.

If the additional context and experimentation collectively influence your assessment, please do update the score, and if not, please do let us know what else we can address. Thank you for your feedback!

Reviewer_WjAiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Hill Sampling: keep one program (the incumbent), sample N edits of it from a frozen LLM each round, run them, and replace the incumbent when an edit scores at least as well (Sec. 3.2, Eq. 3). There is no archive, no population, no crossover and no weight update. The method is tested on three verifiable problems, each with a different open-weight model: circle packing (26 circles, gpt-oss-20b), sums and differences of sets (OLMo-3.1-32B), and the Erdős minimum-overlap problem (Mistral-Small-24B) (Sec. 4.1). Baselines are repeated sampling from the initial program (RS), random weight perturbations without learning (Model Noise), evolution strategies on the model weights (ES, with max@8 and softmax variants), and Hill Sampling combined with ES (Sec. 4.2). On Erdős, a further set of mechanisms is tested as variants of or alternatives to Hill Sampling (Top-K parents, diversity, subpopulations, in-context history, execution feedback, temperature control, weight-space hill climbing) (Sec. 5.3, App. F). The main claims are: a new best published result on circle packing, a result better than AlphaEvolve on Erdős, and a "strong" result on Sets (Table 1); ES with a learning rate of zero gives a better maximum than ES with learning (Fig. 2); ordinary token sampling beats weight-space noise (Fig. 3); and none of the extra mechanisms on Erdős helps (Fig. 5). Results use three seeds per method and per-method temperature tuning (Sec. 4.1, App. C.1). Prompts, hardened verifiers and the best circle-packing program are given in App. B, D and E.

Critical strengths
  • The method is simple and fully specified in a few lines (Sec. 3.2), and the paper is clear about what it does and does not do.
  • Reproducibility materials are better than usual: full prompts (App. B), the verifier code (App. D), the best circle-packing construction and the program that produced it (App. E), and the hyperparameter choices (App. C).
  • The paper is honest in places where it would be easy to hide problems: the footnote on page 6 explains that competing circle-packing results differ at floating-point precision, and the Fig. 3 caption says some bars use one seed.
  • The "no initial information" ablation on circle packing (same best score without the domain-specific prompt) is a useful finding (Fig. 1a, Sec. 5).
  • The ES and model-noise study is the most original part of the paper: it is a large-model test of the idea that useful experts sit close to the pretrained weights, and the finding that learning improves the mean while the maximum falls (Fig. 2) is worth knowing.
  • Hardening the verifiers after observing reward hacking on Sets (App. D) is good practice.
Critical weaknesses
  1. The Erdős improvement over AlphaEvolve is smaller than the tolerances in the paper's own verifier. The claimed gain is 0.380923 − 0.38089767 ≈ 2.5e-5 (Table 1). The verifier in App. D accepts an h whose integral is off by up to 1e-3 (atol=1e-3) and accepts a reported C5 that differs from the recomputed one by up to 1e-4 (atol=1e-4). A numerical check shows that re-optimizing h under a ±1e-3 integral slack moves the achievable overlap by about 1e-4, roughly four times the claimed gain; the reported-value tolerance is another 1e-4 if the reported value is what gets scored. The code in App. D does not show which value is scored and does not renormalize h. The verifier is in the model's context (App. D, first paragraph), and the authors themselves saw reward hacking on Sets. The Erdős construction is absent from the paper (App. E gives only the circle-packing one), so the reader cannot recompute the bound. A public report on the OpenEvolve evaluator the paper starts from (issue #562, filed 30 October 2026, after the deadline) describes exactly this loophole. Please state which value is scored, renormalize h to integrate to exactly 1, give the construction, and report its recomputed value, ideally in exact arithmetic. This is feasible in the discussion period. Until then the Erdo''s claim in the abstract is unverified.

  2. The circle-packing "new state of the art" is a tie at floating-point precision on a benchmark that is exhausted at the record. The improvement over ThetaEvolve is 7.5e-9 (2.635983084917604 vs 2.6359830774, Table 1). The footnote on page 6 gives Evölther's value as 2.6359830849176054, which is higher by 1.4e-15, three units in the last place of a float64, and says that "validity can depend on floating-point evaluation details". Packomania currently lists 2.635983084919 for this instance, above the paper's value, and published systems agree to 7–9 digits. The record program itself (App. E) is a standard pipeline: a hexagonal seed, random restarts, simulated annealing and SLSQP. That is a polishing step inside a basin every leading system has already found, so the last digits say nothing about the search method. The footnote also says the paper's standard procedure gives 2.6359830848923913; which run produced the Table 1 value is unclear. The real result here is Fig. 1a, where Hill Sampling's mean best is about 1% above RS and the ES variants. Please drop "new state of the art" from the title, abstract and conclusion, say that the method matches the best known packing, and report what still discriminates: completions or time to reach it, or a harder instance.

  3. The headline numbers come from runs that have no baseline counterpart. The comparison runs in Fig. 1 use 200/600/80 rounds at N=64 (12,800 / 38,400 / 5,120 completions) with 5 s and 10 s timeouts (Sec. 4.1). The Table 1 numbers come from different runs: Erdo''s at round 70 of Hill Sampling (512), i.e. 35,840 completions over 12 hours with two 20 s timeouts, seven times the Erdo''s comparison budget; Circles at round 14 with a single 100 s timeout; Sets at round 2 (Table 1 caption). No baseline was run under these settings, and the paper does not say how many such runs were launched. A "best found" number grows with the number of attempts. In Fig. 1c, no method's mean best reaches the AlphaEvolve value. Please report the number of headline-style runs, give baselines under the same settings, and separate "best over everything we tried" from the controlled comparison.

  4. "Better than evolution" is never tested under matched conditions. The paper builds on OpenEvolve (Sec. 4.1, Reproducibility statement), yet never runs OpenEvolve's own default search (islands and a MAP-Elites archive), ShinkaEvolve, or any other evolutionary harness with the same model and the same number of completions. The evolution numbers in Table 1 come from other papers with other models (Gemini for AlphaEvolve) and unknown budgets. Sec. 5.3 does test island-like pieces (Subpopulation Evolution: 64 subpopulations of eight, merged after four steps) and parent-selection variants, but only on Erdo''s, where RS and Hill Sampling already tie (Fig. 5), so those tests cannot show that a harness is unnecessary on the one domain where Hill Sampling clearly wins. Since the codebase is OpenEvolve, a matched run on Circles with the same model, completions and timeout is cheap and would settle the second claim of the title.

  5. The method already exists under other names, and the closest prior work is compared only on published numbers. Hill Sampling is (1+N) hill climbing with an LLM as the mutation operator. The paper itself calls it "a simple evolutionary algorithm" (Sec. 3.2). The same loop appears as "HillClimb" in LLM4AD (Zhang et al., PPSN 2024) and as the "greedy protocol" in Xing et al. 2026, who run it on this very circle-packing instance. Gideoni et al. 2026 propose sequential conditioned sampling, which conditions each generation on a random subset of programs that ran without error; the difference here is greedy selection of the single best program. The paper compares with Gideoni et al. on published numbers (Sec. 2) and does not run it, although the change is one line of parent selection. Ye et al. 2026 (SimpleTES, cited) combine parallel exploration, feedback-driven refinement and local selection with the same model family, again with no comparison. The standard test-time-compute literature on repeated sampling and sequential revision is absent: Brown et al. 2024 (Large Language Monkeys), Snell et al. 2025 (sequential vs parallel test-time compute), Madaan et al. 2023 (Self-Refine), Lehman et al. 2022 (Evolution through Large Models). Please state what is new relative to these, cite them, and run sequential conditioned sampling as a baseline.

  6. Three seeds cannot carry the paper's negative claims. Most comparisons sit inside overlapping error bars: on Sets "there is no significant difference between the methods" (Fig. 1 caption); on Erdős all bars except the no-information variant lie within about 0.07% of each other (Fig. 1c, Fig. 5), and the Fig. 5 caption says none of the mechanisms "convey a significant advantage". The contribution list turns this into "show them all to be unnecessary", which is a much stronger statement than "we could not detect a difference with three seeds". Fig. 3 uses one seed per noise level. Temperature is tuned per method over the three reported seeds (App. C.1); the grid is small and the same for every method, so this adds only a small, symmetric optimism, but there are no held-out seeds. The clear positive result is circle packing (Fig. 1a); the rest should be stated as inconclusive or run with more seeds.

  7. The ES negative result is tied to a very small ES budget. Main ES runs use a learning rate of 1e-7 with sigma 3e-4 and 32 antithetic pairs (App. C.2). Because compute is matched by completions, ES (max@8) receives only about 25, 75 and 10 updates on Circles, Sets and Erdős. ES does learn under these settings: mean fitness rises in Fig. 6a and strongly in Fig. 2a (at sigma 1e-3), while the maximum falls with the learning rate (Fig. 2b). That is an interesting finding, but it is a finding about this budget. Qiu et al. 2026 ran ES with a learning rate about 5,000 times larger for 500–2,500 iterations. Whether a budget of that size reverses the picture is untested, and the abstract's "learning model weights via ES is worse" should be qualified by the budget. Two small inconsistencies to fix: Fig. 2 and Fig. 6 use sigma 1e-3 and temperature 0, while the main runs use sigma 3e-4 and temperature about 1; App. C.1 says "0.0001" where Fig. 6 and App. C.2 say 1e-3.

  8. One model per domain, so model and method are confounded. Each domain uses a different model (Sec. 4.1, "we arbitrarily divide"). The abstract's "using three open-weight models" reads as if each method were tested on three models. The only cross-model evidence is two single Hill Sampling runs with gpt-oss-20b (Erdo''s, 0.38088232; Sets, 99% of AlphaEvolve with a different prompt and no initial code), each confounded with other changes. No method-by-model comparison exists, so nothing can be said about whether the circle-packing advantage transfers, or whether the Sets result (about 96% of AlphaEvolve) is a model effect. Calling the Sets result "strong" (abstract) is an overstatement.

Questions for authors
  1. Erdo''s: is the score computed from the recomputed C5 or from the value the program reports? Is h renormalized to integrate exactly to 1 before the overlap is computed? Please provide the best construction and its recomputed value, verified in exact arithmetic.
  2. How many runs of the Table 1 kind (Erdős 70 rounds at N=512 with 20 s timeouts; Circles with a 100 s timeout) were launched, over how many seeds and temperatures? What do RS and the ES variants score under those same settings?
  3. Please run OpenEvolve's default configuration (islands and MAP-Elites), and ShinkaEvolve if possible, with the same model, completions and timeout, at least on Circles.
  4. Please run sequential conditioned sampling (Gideoni et al. 2026) under the same budget. What exactly is new in Hill Sampling relative to it, to LLM4AD's HillClimb, and to SimpleTES?
  5. Circle packing: please show score against completions for each method up to the best known packing. The record itself no longer separates methods; the path to it does.
  6. Which run produced the Table 1 circle-packing value, given that the footnote says the standard procedure gives 2.6359830848923913?
  7. ES: App. C.1 says 1e-7 was chosen for stability. With about 10–75 updates per run, how far do the weights move in total, and would a budget closer to Qiu et al. 2026 change the maximum-fitness conclusion?
  8. Temperature is tuned over the reported seeds. What are the results on fresh seeds at the chosen temperatures?
  9. Sets: is Hyra's 1.21 the same quantity as Eq. 7? Lin & Li 2026 define a different exponent. On the paper's own scale, Gerbicz 2025 (cited) reports about 1.173, above AlphaEvolve's 1.158. Which value should Table 1's "best prior" show?
  10. The incumbent is never re-evaluated (Sec. 3.2). For stochastic programs a lucky score can block later improvements. How often did accepted incumbents have a score equal to the previous one, and did re-evaluating the incumbent change anything?
  11. Will the code be released? The reproducibility statement describes documentation only.
8 comments

Thank you for your feedback and for noting the general reproducibility and the originality of the ES and model-noise study. We address all points below, including additional experiments, updates to the text, and new plots, to strengthen the paper. If you agree with the responses, please do increase our score, and if not, please let us know where you still disagree:

Questions For Authors:

  1. ”Erdo''s: is the score computed from the recomputed C5 or from the value the program reports? Is h renormalized to integrate exactly to 1 before the overlap is computed? Please provide the best construction and its recomputed value, verified in exact arithmetic.” Our paper uses the evaluation code from OpenEvolve, which has slack allowed both from normalization and from reporting, as of time of submission. While this does mean the discovered Erdos solution may not be exactly realizable, other papers also evaluate with slack (e.g., TTT-discover also has the reporting slack, potentially also inherited from the OpenEvolve benchmark). As requested, we have re-normalized and evaluated without slack. Doing so yields a zero-slack bound of 0.38096501243139724 with an associated score of 0.999889813180798. Again, we would like to emphasize that other papers evaluate with slack and the existence of slack does have implications for feasibility, but does not invalidate the evaluation as a useful benchmark for comparing methods. We will note the slack in the paper and how this changes the result. Thank you for pointing this out and helping to strengthen the paper.
  2. ”How many runs of the Table 1 kind (Erdős 70 rounds at N=512 with 20 s timeouts; Circles with a 100 s timeout) were launched, over how many seeds and temperatures? What do RS and the ES variants score under those same settings?” We generally ran 3 seeds, and 1 or 2 temperatures, depending on the problem. However, we evaluated several different timeouts and durations, and chose favorable temperatures, so simply reporting 3 seeds may not capture the whole picture. The point of this table was not to provide a measure of variation or matched evaluation, but rather to establish our result that scaling a very simple algorithm can achieve top results among published methods. We do not have computational resources to run each of these under matched settings, nor do we want to bias toward a timeout which worked well for our method, which is why we additionally provide controlled experiments under our initial and fair conditions.
  3. ”Please run OpenEvolve's default configuration (islands and MAP-Elites), and ShinkaEvolve if possible, with the same model, completions and timeout, at least on Circles.” We have done so and report results below. We also note that Hill Sampling outperforms every prior result with a documented method on Circles, including evolutionary algorithms that outperform OpenEvolve (e.g., ThetaEvolve). Collectively, we believe that our claims are more than justified.
  4. ”Please run sequential conditioned sampling (Gideoni et al. 2026) under the same budget. What exactly is new in Hill Sampling relative to it, to LLM4AD's HillClimb, and to SimpleTES?” Gideoni published their method concurrently while we were working on ours, and have weaker results, so we do not feel the result would change the conclusion of our paper, and we cannot evaluate every baseline, even if it hadn’t been concurrent. More specifically, given the results of the simple baseline, Gideoni et al. propose a method that requires additional complexity (e.g., resets and phased evaluation), and they do not consistently surpass both AlphaEvolve and ShinkaEvolve on any of the domains they evaluated, and are worse on the three domains we evaluate. Specifically, the simple algorithm in Gideoni’s work uses random parent selection, LLM4AD uses one-at-a-time (N=1) hill climbing, and SimpleTES is a more general framework. Hill Sampling can be seen as a simplification of many existing algorithms. Our contribution is not primarily methodological novelty, but rather simplification, scale, and empirical verification. No prior work has demonstrated that such a simple algorithm can achieve top results among published methods when scaled appropriately, which is a major contribution to future research efforts.
  1. ”Circle packing: please show score against completions for each method up to the best known packing. The record itself no longer separates methods; the path to it does.” In response, we have generated best-so-far plots for each of the three main tasks. We have generated best-so-far plots for all three tasks and summarize their findings below. We will include these plots in the final manuscript. Circles: HS and RS curves have similar shape and overlap significantly until halfway through the evaluation budget (6K responses), at which point HS overtakes both RS and the AlphaEvolve benchmark, while RS remains substantially and significantly lower. Sets: Both curves are nearly identical. Erdos: HS remains above RS at every point on the plot, and is strictly superior, with an especially large gap before 3K responses. After this, the results converge quickly to the final scores shown in the main results (Figure 1).
  2. ”Which run produced the Table 1 circle-packing value, given that the footnote says the standard procedure gives 2.6359830848923913?” This run was a separate 3-seed run that we allowed more time for evaluation than in our main experiments, and that we ran with no slack for easy comparison.
  3. ”ES: App. C.1 says 1e-7 was chosen for stability. With about 10–75 updates per run, how far do the weights move in total, and would a budget closer to Qiu et al. 2026 change the maximum-fitness conclusion?” We ran ES beyond convergence (roughly twice as long) in early experiments, and results were consistent.
  4. ”Temperature is tuned over the reported seeds. What are the results on fresh seeds at the chosen temperatures?” We choose the set of three seeds with the best temperature using nine initial seeds. We chose this setup such that all results, for our proposed method and baselines, incorporate the same advantage from temperature selection.
  5. ”Sets: is Hyra's 1.21 the same quantity as Eq. 7? Lin & Li 2026 define a different exponent. On the paper's own scale, Gerbicz 2025 (cited) reports about 1.173, above AlphaEvolve's 1.158. Which value should Table 1's "best prior" show?” Thank you for pointing this out. It appears Gerbicz 2025 and Zheng 2025 achieve the best comparable result, which we have noted in the text, while MLEvolve (Du et al., 2026) currently achieves the best result for fully autonomous discovery, which we have placed in our table.
  6. ”The incumbent is never re-evaluated (Sec. 3.2). For stochastic programs a lucky score can block later improvements. How often did accepted incumbents have a score equal to the previous one, and did re-evaluating the incumbent change anything?” This is true, and we do not have a measure of how often equal scores are replaced. We primarily chose to not reevaluate so that scores monotonically improved, given results on initial toy experiments. Whether it is useful to accept a lower-scoring solution in order to aid exploration on more challenging domains is a good question for future research.
  7. ”Will the code be released? The reproducibility statement describes documentation only.” We have documented as much as possible so that reproduction of this simple baseline should be straightforward, but are not at liberty to share the code.

Detailed Responses:

  1. ”The Erdős improvement over AlphaEvolve is smaller than the tolerances in the paper's own verifier.... A public report on the OpenEvolve evaluator the paper starts from (issue #562, filed 30 October 2026, after the deadline) describes exactly this loophole. Please state which value is scored, renormalize h to integrate to exactly 1, give the construction, and report its recomputed value, ideally in exact arithmetic. This is feasible in the discussion period. Until then the Erdo''s claim in the abstract is unverified.” As noted, our code uses the evaluation from OpenEvolve, which has the slack both from normalization and from reporting, as of time of submission. While this does mean the discovered Erdos solution may not be exactly realizable, other papers also evaluate with slack (e.g., TTT-discover also has the reporting slack, likely also inherited from the OpenEvolve benchmark). As requested, we have re-normalized and evaluated without slack. Doing so yields a zero-slack bound of 0.38096501243139724 with an associated score of 0.999889813180798. Again, we would like to emphasize that other papers evaluate with slack and the existence of slack does have implications for feasibility, but does not invalidate the evaluation as a highly useful benchmark for comparing methods. We will note the slack in the paper and how this changes the result. Thank you for pointing this out and helping to strengthen the paper.
  2. ”The circle-packing "new state of the art" is a tie at floating-point precision on a benchmark that is exhausted at the record. The improvement over ThetaEvolve is 7.5e-9 (2.635983084917604 vs 2.6359830774, Table 1). The footnote on page 6 gives Evölther's value as 2.6359830849176054, which is higher by 1.4e-15, three units in the last place of a float64, and says that "validity can depend on floating-point evaluation details". Packomania currently lists 2.635983084919 … Please drop "new state of the art" from the title, abstract and conclusion, say that the method matches the best known packing, and report what still discriminates: completions or time to reach it, or a harder instance.” Packomania currently lists Hyra as the top performance, and it achieves lower score under our evaluation. Hyra reports a score of 2.6359830951068446 (greater than ours); however, their given construction does not pass our no-slack check. To be fair, we take their solution and decrease the slack by 1e-1, 1-2, and so on, until failure. Their solution is valid after subtracting 1e-9, yielding 2.6359830691068447, which is lower than the ThetaEvolve result we report in our table. Packomania lists Hyra's score at yet another value (2.635983084919), which is inconsistent with Hyra’s own report and our own evaluation, and it is unclear how this figure was attained. Evolther does achieve a higher score, but no details are documented on the method used or how to implement, limiting its scientific utility. We therefore stand by the claim that our score is the best of all no-slack results that include published or documented details of the methods used to obtain them. That said, our primary contribution is to achieve comparable performance to top performing methods, with a far simpler algorithm, within a reasonable compute budget. This is the core motivation and contribution of our paper. No prior work shows that such a simple algorithm can achieve top results when scaled appropriately, which we do. This is a major contribution to future research efforts.

Comments on the paper

6 comments

Thank you for your engagement. Answers are below. If you agree with the responses, please do increase our score, and if not, please let us know where you still disagree:

  1. “Could you provide the initial verified score and the numerical mean final best-so-far score with its standard error for HS on each task here during the discussion?“ We report these statistics here. Note that the initial score of -1 on Sets reflects a penalty for a consistent timeout. Removing the timeout produces a score of 0.8671824336051941±0.0.
    TaskInitial score (mean ± SE)Final best-so-far score (mean ± SE)
    Circles0.9597642169962065±0.01.0000345706939697±1.1007026831906717e-05
    Erdős0.7694517902052751±0.00.99922776222229±0.00017724917612279762
    Sets-1.0±0.00.9575090408325195±0.00030066570914450395
  2. ”Is the recommendation to use pure HS first supported by a measured trade-off between search quality and computational cost, or primarily by implementation simplicity?” The recommendation is primarily to favor implementation simplicity. There is additional non-negligible cost (e.g., sampling full model perturbations at the scale of billions of parameters), but the more significant barrier is likely to be the implementation of this algorithm in systems like vLLM, which is not designed for this purpose, while we also have no significant evidence that it performs any better. We also note that HS + ES is a novel combination, and so there is not a strong precedent for favoring this algorithm.
  3. “I appreciate the clarification that the contribution is primarily empirical. Closely related search loops appear in (1+1)-EPS (arXiv 2407.10873) and Sequential Best-of-N in Gupta et al. (arXiv 2607.18235), which also report competitive results from simple search. Is it accurate to view the main contribution as scaling this search principle to obtain strong mathematical results and systematically evaluating additional search mechanisms? Please clarify if this overlooks a conceptual distinction or a new empirical insight.” That is largely correct. In those papers they do not achieve state-of-the-art on any domain among published methods, and find the performance to vary more significantly. Specifically, Gupta et al. is contemporary work that advocate for a complex adaptive allocation across harnesses, and they do not ever match or exceed any scores from AlphaEvolve, unlike our results. (1+1)-EPS likewise finds methods to vary and suggests EoH is more robust to model choice. We find that our simple method, scaled and tuned over temperature, breaks that trend. This is primarily an empirical insight that would be highly valuable to the scientific community. We additionally provide insight into ES training at the largest scale recorded, and the use of token-sampling stochasticity compared to model perturbations.

Additionally, we have run a couple more experiments with positive results for OpenEvolve on Circles, which we will include in the final manuscript.

Setup: A) We use the default configuration from OpenEvolve over three seeds. We also increase the iterations of their “Phase 2” to match the total number of LLM responses as in Figure 1 (a). We chose to increase Phase 2 since Phase 1 is responsible for their rough initialization, while Phase 2 results in breakthrough performance, and Phase 2 matches the prompt in our evaluation. Note that we also tune a temperature parameter in our experiments, and OpenEvolve is tuned with an unknown budget for each domain, including circles. That said, we try to make the runs as comparable as possible. We use the same evaluation, increase the parallel evaluations from 4 to 8 to fully utilize our GPUs (keeping total completions constant and matched), and change the timeout to 5 seconds to match our allocated time for evaluation. B) We also run a supplemental no-slack experiment similar to the setup in Table 1 (including no slack an a 100-second evaluation). The score we report in Table 1 comes from round 14 (7,168 LLM completions) of the best single seed, and running longer did not seem to improve this seed. We report the same result for a single seed of OpenEvolve with no slack at the equivalent number of responses to round 14 and 25. We note that for this supplemental experiment, our best seed was selected from multiple seeds, with more tuning and longer runs, so this result is not controlled for compute, as in the rest of Table 1. This experiment is primarily to illustrate a sense of a similar no-slack seed. The difficulty of controlling for compute here are why we report compute matched experiments in Figure 1.

Results (Normalized over AlphaEvolve’s score of 2.6358627564136983; ↑ higher is better)

A)

TaskOpenEvolve (mean ± SE) ↑Hill Sampling 512 Temp 1 (mean ± SE) ↑Hill Sampling 512 Temp Tuned (mean ± SE) ↑
Circles0.9862837151535512±0.0110771881233231431.0000345706939697±0.0000110070268319067171.0000345706939697±0.000011007026831906717

B)

MethodBest No-Slack Score at Round 14Best No-Slack Score at Round 25
OpenEvolve0.99704348040.9978084468
Hill Sampling 5121.00004565051941.0000456505194

Another experiment has just completed: OpenEvolve on Erdos, which we will also add to the final manuscript to strengthen our comparison to evolutionary algorithms in the paper.

Setup: We use the default config from OpenEvolve, compare to results from Figure 1 (c), and set up the experiment as in A) in the prior comment. We additionally change the number of total tries at evaluation to 2, to match our experimental setup in Figure 1 (c).

Results (Normalized by dividing AlphaEvolve’s score of 0.38092303510845016 by the result; ↑ higher is better)

TaskModelOpenEvolve (mean ± SE) ↑Hill Sampling 512 Temp 1 (mean ± SE) ↑Hill Sampling 512 Temp Tuned (mean ± SE) ↑
ErdosMistral-24B0.9846457448953535±0.00257741918240659470.9985832992835055±0.000067466767853773710.99922776222229±0.00017724917612279762

If the additional context and experimentation collectively influence your assessment, please do update the score, and if not, please do let us know what else we can address. Thank you for your feedback!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.