acceptodds
Under review as a conference paper at ICLR 2027

Initialization Improves LLM-Driven Discovery

Abstract

Large Language Models (LLMs) have been used for novel discovery of algorithms, theorems, drugs, and other tasks through the use of harnesses that prompt an LLM to iteratively optimize an objective. In this work, we study the relationship between the population of previous iterates and eventual discovery success. We generalize past work on harness design to develop a suite of 12 harnesses called 'Modular' and characterize their performance across 5 diverse discovery tasks, finding that discovery success is brittle and sensitive to harness design. We uncover mode collapse, characterized by a dramatic drop in the diversity of iterates, as a common failure mode. We find that popular state-of-the-art harnesses and diversity-inducing harness interventions, which aim to prolong this collapse, yield inconsistent gains. Our results instead uncover that the performance of early discoveries is predictive of eventual success. We therefore propose a universally applicable intervention that performs an initial stage of parallel exploration in order to initialize subsequent iterative optimization. Our method provides consistent gains across many harnesses and target applications, confirming the importance of initialization in LLM-driven discovery.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Driven Discovery, Discovery Harnesses, Scientific Discovery, Auto-research, LLMs, Inference-time compute

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XpnbOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how initialization affects LLM-driven discovery. The authors develop 12 Modular discovery harnesses and evaluate them across five tasks. They find that sequential discovery often suffers from mode collapse and that diversity-oriented interventions do not consistently improve final discovery quality. Instead, early discovery quality is predictive of downstream success. Based on this observation, the paper proposes a parallel exploration stage to construct a strong initial population before sequential optimization, improving several existing harnesses by 2–21%.

Critical strengths
  1. Clear and practically useful insight about initialization. The paper identifies early population quality as a more reliable predictor of downstream discovery than simply maintaining diversity, and proposes a simple intervention applicable across different harnesses. This is a useful design principle for LLM-driven discovery systems.
  2. Broad empirical study across tasks and harnesses. The analysis covers multiple discovery tasks, LLMs, 12 Modular variants, and three state-of-the-art harnesses. The consistent gains from initialization across these settings make the main empirical observation reasonably convincing.
Critical weaknesses
  1. The initialization-budget selection is partly heuristic and potentially post-hoc. The convergence thresholds are task-specific, and the parameter N is selected through visual analysis of the discovery curves. This makes the claim of a generally applicable automatic initialization strategy less convincing.
  2. Evaluation artifacts affect at least one discovery task. The authors identify reward hacking in CloudCast and apply a post-hoc score cutoff to remove such discoveries. While transparently reported, this raises some concern about the robustness of the task evaluators and whether similar undetected artifacts may influence other tasks.
Questions for authors
  1. How sensitive are the results to the manually chosen convergence parameters (\delta, N)?
  2. Can the authors clarify how the convergence criterion would be selected for a completely new task without examining its discovery curves?
  3. How confident are the authors that reward-hacking behavior is absent from the other discovery tasks?
2 comments

Thank you for your review! We are glad you find our initialization method to be practically useful and our empirical analysis to be broad and comprehensive. We respond to your concerns in turn below:


Initialization Budget Selection

The initially proposed convergence criterion was intended as a proof of concept, and we previously noted that it requires calibration by a domain expert, which may be reasonable to expect of practitioners designing a discovery task. However, we agree that domain-specific convergence hyperparameters are challenging to set.

To address your concern, we revise our recommendation: instead of performing a convergence analysis to determine the initialization-to-discovery budget ratio rr, we simply treat it as a hyper-parameter. Across the five (task, LLM) settings in the main paper, we find that discovery trajectories initialized with 25% of the discovery budget (r=0.25) consistently outperform uninitialized discovery.

We then also validate whether setting r = 0.25 generalizes to 15 total unique (task, LLM) settings in Appendix C, Fig 12, including two newly introduced (task, LLM) settings, DeepSeekV4 Flash (a new model family) for the CloudCast and CantBeLate task that we previously did not test. We find that serial trajectories initialized with 25% of the discovery budget beat their uninitialized counterparts (r = 0) in 12/15 settings. In the three settings where r = 0.25 did not actively beat uninitialized discovery, initialization either had a negligible impact or the LLM itself was not suited to the discovery task (i.e., lacked relevant parametric knowledge; we elaborate in Section C.3). In other words, using 25% of the discovery budget doesn’t harm the performance of discovery harnesses and in most cases improves it.

In conclusion, we find that simply using a predetermined 25% of the discovery budget for initializing serial trajectories consistently improves discovery performance across a wide range of settings, and sidesteps the problem of needing to determine task-specific convergence criteria. This validates the simplicity and utility of our initialization method.

Main paper Section 5 and Appendix C.3 have been updated with the revised r=0.25 recommendation, additional DeepSeek experiments (Figure 12), and subsequent analysis.


Reward Hacking

You raise reasonable concerns about task evaluators, and we agree! In this work, we use (and cite) the task definitions and their evaluation criteria as defined by prior published literature. For the task in question, CloudCast, we directly used the evaluation suite provided by (Agrawal et al., 2026a).

We notice, and raise concerns about, instances of reward hacking whereby the LLM (Gemini 3.5 Flash), rather than only implementing a cloud scheduling algorithm with the provided information, instead attempts to evaluate different discoveries locally and greedily choose the highest performing one. This behavior does result in a legitimate discovery, but may be undesirable in settings where evaluation should be independent of the discovery procedure. We leave this choice to the task-designer as it is outside the purview of improving discovery harnesses via interventions.

In this work, we assume this behavior is unwanted and subsequently detected reward hacking via anomalously high-scoring candidates. Therefore, we minimally augment Agrawal et al. (2026a)’s task evaluation by applying a max-score filter of 0.015 to CloudCast discoveries; only 12 out of 48817 CloudCast discoveries in this work (less than 0.025% of discoveries) exceeded this threshold. Our conclusions about the utility of initialization are unaffected by the inclusion/exclusion of these 12 candidate discoveries. We did not detect any such anomalously high-scoring candidates for other (task, LLM) pairs.

The responsibility of catching/evaluating reward hacking falls on the discovery-task designer, and appropriate edge cases should be covered in discovery evaluation. We point out that reward hacking is not a weakness of our initialization method (i.e., the scope of our paper), but rather a broader concern that task evaluations should account for.

We have updated Appendix B.1 with the analysis of reward hacking for CloudCast.


We have incorporated all updates from the review process into the paper PDF (demarcated in blue text).

[Agrawal et al., 2026a] optimize_anything: Unified text optimization can outperform specialized systems. In Proceedings of the ACM Conference on AI and Agentic Systems, 2026.

With the rebuttal period coming to an end in 2 days, we thank you for your review and are wondering if you have any outstanding questions? We have updated the manuscript with new experiments and summarized said updates above.

Reviewer_Bx1MOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the role of initialization in LLM-driven discovery tasks. The authors developed the Modular suite to implement 12 distinct iterative discovery harnesses and compared serial iteration against parallel sampling strategies across five tasks, identifying mode collapse as a common failure pattern. They found that interventions aimed solely at increasing population diversity failed to yield consistent performance improvements, whereas early population quality proved predictive of subsequent discovery outcomes. The paper addresses topics highly relevant to the burgeoning fields of LLM agents and AI-driven scientific discovery. It features an extensive experimental scope, covering various harnesses, multiple LLMs, and five distinct task categories.

Critical strengths
  1. Prior work on LLM discovery harnesses mostly focuses on refining iterative updates and boosting generative diversity. This paper shifts attention to initialization and empirically demonstrates that early population quality matters more than later-stage diversity interventions, offering a useful takeaway for the community.
  2. The authors developed a modular testbed supporting 12 harness variants, benchmarked against three established SOTA methods. Experiments span five diverse tasks and multiple LLMs, with a large empirical evaluation scale.
  3. The work verifies that serial iteration pipelines are prone to mode collapse, and standalone online diversity interventions cannot stably improve discovery performance. It identifies a correlation between early population quality and final task outcomes.
  4. The proposed parallel-then-serial strategy is harness-agnostic and conceptually simple, requiring minimal modification to existing discovery pipelines.
Critical weaknesses
  1. The innovation of this article is limited, and parallel sampling+screening of high-quality samples is restarted. This paradigm is a common approach in reinforcement learning and evolutionary algorithms; This article only transfers this idea to the large model harness discovery framework, lacking essential new ideas.
  2. The convergence criterion relies on manually setting hyperparameters to limit the degree of automation of the method; On a brand new unknown task, users do not know how to set $$\delta,N, N.
  3. The paper observed a correlation between early population quality and final performance, but did not fully rule out alternative explanations: the performance improvement may only come from more sampled tokens in the early stage, rather than the "high-quality initialized population" itself. The main limitation of this article is the lack of key baselines to separate the effects of "increasing the number of samples" and "initializing warm start", and to directly deduce the correlation as causality. The causal argument is weak and there is over interpretation.
  4. Some key conclusions are not consistently supported by the experimental results. Figure 7 claims to compare initialized and uninitialized SOTA harnesses, but duplicates Figure 1 and still shows Modular variants in the legend. The claimed 2%-21% improvements therefore cannot be verified from this figure.
Questions for authors
  1. Can you add a control baseline that fixes total token budget and allocates all initialization tokens directly to the serial stage, to separate gains from warm-start initialization versus extra sampling tokens?
  2. Without prior experimental curves for unseen tasks, how should practitioners select the convergence hyperparameters $$\delta,N),N)? Can you show the method still works under such settings?
  3. Why fix initial population size m=15m=15? Have you run systematic ablations over m across different tasks and LLMs to verify the robustness of this choice?
  4. Why is the baseline with continuously interleaved parallel sampling and serial iteration excluded? Could this simpler mixed strategy achieve comparable performance to your parallel-then-serial switching pipeline?
  5. How do you rule out that the observed mode-collapse signal is an artifact from the embedding model or amplified by pruning low-quality samples from the population?
6 comments

Thanks for your review! In response, we have:

  • Clarified how all of our baselines are budget-matched for fair comparison (Section 5)
  • Revised our recommendation to do a convergence analysis, and instead recommend treating the initialization budget as a hyper-parameter, which we empirically find can be set to r=0.25 (Section 5)
  • Run additional experiments to validate that the recommended initialization budget (r=0.25) generalizes to a new model family (DeepSeek V4 Flash); Appendix C.3 Fig 12.
  • Fixed Figure 7 to reflect the intended experiment
  • Run additional experiments to assess the impact of population size (Appendix C.3 Table 13)
  • Run additional control experiments to compare greedily initialized vs. randomly initialized populations under fixed token budgets (Appendix C.3, Table 13)
  • Run an additional analysis of discovery diversity w/ additional embedding models (Appendix C.2 Figs 10 & 11)
  • We have incorporated all updates from the review process into the paper PDF (demarcated in blue text).

We also respond in detail to your questions below:


Novelty

We agree that parallel sampling is a general paradigm in RL and evolutionary algorithms. Our contribution is not the paradigm itself, but its role in our unique study of how discovery trajectories should be initialized, which current LLM-driven discovery harnesses do not address.

We first confirm that discovery harnesses can outperform naive parallel sampling of an LLM as a baseline, but are often brittle across (LLM, task) settings. We then conduct a unique analysis into the reasons for brittleness by testing a wide range of harness interventions, ultimately finding that harnesses are sensitive to their initialization. This is our key contribution, as it enabled us to propose a simple initialization baseline which utilizes parallel sampling.

The fact that our initialization method improves the performance of a wide range of discovery harnesses (15 unique harnesses) validates our choice to study initialization as a key design choice in discovery harnesses. As you note, our approach is “simple, requiring minimal modification to existing discovery pipelines,” which we take to be a strength of our contribution as it is easy to adopt and build on.


Initialization Budget Selection

The initially proposed convergence criterion was intended as a proof of concept, and we previously noted that it requires calibration by a domain expert, which may be reasonable to expect of practitioners designing a discovery task. However, we agree that domain-specific convergence hyperparameters are challenging to set.

To address your concern, we revise our recommendation: instead of performing a convergence analysis to determine the initialization-to-discovery budget ratio rr, we simply treat it as a hyper-parameter. Across the five (task, LLM) settings in the main paper, we find that discovery trajectories initialized with 25% of the discovery budget (r=0.25) consistently outperform uninitialized discovery.

We then also validate whether setting r = 0.25 generalizes to 15 total unique (task, LLM) settings in Appendix C, Fig 12, including two newly introduced (task, LLM) settings, DeepSeekV4 Flash (a new model family) for the CloudCast and CantBeLate task that we previously did not test. We find that serial trajectories initialized with 25% of the discovery budget beat their uninitialized counterparts (r = 0) in 12/15 settings. In the three settings where r = 0.25 did not actively beat uninitialized discovery, initialization either had a negligible impact or the LLM itself was not suited to the discovery task (i.e., lacked relevant parametric knowledge; we elaborate in Section C.3). In other words, using 25% of the discovery budget doesn’t harm the performance of discovery harnesses and in most cases improves it.

In conclusion, we find that simply using a predetermined 25% of the discovery budget for initializing serial trajectories consistently improves discovery performance across a wide range of settings, and sidesteps the problem of needing to determine task-specific convergence criteria. This validates the simplicity and utility of our initialization method.

Main paper Section 5 and Appendix C.3 have been updated with the revised r=0.25 recommendation, additional DeepSeek experiments (Figure 12), and subsequent analysis.

Controlling for the Token Expenditure of Different Baselines

When comparing discovery trajectories from different harnesses/interventions, we always enforce a fixed token budget to ensure fair comparison of all baselines. The token budget of each discovery task is fixed a priori in our experimental methodology (Table 6).

“the performance improvement may only come from more sampled tokens in the early stage, rather than the "high-quality initialized population" itself.”

We agree this is an important distinction, and we clarify how our experiments already control for it. All harnesses, with and without initialization, run under the same fixed token budget, so that design choices that consume more tokens per discovery (e.g., LLM, mutator prompt, LLM-judge, initialization, etc.) are compared fairly. The token cost of initialization is deducted from the harness’s overall budget, so initialized trajectories receive no additional tokens compared with uninitialized ones.

“Can you add a control baseline that fixes total token budget and allocates all initialization tokens directly to the serial stage, to separate gains from warm-start initialization versus extra sampling tokens?”

This baseline is already included in the paper. In Figure 5, the gray bar (at initialization budget 0) allocates the entire token budget to the serial stage.

The initialized trajectories (blue bars) in Figure 5 are budget-matched to the uninitialized trajectories (gray bar); this means that we explicitly account for ALL of the tokens consumed while producing the initial parallel population plus the subsequent serial iterations.

“The main limitation of this article is the lack of key baselines to separate the effects of "increasing the number of samples" and "initializing warm start”

We already demonstrate the token-efficiency of our proposed initialization method by directly comparing it to budget-matched serial sampling.

However, to answer your question and separate these effects, we add a random initialization baseline. We compare uninitialized harnesses directly to harnesses that spend 25% of their token budget on greedy initialization (our method) and random initialization. The initialization methods differ only in the manner in which they subsample the parallel pool (e.g., greedily vs randomly). Both initialization methods therefore explore the same number of samples, and differ only in the quality of the initialized population. Greedy initialization consistently outperforms random (and uninitialized) baselines, which indicates that the initialization scheme itself drives better future discovery (Appendix C.3, Table 12).

Population SizeUninitializedGreedy Init. (s=0.98)Random Init.
150.51 ± 0.090.54 ± 0.05 (+6.2%)0.52 ± 0.07 (+2.6%)
300.49 ± 0.080.53 ± 0.04 (+7.2%)0.52 ± 0.06 (+5.9%)
450.51 ± 0.100.54 ± 0.04 (+7.3%)0.52 ± 0.07 (+2.1%)
Average0.50 ± 0.090.54 ± 0.04 (+6.9%)0.52 ± 0.07 (+3.5%)

General Questions

“How do you rule out that the observed mode-collapse signal is an artifact from the embedding model or amplified by pruning low-quality samples from the population?”

Good question; we find that there is variance in the diversity of samples across different harness designs, and we detail the effect of each of these design choices in Table 9. However, we don’t find that the presence or absence of mode collapse is indicative of discovery performance. More specifically:

  • Artifacts from embedding model: We re-assess the diversity of trajectories under additional embedding model choices in response to your query. In Appendix C.2 Fig 10, we use a general embedding model (all-MiniLM-L6-v2) and a code embedding model (nomic-ai/CodeRankEmbed) for natural language and coding tasks, respectively (per prior work), and in Fig 11 we use a general embedding model (BAAI/bge-small-en-v1.5) for natural language tasks and a code embedding model (microsoft/unixcoder-base) for coding tasks. We observe the same mode collapse trends in both Fig 10 and Fig 11.
  • Pruning low-quality samples is common in discovery harnesses, and is a design choice we adopted directly from [Yang, Chengrun, et al.] for our Modular harness suite; without pruning low performers, harnesses are likely to continue to generate low-performing future discoveries. To further assess the impact of varying the pruning strategy, we include SOTA harnesses (OE/SE/OA) which make more sophisticated pruning design choices, and these do indeed result in more diverse samples (See Table 2 in paper). However, we don’t notice that these harnesses consistently outperform less diverse harnesses in terms of discovery performance (see Table 4 in paper).

[Yang, Chengrun, et al.] "Large language models as optimizers." International Conference on Learning Representations. Vol. 2024. 2024.

General Questions Continued

“Why fix initial population size (m=15)?”

Good question. We fixed the initial population size of the modular harnesses to m=15 as we initially suspected that since discovery harnesses prioritize sampling high-performing discoveries for mutation, the lower-performing candidates in the population wouldn’t be consequential to discovery. As a result, we suspected that increasing population size wouldn't affect discovery. In response to your question, we directly confirm this hunch by running additional multi-seeded experiments where we vary m = (15, 30, 45) for the prompt optimization task across different population sizes (reported in Appendix C.3.1, Table 11). Population size does not seem to have a significant impact on discovery performance. (See table above, initialization gains are stable at 6-7% across population sizes.)

Further, we point out that we adopted the default recommended hyper-parameters for the SOTA (OA/OE/SE) harnesses (Appendix A.3); and each of these harnesses also has a different default population size. Ultimately, we found that initialization was beneficial across discovery harnesses regardless of individual design choices (e.g., population size).

“Why is the baseline with continuously interleaved parallel sampling and serial iteration excluded? Could this simpler mixed strategy achieve comparable performance to your parallel-then-serial switching pipeline?”

Good question. In this work, we study the role of initialization on discovery performance; the strategy you suggest is a generalization of our approach whereby a discovery trajectory is re-initialized periodically. As such, this is an extension of the initialization strategy rather than a baseline for it.

While we agree that it is interesting to study, this approach would introduce substantially more hyperparameters, such as how much compute to allocate to the serial and parallel phases and how frequently to switch between them. Designing an effective switching schedule between the parallel and serial phases is a substantial research question in its own right, beyond the scope of this work. We have added it to the discussion in the future work.

Clarifying Figure 7:

There was a typo in the latex figure code; thanks for catching this! We have corrected the typo so figure 7 now reports the gains of the three SOTA harnesses.

Reviewer_uvz7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how discovery harnesses manage the population of candidates used in LLM-driven optimization. It compares twelve combinations of mutation and parent selection with parallel sampling at matched token budgets, then tests diversity interventions and three existing harnesses. The experiments span five task families and thirteen task/model pairings. Sequential trajectories often lose embedding diversity, while interventions that preserve diversity do not consistently improve scores. The proposed method spends part of the budget on parallel sampling, selects a high-scoring population, and uses it to start sequential refinement. A convergence criterion chooses when to switch. Budget sweeps and comparisons across harnesses examine the effect of this initialization.

Critical strengths

The comparisons show how much harness design matters. The 12 Modular variants (Appendix A.1/Table 5), parallel baselines, five diversity interventions, and three existing harnesses cover several reasonable design choices. Figures 2-3, Tables 3-4, and Appendix C show that more embedding diversity does not reliably mean better discoveries. The result depends on both task and model.

Parallel initialization is easy to apply to an existing harness. Figure 5 compares intermediate budget allocations with fully sequential and fully parallel search. Figure 12 extends this to thirteen task/model pairings, including cases where the method works less well. Appendix D provides the mutation prompts needed to understand the Modular variants.

Critical weaknesses

The budget-selection rule has only been tested retrospectively. Appendix B.2 says N was chosen by visually inspecting completed parallel curves, with task-specific delta. Section 5 then orders the completed pool to detect convergence. This does not show how a rule fixed in advance would perform during a new search. Figure 12 matches the best budget in 9/13 settings; for GPT-OSS TSP and Qwen circle packing, the marked early budget is much worse than later allocations or pure parallel search. The claims about consistent near-ideal prediction and uniformly beating both endpoints should reflect these failures.

The uncertainty estimates and solution checks need more detail. In Tables 2-4, n=12, n=60, and n=3 count different harness configurations. Appendix C/Table 10 gives one score per named existing harness, without specifying independent repeats per configuration or how the 95% intervals were calculated. An average across configurations does not establish repeatable gains for each harness. CloudCast also has known reward hacking, handled by a post-hoc cutoff of 0.015 (Appendix B.1). Invalid-candidate rates and checks on candidates below the cutoff are missing, leaving their effect on the comparison unclear.

Questions for authors

How does batched switching perform on fresh trajectories when delta and N are fixed in advance? Please compare with fixed-fraction budgets and the best allocation, including the Figure 12 failures.

How many independent runs were made per configuration, and how were the intervals calculated? Please report paired initialization gains for each OA/OE/SE harness. Also clarify Figure 7: its caption names these three harnesses, but its legend says Modular, n=12.

How often were CloudCast candidates invalid, and do the results survive task-validity checks beyond the score cutoff? Are GSM8K selection and final evaluation examples disjoint, and are judge/evaluation tokens included in the budget?

4 comments

Thanks for your review! In response, we have:

  • Run additional multi-seeded experiments (Appendix C.3.1 Tables 13-16)
  • Revised our recommendation to do a convergence analysis, and instead recommend treating the initialization budget as a hyper-parameter, which we empirically find can be set to r=0.25 (Section 5)
  • Run additional experiments to validate that the recommended initialization budget (r=0.25) generalizes to a new model family (DeepSeek V4 Flash); Appendix C.3 Fig 12.
  • Fixed Figure 7 to reflect the intended experiment
    • Disaggregated results from Fig 7 into Table 10 in Appendix C.3
  • We have incorporated all updates from the review process into the paper PDF (demarcated in blue text)

We also respond in detail to your questions below:


Uncertainty Estimates and Sensitivity Analysis

“How many independent runs were made per configuration, and how were the intervals calculated?”

In the original submission, each configuration was run once. The nn in Tables 2–4 counts harness configurations within a family: all modular harnesses (n=12n=12, Table 5), modular harnesses with the five diversity interventions of Section 4 (n=60n=60), and SOTA harnesses (OA/SE/OE, n=3n=3). Within each family, the 95% confidence intervals were computed over these configurations, reflecting variation across harness designs. We have clarified this in the captions of Tables 2-4.

“An average across configurations does not establish repeatable gains for each harness.”

Because our goal is a harness-agnostic intervention, our primary analysis aggregates within each family (Table 3), and the original single-seed experiments show the same trend across (task, LLM) settings. This shows that initialization is robust to different harness design choices.

We agree, however, that this does not establish repeatable gains for each harness, and we therefore ran a new multi-seed analysis (now in Appendix C.3.1). It covers all 15 harnesses (12 modular, 3 SOTA) in all five (task, LLM) settings of Table 1, with and without initialization, over three seeds. Seeds control both the initial population and any non-deterministic parts of the harnesses. Results are disaggregated by harness ID for both Modular and SOTA harnesses, and since this involved 450 unique discovery trajectories, the tables are too large to include here; we report the mean and standard deviation over these experiments in Appendix C.3.1, Table 13-16. These results continue to show that initializing harnesses with 25% of the discovery budget consistently outperforms uninitialized counterparts.


Initialization Budget Selection

The initially proposed convergence criterion was intended as a proof of concept, and we previously noted that it requires calibration by a domain expert, which may be reasonable to expect of practitioners designing a discovery task. However, we agree that domain-specific convergence hyperparameters are challenging to set.

To address your concern, we revise our recommendation: instead of performing a convergence analysis to determine the initialization-to-discovery budget ratio rr, we simply treat it as a hyper-parameter. Across the five (task, LLM) settings in the main paper, we find that discovery trajectories initialized with 25% of the discovery budget (r=0.25) consistently outperform uninitialized discovery.

We then also validate whether setting r = 0.25 generalizes to 15 total unique (task, LLM) settings in Appendix C, Fig 12, including two newly introduced (task, LLM) settings, DeepSeekV4 Flash (a new model family) for the CloudCast and CantBeLate task that we previously did not test. We find that serial trajectories initialized with 25% of the discovery budget beat their uninitialized counterparts (r = 0) in 12/15 settings and purely parallel discovery (r = 1) in 9/15 settings. Since initialized trajectories outperform parallel discovery in the majority of settings, we conclude that initialized serial discovery harnesses are predictably better than parallel discovery. In the three settings where r = 0.25 did not actively beat uninitialized discovery, initialization either had a negligible impact or the LLM itself was not suited to the discovery task (i.e., lacked relevant parametric knowledge; we elaborate in Section C.3). In other words, using 25% of the discovery budget doesn’t harm the performance of discovery harnesses and in most cases improves it.

In conclusion, we find that simply using a predetermined 25% of the discovery budget to initialize serial trajectories consistently improves discovery performance across a wide range of settings, and sidesteps the need to determine task-specific convergence criteria. This validates the simplicity and utility of our initialization method.

Main paper Section 5 and Appendix C.3 have been updated with the revised r=0.25r=0.25 recommendation, additional DeepSeek experiments (Figure 12), and subsequent analysis.

Reward Hacking

We only noticed the CloudCast reward hacking behavior in 12 out of 48817 total discoveries (less than 0.025% of discoveries) for the Gemini 3.5 model; our conclusions about the utility of initialization are unaffected by the inclusion/exclusion of these 12 candidate discoveries.

We point out that reward hacking is not a weakness of our initialization method (i.e., the scope of our paper), but rather a broader concern that discovery-task designers should account for in their evaluation methods. In this work, use (and cite) the task definitions and their evaluation criteria as defined by prior published literature. For the task in question, CloudCast, we directly used the evaluation suite provided by (Agrawal et al., 2026a). We detected reward hacking via anomalously high-scoring candidates. We did not detect any such anomalously high-scoring candidates for other (task, LLM) pairs. The responsibility of catching/evaluating reward hacking falls on the discovery-task designer, and appropriate edge cases should be covered in discovery evaluation.


General Questions

“Please report paired initialization gains for each OA/OE/SE harness. ”

Here are the paired initialization gains for each OA/OE/SE harness (and updated in the paper in Appendix C.3.1, Table 10).

HarnessCan't Be Late (Gemini 3.7 Flash)CloudCast (Gemini 3.5 Flash)TSP (Gemini 3.5 Flash)Circle Packing (GPT-OSS 120B)Prompt Optim. (Llama 3.1 Instruct 8B)
SEUninitialized-88.600.0085-18792.510.46
Init. (s = 0.98)-90.090.0100-16832.500.46
% Change-1.68%18.14%10.43%-0.39%0.00%
OEUninitialized-95.060.0100-17552.540.44
Init. (s = 0.98)-90.090.0100-16792.540.56
% Change5.23%0.33%4.33%0.06%27.27%
OAUninitialized-90.280.0099-16802.510.54
Init. (s = 0.98)-89.920.0100-16782.570.66
% Change0.40%0.49%0.11%2.22%22.22%
Average (n = 3)Uninitialized-91.32 ± 8.330.0095 ± 0.0021-1772 ± 2502.52 ± 0.040.48 ± 0.13
Init. (s = 0.98)-90.03 ± 0.250.0100 ± 0.0000-1680 ± 62.54 ± 0.090.56 ± 0.25
% Change1.40%5.70%5.15%0.63%16.67%

Also clarify Figure 7: its caption names these three harnesses, but its legend says Modular, n=12.

Thank you for catching this! There was a typo in the latex figure code, which we have corrected, so Figure 7 now reports the gains of the three SOTA harnesses.

“are judge/evaluation tokens included in the budget?”

Yes, judge/evaluation tokens are included in the discovery budget as they consume resources during a discovery trajectory.

“Are GSM8K selection and final evaluation examples disjoint”

Yes, the selection and final evaluation examples are disjoint. In checking this, we found that the original figures in the paper mistakenly reported results with the selection examples. All figures/tables in the paper have now been updated to reflect the GSM8K results with the held-out examples. The initial reported trends remain the same.


[Agrawal et al., 2026a] optimize_anything: Unified text optimization can outperform specialized systems. In Proceedings of the ACM Conference on AI and Agentic Systems, 2026.

With the rebuttal period coming to an end in 2 days, we thank you for your review and are wondering if you have any outstanding questions? We have updated the manuscript with new experiments and summarized said updates above.

thanks for the detailed response and the new experiments. The three-seed results address my concern about whether the gains were repeatable within each harness. I also appreciate the move to a fixed 25% initialization budget and the additional tests on DeepSeek V4 Flash. These changes make the recommendation easier to evaluate and apply.

I’d still frame 25% as a starting point. It beats pure parallel search in 9 of 15 settings, so the best allocation can still depend on the task and model. With that caveat, the new evidence addresses my main concerns, and I’m raising my rating from 2 to 3 (weak acceptance).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Updates

Dear AC, SAC, and PCs,

We would like to highlight that each reviewer has noted the work as an important contribution to understanding the role of initialization in LLM-driven discovery and has acknowledged the simplicity and utility of our proposed solution.

To help facilitate the decision process, we outline the changes to our revised manuscript here:

  • Section 5: Clarified how all of our baselines are budget-matched for fair comparison.
  • Section 5/Appendix C.3: Revised our recommendation to do a convergence analysis, and instead simply recommend treating the initialization budget as a hyper-parameter; finding that using 25% of discovery budget (r=0.25) enables consistent gains over uninitialized harnesses.
  • Appendix C.3 Fig 12: Additional DeepSeek V4 Flash experiments; validates that the recommended initialization budget (r=0.25) generalizes to a new model family.
  • Appendix C.3.1: Multi-seeded experiments and sensitivity analysis to confirm robustness of initialization to performance variation of individual harnesses
  • Appendix C.3, Table 13: Additional control experiments to compare greedily initialized vs. randomly initialized populations under fixed token budgets (finding that proposed greedy initialization is constantly beneficial)
  • Appendix C.3, Table 13: Additional control experiments to compare varying population sizes (finding little difference between sizes).
  • Appendix C.2 Figs 10 & 11: Additional analysis of discovery diversity w/ additional embedding models, confirming the results are invariant to choice of embedding model
  • Figure 7: Fixed figure to reflect intended experiment
    • Disaggregated results from Fig 7 into Table 10 in Appendix C.3
  • We have incorporated all updates from the review process into the paper PDF (demarcated in blue text).

We believe the paper has been substantially strengthened through the rebuttal and meets the quality and significance expected for ICLR. Thank you again for your time and consideration.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.