acceptodds
Under review as a conference paper at ICLR 2027

Online Surrogate Repair: Decoupling High-Fidelity Feedback from Search Length in Closed-Loop Discovery

Abstract

Closed-loop AI scientists can generate candidate designs at low marginal computational cost, whereas reliable feedback may require wet-lab synthesis, characterization, or high-fidelity computation. Addressing this imbalance through custom laboratory automation remains infrastructure-intensive and costly, while replacing new experiments with a fixed surrogate leaves persistent model errors that can be amplified by optimization. We propose online surrogate repair (OSR), a closed-loop algorithm that uses sparse high-fidelity evaluations to update the surrogate throughout a longer agent search conducted primarily with inexpensive surrogate feedback. An acquisition rule selects which designs from the agent's accumulated proposals receive high-fidelity evaluation, and the resulting labels update the surrogate used in subsequent episodes. Across controlled synthetic environments, we demonstrate that improving global surrogate fit does not necessarily reduce maximum regret, whereas Q90-UCB and expected improvement (EI) substantially reduce regret by directing evaluations toward regions that determine the optimizer's decisions. On MADE, controls receiving high-fidelity feedback after every episode require – more oracle queries to match Online EI under two LLM orchestrators and more under the non-LLM Chemeleon+MLIP workflow. Online surrogate repair introduces a novel third feedback regime between fixed-surrogate operation and high-fidelity feedback after every episode, separating the frequency of high-fidelity evaluation from the duration of the agent's search.

Primary area:
optimization
Keywords:
AI for science, closed-loop AI scientists, online surrogate repair, surrogate modeling, active learning, expected improvement, materials discovery

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

The paper presents Online Surrogate Repair (OSR), a system in which a scientific agent searches using inexpensive surrogate feedback, while an acquisition rule periodically selects archived candidates for oracle evaluation. The resulting labels are appended to the surrogate’s context. Experiments on synthetic tasks and the MADE materials-discovery benchmark compare acquisition rules and feedback schedules. The paper aims to show that sparse online updates can support longer searches with fewer oracle feedback queries.

Critical strengths

The distinction between global prediction error and optimization regret is useful, and the experiments illustrate that improving one need not improve the other. This is arguably the paper’s most informative observation, although it is not developed into the central contribution.

Critical weaknesses
  1. The methodological novelty is severely limited. Searching with surrogate predictions, selectively evaluating candidates with the true objective, and incorporating those observations into subsequent model updates are established components of surrogate-assisted optimization. For example, Pitra et al., Doubly Trained Evolution Control for the Surrogate CMA-ES (2016), already combine optimizer-generated candidates, acquisition-based selection including EI and model updates using true evaluations. OSR’s “repair” consists of appending labelled observations to an existing model’s context, without introducing a new update mechanism, acquisition function, or adaptive schedule. Proposal-policy independence and modular separation of search from evaluation do not, by themselves, establish a substantive distinction from this literature.

  2. The experiments do not adequately establish incremental value over comparable existing methods. Most comparisons involve restricted variants of the proposed pipeline, such as no repair, one-shot static acquisition, or oracle feedback after every episode. These comparisons largely confirm the expected benefits of adaptive sampling and additional inexpensive search. Archive-BO is relevant, but retains the same archive, schedule, and EI rule while changing the surrogate; it therefore primarily tests surrogate choice rather than the contribution of the OSR framework. The paper lacks convincing comparisons against established methods that likewise permit extended surrogate-guided search and sparse true evaluations under matched resource budgets. Consequently, the results provide limited evidence of an advance beyond established practice.

Questions for authors
  1. What specific algorithmic contribution distinguishes OSR from existing surrogate-assisted optimization and evolution-control methods, beyond using an agent and an in-context surrogate?

  2. Can the authors provide comparisons against closely related methods with matched initial information, oracle budgets, and access to inexpensive search, while reporting candidate-generation and surrogate-evaluation costs?

  3. Beyond confirming that global prediction accuracy and optimization regret are different objectives, what new conclusion about their relationship is established, and how does it inform a method beyond standard acquisition rules?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Chng 1/2

Dear Reviewer Chng,

We thank you for the rigorous evaluation, for recognizing the value of our distinction between global prediction error and optimization regret, and for pointing us to Doubly Trained Evolution Control (DTS-CMA-ES, Pitra et al., 2016), which we now explicitly cite and benchmark against in Sections 2.2 and 4.3 (Table 3). We have substantially improved our paper, and all revised content is marked in red. We provide the following response:

[Q1 & W1] Algorithmic Contributions Distinguishing OSR from Evolution-Control Methods (DTS-CMA-ES, lq-CMA-ES).

We thank the reviewer for asking what OSR adds to evolution control. OSR integrates retrospective archive selection and in-context repair into an agentic discovery loop whose oracle budget is independent of search length. This gives two concrete distinctions:

  1. Cross-Episode Retrospective Archive Selection (AtA_t) vs. Intra-Generation Population Control (CtC_t): DTS-CMA-ES and lq-CMA-ES reuse previously measured points to fit their surrogates, while selecting new true evaluations from the current CMA-ES population. OSR instead retains unmeasured proposals across episodes. At each scheduled repair, EI selects from this archive and the new label updates feedback for subsequent episodes. This distinction has a large effect in the new MADE ablation at T=100T=100, Q=10Q=10: mean Top-1 is 128.3 for Online EI, versus 78.4 for periodic current-proposal evaluation, 79.9 for random archive selection, and 78.7 with a fixed surrogate (Table 5):
Method (T=100,Q=10T=100, Q=10, 30 systems)Top-1Top-5Top-5 AUCEF-topAF-topUnmatched
Online EI (Archive AtA_t)128.3(92)85.1(53)48.3(39)1.400(114)5.94(68)x12/30
Q90-UCB (Archive AtA_t)108.1(85)66.7(63)36.1(40)1.231(108)4.72(77)x10/30
Original Control (Q=TQ=T)99.7(73)57.9(45)32.6(30)1.00 (ref)1.00 (ref)0/30
Random archive79.9(71)33.3(45)18.4(26)0.880(79)2.06(55)x2/30
Fixed surrogate (Q=0Q=0)78.7(115)40.0(70)26.7(51)0.864(130)4.17(81)x9/30
Periodic proposal (Current CtC_t)78.4(51)40.9(37)22.3(24)0.906(85)1.17(33)x0/30
  1. In-Context Multi-Surrogate Fusion: The same repair loop can use LQ predictions or DTS GP means and variances as additional TabPFN features. The LQ combination attains mean normalized gaps of 0.304 and 0.405, versus 0.363 and 0.537 for native lq-CMA-ES at N=32N=32 and N=1000N=1000, respectively. The GP combination also improves on both DTS variants in mean normalized gap (Table 3).
AuthorsOpenReview ↗

Response to Reviewer Chng 2/2

[Q2 & W2] Comparisons Against Established Methods Under Matched Budgets and Candidate/Surrogate Evaluation Costs.

We thank the reviewer for requesting established baselines and explicit evaluation budgets. At matched oracle-query allowances, LQ + TabPFN + online EI achieves the lowest mean normalized gap at both context sizes: 0.304 and 0.405, versus 0.363 and 0.537 for native lq-CMA-ES. The comparison covers all 24 BBOB functions in 5D and 10D, with three instances per task, shared initial observations (N∈{32,1000}N\in\{32,1000\}), and an 32 additional oracle evaluations. Baselines include CMA-ES, TuRBO-1, native lq-CMA-ES, and both DTS-CMA-ES variants (Table 3):

MethodNorm. Gap (N=32N=32)Log Gap (N=32N=32)Norm. Gap (N=1000N=1000)Log Gap (N=1000N=1000)
TuRBO-10.326(16)1.820(33)0.506(17)1.573(36)
DTS-CMA-ES (EI)0.348(15)1.774(63)0.556(20)1.413(82)
Native lq-CMA-ES0.363(16)1.401(26)0.537(14)1.071(29)
DTS-CMA-ES (variance)0.460(17)1.976(64)0.748(19)1.767(62)
CMA-ES0.568(22)2.304(28)0.856(16)1.959(15)
Fixed TabPFN + EI0.693(21)2.356(28)0.541(17)1.580(26)
LQ + TabPFN + online EI0.304(15)1.423(29)0.405(15)1.051(22)
GP + TabPFN + online EI0.333(18)1.476(28)0.433(18)1.073(24)
Online TabPFN + EI0.376(17)1.958(29)0.439(17)1.465(24)
Online TabPFN + Peak0.428(17)2.083(26)0.465(16)1.499(22)

The paired normalized-gap improvements over native lq-CMA-ES are 0.058 [0.026, 0.091] and 0.131 [0.099, 0.164] at N=32N=32 and N=1000N=1000 (95% intervals; Appendix H). The GP combination also has positive paired intervals against DTS-CMA-ES with variance selection. Standalone Online EI reaches 0.439 at N=1000N=1000, below every standalone optimizer baseline in mean normalized gap.

The additional LLM cost is USD 0.33-0.68 per campaign at matched mean audited discovery counts. Online EI costs USD 1.35 versus 1.02 for DeepSeek and USD 2.02 versus 1.34 for Luna. Reported GPU times are 922 versus 813 and 342 versus 387 H100-seconds, respectively. We note that the additional costs are minimal compared to costly oracles such as wet-lab verification, justifying potential application of OSR in such cases (Appendix E, Table 9):

WorkflowOperating PointOracle Queries (QQ)LLM Cost ($)GPU Time (H100-s)Discoveries
DeepSeek-V4-FlashOnline EI (300 ep.)321.3592253.9
Control, catch-up (224.5 ep.)224.51.0281353.9
Control, matched budget320.1210213.07
GPT-5.6-LunaOnline EI (300 ep.)322.0234256.9
Control, catch-up (190.6 ep.)190.61.3438756.9
Control, matched budget320.16639.63

Seperately, the strictly search-time Top-1 comparison gives an acceleration factor of 4.52-5.44x less oracle queries (Table 4).

[Q3 & W1] New Conclusions on the Accuracy-vs-Regret Relationship and How They Inform Our Method.

We thank the reviewer for asking how the findings inform OSR. They guide optimization-directed label selection and motivate hybrid in-context repair:

  1. Global fit is an insufficient repair objective: Global-UQ reduces mean NRMSE by 0.0046 but mean regret by only 0.0631; EI reduces mean NRMSE by 0.0011 and mean regret by 0.4093 (Table 1). In the agent loop, our adapted Tab-AICL baseline reaches mean regret 0.1169 versus 0.0552 for Online EI. This supports selecting context labels for their effect on optimization.
  2. Query dispersion alone does not determine performance: EI achieves mean regret 0.0552 at normalized step distance 0.356, between Peak (step 0.079, regret 0.0976) and Tab-AICL (step 0.915, regret 0.1169). The geometry measurements cover 620 successive acquisition pairs per method (Appendix C.2):
MethodStep/d0\text{Step} / d_0Spread/d0\text{Spread} / d_0Incumbent Updates / WorldMean Regret (↓\downarrow)
Online Peak0.0790.1896.10.0976
Online Q90-UCB0.1890.2556.00.0600
Online EI0.3560.4485.60.0552
Tab-AICL0.9150.9161.10.1169

The online comparisons also favor EI over Q90-UCB: mean regret is 0.0552 versus 0.0600 in the synthetic campaigns, and mean Top-1 is 128.3 versus 108.1 on MADE. Q90-UCB's slight lead in the pooled static study therefore does not carry over to these closed-loop comparisons. 3. Constructive extension through hybrid repair: We add classical predictions as TabPFN features and update its context through Online EI. The LQ hybrid lowers mean normalized gap by 16% at N=32N=32 and 25% at N=1000N=1000 versus native lq-CMA-ES (Table 3).

Reviewer_ChngOpenReview ↗

Response to rebuttal

Thank you for providing additional experiments. However, my main concern is not that the paper combines existing techniques. A meaningful integration can certainly be a valuable contribution. Rather, the paper does not sufficiently establish a non-trivial insight connecting its motivation, design choices, and empirical findings. While the additional experiments demonstrate that the proposed combination can be effective, it remains unclear what generalizable methodological understanding the reader gains beyond the established principles of surrogate-assisted optimization and acquisition-based evaluation.

I would therefore appreciate it if the authors could clarify what genuinely new insights, empirical observations, or algorithmic contributions this paper offers beyond the existing literature.

Reviewer_fTTVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the trade-off between expensive real-world evaluation and lightweight surrogate evaluation in closed-loop AI Scientist systems, this paper proposes Online Surrogate Repair (OSR). Its core idea is to select only a subset of candidates with high optimization value for real-world evaluation during training, and to use the evaluation results to update the surrogate online. The main contribution of this paper is to provide pipeline and model adaptation for the online surrogate management and evaluation paradigm within a closed-loop AI Scientist system, and to demonstrate through experiments that global prediction error and maximum regret are not equivalent in surrogate optimization problems. Experimental results show that the proposed OSR is competitive on selected baselines, and that the optimization-oriented sampling strategy can effectively reduce maximum regret without significantly improving global prediction error.

Critical strengths

(1) This paper organizes established surrogate-assisted optimization concepts into relatively independent feedback management modules, enabling their stable integration into the closed-loop AI Scientist system.

(2) This paper compares different sampling methods and, through experiments, distinguishes between and optimizes global prediction error and maximum regret while presenting the metrics, which holds certain methodological significance.

Critical weaknesses

(1) The paradigm of online agent updating presented in this paper is not particularly innovative. It overlaps heavily with existing research in the field of surrogate-assisted optimization and fails to properly cite relevant papers. Though using different sampling methods and agent bases, this contribution has been classified in previous literature as an adaptation rather than an innovation.

(2) In Section 4.1 and Section 4.2, the test environment itself is generated by a Sampled Directed Acyclic Graph (SCM) that is highly consistent with the TabPFN pre-training prior, which introduces bias. A more generalized experimental environment is needed to further validate the algorithm’s advantages.

Questions for authors
  1. Novelty and Related Work: The online surrogate updating paradigm presented in this paper overlaps significantly with existing research in surrogate-assisted optimization and fails to properly cite relevant prior work. Could you clarify the specific distinctions between OSR and previous model management strategies?

  2. Experimental Bias and Generalization: In Sections 4.1 and 4.2, the test environments are generated by a Sampled Directed Acyclic Graph (SCM) that aligns closely with the TabPFN pre-training prior. How do you rule out the possibility that this structural consistency artificially inflates the performance of your method?

  3. Methodological Significance: Your comparison of different sampling methods and the experimental distinction between global prediction error and maximum regret hold certain methodological significance. Could you elaborate on how the proposed metrics and the optimization-oriented sampling strategy (Q90-UCB and EI) can serve as practical guidelines for other researchers in surrogate-assisted optimization?

3 comments
AuthorsOpenReview ↗

Response to Reviewer fTTV 1/2

Dear Reviewer fTTV,

We thank you for the insightful feedback and for highlighting the methodological significance of our distinction between global prediction error and maximum regret, as well as the modular integration of online feedback management into closed-loop AI Scientist workflows. We have substantially improved our paper, and all revised content is marked in red. We provide the following response:

[Q1 & W1] Novelty, Related Work, and Distinctions Between OSR and Prior Model Management Strategies.

We thank the reviewer for asking how OSR differs from earlier model-management methods. The contribution is the integrated agentic discovery loop: proposal generation continues between oracle measurements, while acquisition over retained proposals repairs the surrogate guiding that search. Sections 2.2 and 4.3 now cite and compare with lq-CMA-ES (Hansen, 2019), DTS-CMA-ES (Pitra et al., 2016), and TuRBO-1 (Eriksson et al., 2019). The formulation has three features:

  1. Decoupled Retrospective Archive Repair (AtA_t) vs. Intra-Generation Evolution Control (CtC_t): DTS-CMA-ES and lq-CMA-ES reuse previously measured points to fit their surrogates, while selecting new true evaluations from the current CMA-ES population. OSR retains unmeasured proposals across episodes, allowing a measurement to revisit an earlier design after search has moved on. In the new MADE ablation at T=100T=100, Q=10Q=10, archive EI achieves mean Top-1 of 128.3, versus 78.4 for periodic current-proposal evaluation and 79.9 for random archive selection (Table 5): [3.1] | Method (T=100,Q=10T=100, Q=10, 30 systems) | Top-1 | Top-5 | Top-5 AUC | EF-top | AF-top | Unmatched | | :--- | :---: | :---: | :---: | :---: | :---: | :---: | | Online EI (Archive AtA_t) | 128.3(92) | 85.1(53) | 48.3(39) | 1.400(114) | 5.94(68)x | 12/30 | | Q90-UCB (Archive AtA_t) | 108.1(85) | 66.7(63) | 36.1(40) | 1.231(108) | 4.72(77)x | 10/30 | | Original Control (Q=TQ=T) | 99.7(73) | 57.9(45) | 32.6(30) | 1.00 (ref) | 1.00 (ref) | 0/30 | | Random archive | 79.9(71) | 33.3(45) | 18.4(26) | 0.880(79) | 2.06(55)x | 2/30 | | Fixed surrogate (Q=0Q=0) | 78.7(115) | 40.0(70) | 26.7(51) | 0.864(130) | 4.17(81)x | 9/30 | | Periodic proposal (Current CtC_t) | 78.4(51) | 40.9(37) | 22.3(24) | 0.906(85) | 1.17(33)x | 0/30 |

  2. In-Context Repair of Pretrained Surrogates: Each new measurement becomes a conditioning-context row for TabPFN-3 or TabICLv2, updating subsequent predictions without training model parameters.

  3. Using Classical Surrogate Predictions in the Repair Model: LQ predictions or DTS GP means and variances become additional TabPFN inputs. The LQ combination lowers mean normalized gap from 0.363 to 0.304 at N=32N=32 and from 0.537 to 0.405 at N=1000N=1000, relative to native lq-CMA-ES (Table 3).

AuthorsOpenReview ↗

Response to Reviewer fTTV 2/2

[Q2 & W2] Experimental Bias from the SCM Prior and Generalization to Non-SCM Environments

We thank the reviewer for requesting an independent benchmark. The BBOB results show that the gains extend beyond our SCM generator. Online EI reduces mean normalized gap from 0.693 to 0.376 at N=32N=32 and from 0.541 to 0.439 at N=1000N=1000 relative to fixed TabPFN + EI. The LQ combination has the lowest mean normalized gaps, 0.304 and 0.405; standalone Online EI also leads all standalone optimizer baselines on normalized gap at N=1000N=1000. These comparisons cover all 24 independently defined BBOB functions in 5D and 10D, three instances each, with shared initial observations and a 32 oracle queries (Table 3; Appendix H):

MethodNorm. Gap (N=32N=32)Log Gap (N=32N=32)Norm. Gap (N=1000N=1000)Log Gap (N=1000N=1000)
TuRBO-10.326(16)1.820(33)0.506(17)1.573(36)
DTS-CMA-ES (EI)0.348(15)1.774(63)0.556(20)1.413(82)
Native lq-CMA-ES0.363(16)1.401(26)0.537(14)1.071(29)
DTS-CMA-ES (variance)0.460(17)1.976(64)0.748(19)1.767(62)
CMA-ES0.568(22)2.304(28)0.856(16)1.959(15)
Fixed TabPFN + EI0.693(21)2.356(28)0.541(17)1.580(26)
LQ + TabPFN + online EI0.304(15)1.423(29)0.405(15)1.051(22)
GP + TabPFN + online EI0.333(18)1.476(28)0.433(18)1.073(24)
Online TabPFN + EI0.376(17)1.958(29)0.439(17)1.465(24)
Online TabPFN + Peak0.428(17)2.083(26)0.465(16)1.499(22)

[Q3 & W1] Methodological Significance and Practical Guidelines for Surrogate-Assisted Optimization.

We thank the reviewer for asking about practical implications. Repair labels should be selected for their effect on optimization, and acquisition rules should be evaluated inside the continuing search. The results give three concrete guidelines:

  1. Evaluate repair by regret as well as global error: Global-UQ gives the largest mean NRMSE reduction, 0.0046 versus 0.0011 for EI, but its mean regret reduction is only 0.0631 versus 0.4093 (Table 1). In the agent loop, our regression adaptation of Tab-AICL reaches mean regret 0.1169 versus 0.0552 for Online EI (Table 2). These comparisons show why global coverage is an insufficient objective for optimization-directed repair.
  2. Select acquisition rules using closed-loop outcomes: Q90-UCB has a slight lead in the pooled static results, with mean regret reduction 0.4236 versus 0.4093 for EI. Online EI, however, obtains lower mean synthetic regret, 0.0552 versus 0.0600, and higher mean MADE Top-1, 128.3 versus 108.1. The pooled static ranking alone therefore does not identify the best online policy.
  3. Check acquisition geometry: Neither the narrowest nor the broadest query dispersion gives the lowest regret. EI's normalized step distance is 0.356, between Peak's 0.079 and Tab-AICL's 0.915, while their mean regrets are 0.0552, 0.0976, and 0.1169, respectively (Appendix C.2):
MethodStep/d0\text{Step} / d_0Spread/d0\text{Spread} / d_0Incumbent Updates / WorldMean Regret (↓\downarrow)
Online Peak0.0790.1896.10.0976
Online Q90-UCB0.1890.2556.00.0600
Online EI0.3560.4485.60.0552
Tab-AICL0.9150.9161.10.1169
Reviewer_fTTVOpenReview ↗

I'm holding my score.

Thanks to the authors for the detailed rebuttal. I have read it in full, and gone through the other reviewers' comments as well. Like them, I still see the same two weak spots doing most of the damage: novelty, and how far the evidence actually carries the claims. I'm holding my score.

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

This paper tries to address how a closed-loop AI scientist can use expensive high-fidelity feedback sparingly. The authors propose online surrogate repair (OSR): an agent searches mainly with cheap feedback from a tabular foundation model surrogate (TabPFN), and at a fixed number of evenly spaced episodes an acquisition rule, mainly expected improvement (EI), picks a design from the archive of all past proposals for an oracle evaluation, whose label is added to the surrogate's context. On synthetic worlds, they show that rules targeting the optimum (EI, Q90-UCB) reduce regret much more than rules targeting global fit, and that Online EI with 32 queries reaches lower regret than a control that queries the agent's choice in every one of 200 episodes. On the MADE materials discovery benchmark, they report that full-feedback controls need 6.36-7.23x (two LLM agents) and 10.27x (a non-LLM generator) as many oracle queries to match Online EI's discovery count.

Critical strengths

I think the problem is practical. In closed-loop discovery, proposing candidates is cheap while reliable evaluation is expensive, and separating how often the oracle is queried from how long the search runs is a reasonable way to frame this (pp. 1-2). Under the paper's own budgeted accounting there is also one result that holds up: with only 32 run-time oracle queries, the best material found reaches 95.0-98.4% of the Top-1 score that Control reaches with 300 queries (Table 8, pp. 19-20). The static study also makes a useful point: rules that target the optimum reduce maximum regret by 0.41-0.42, while rules that target global uncertainty reduce it by only 0.05-0.06 (Table 1, p. 6).

Critical weaknesses
  1. I feel the novelty is limited and the evidence is narrow. The core of OSR is to select points from a candidate pool by expected improvement and update the surrogate, which is standard Bayesian optimization; what is new is mainly that the candidates come from an LLM agent. Calling it "a novel algorithm" and "a novel third feedback regime" (pp. 1, 9) overstates this. The synthetic results even suggest that the gain comes from EI selection over the archive rather than from the agent: with the same 200 queries, EI-selected repairs reach a regret of 0.0278, while querying the agent's own choice reaches 0.0755 (Table 2, p. 7). Yet this strongest baseline, applying EI selection online in every episode, is not run on MADE, where Table 3 only compares Online EI with Vanilla, Control and Random. On the real side, MADE is the only benchmark, so I find it hard to judge whether the method would work in other discovery settings; the authors themselves note untested environments and surrogate classes (p. 9).

  2. I think the headline numbers depend on a post hoc audit, and the fair accounting is much weaker. The 6.36-7.23x and 10.27x in the abstract count discoveries found by auditing all 300 episodes' proposals with the oracle after the run, and these audit calls are not counted in the query budget (pp. 8, 19). In a real deployment only the 32 verified designs would be known. At the same number of search episodes, Online EI actually finds fewer discoveries than Control (82% and 67% of Control's count, p. 9; Fig. 3, left). Under the budgeted accounting that only uses run-time queries, the acceleration factor drops to 2.04 and 2.19, and to 0.99 for the non-LLM generator (Table 9, p. 20), and these numbers appear only in the appendix. I would expect the budgeted results to be the main ones.

Other weaknesses:

  • The synthetic study uses 20 worlds and a single LLM, apparently with one run per world, and reports no significance tests; the intervals of Online EI (Q=32) and Oracle Control (Q=200) overlap widely (0.0298-0.0843 vs. 0.0468-0.1069, Table 2).
  • Several definitions and descriptions are inconsistent. Regret is defined on the surrogate-searched design in Eq. 1 (p. 4) but on the best oracle-verified design in Table 2 (pp. 7-8). Static EI is sequential in Section 3.3 (pp. 5-6) but a single batch from a fixed pool in Appendix B (p. 15). Section 4.1 says Table 1 reports final values, but it only reports changes (p. 6). The limitations say no other surrogate class was tested (p. 9), while Archive-BO uses a GP (p. 7).
  • Details needed to reproduce the MADE results are missing, including the tabular featurization, the list of the 30 chemical systems, and the agent prompts.
Questions for authors
  1. Why are the audited discovery counts, rather than the budgeted ones, used for the headline claims in the abstract and Table 3?
  2. How does Online EI with 32 queries compare on MADE with applying EI selection over the archive in every episode, the strongest baseline in the synthetic study?
  3. Is there evidence beyond MADE that the method transfers to other discovery problems, or to surrogates other than tabular in-context models?
3 comments
AuthorsOpenReview ↗

Response to Reviewer WpnS 1/2

Dear Reviewer WpnS,

We thank you for your constructive feedback and for recognizing the practical value of decoupling oracle query frequency from search length, our static regret-vs-error findings (Table 1), and the strong search-time Top-1 material quality achieved with only 32 oracle queries. We have substantially improved our paper, and all revised content is marked in red. We provide the following response:

[Q1 & W2] Audited discovery counts vs. strictly budgeted search-time results in the Abstract and main text.

We thank the reviewer for emphasizing strictly budgeted evaluation. Online EI retains its advantage using only 32 search-time oracle queries: AF-top is 4.52-5.44x, corresponding to 77.9-81.6% fewer queries. Mean Top-1 is 126.0 versus 114.3 for Control with DeepSeek, and 133.7 versus 93.9 with Luna. These quality metrics use no post-hoc labels and are now written in the Abstract and main Table 4, search-time discovery counts are kept in Appendix G, Table 10:

WorkflowMethodTop-1Top-5Top-5 AUCEF-top (@32)AF-topUnmatched
DeepSeek-V4-FlashOnline EI (Q=32Q=32)126.0(63)104.7(43)77.1(34)1.119(37)4.78(43)x38/90
Vanilla (Q=32Q=32)110.8(65)78.3(39)54.7(33)0.968(20)2.67(47)x19/90
Control (Q=32Q=32 ref)114.3(60)84.1(34)59.8(30)1.00 (ref)1.00 (ref)0/90
GPT-5.6-LunaOnline EI (Q=32Q=32)133.7(64)111.9(49)87.3(41)1.546(119)5.44(42)x36/90
Vanilla (Q=32Q=32)94.7(48)67.2(39)39.2(29)1.082(74)2.21(29)x9/90
Control (Q=32Q=32 ref)93.9(50)66.3(39)36.8(28)1.00 (ref)1.00 (ref)0/90
Chemeleon+MLIPOnline EI (Q=32Q=32)125.5(57)106.1(38)89.1(34)1.005(14)4.52(50)x41/90
Vanilla (Q=32Q=32)124.8(56)103.0(37)85.1(37)1.000(16)4.28(55)x39/90
Control (Q=32Q=32 ref)124.9(54)106.1(36)90.2(32)1.00 (ref)1.00 (ref)0/90

The separately reported discovery yield (Disc.) measures all proposals across the MADE campaign. Post-hoc audit labels never enter the search or surrogate context, and audit oracle evlauations are separate from the feedback budget QQ. Without post-hoc evaluation, we would leave 268 of the 300 proposed structures unevaluated and artificially cap the reported discovery count at 32. We therefore report audited discovery yield alongside strictly search-time quality metrics, which use only the 32 search-time oracle measurements.

[Q2 & W1] Novelty, the role of the agent, and comparison with full-frequency archive EI.

We thank the reviewer for asking us to clarify the contribution and the agent's role. OSR's novelty lies in integrating retrospective archive acquisition and in-context surrogate repair into a continuing agentic discovery loop. The agent generates proposals and uses the repaired feedback, and acquisition selects which retained, unmeasured proposal receives an oracle label, which updates the surrogate's context without parameter training. DTS-CMA-ES and lq-CMA-ES select new evaluations from the current population, whereas OSR keeps earlier proposals eligible as search continues. This separates the oracle budget from search length and supports both LLM and Chemeleon proposal policies. The requested full-frequency archive-EI comparison is already our MADE Control. MADE generates one proposal per episode; with Q=T=300Q=T=300, each is immediately measured, so the unmeasured archive contains only the current proposal. Control is therefore exactly the Q=TQ=T case of archive EI. Online EI uses the same mechanism with Q=32Q=32. At T=100T=100, Q=10Q=10, the new ablation gives mean Top-1 128.3 for Online EI, versus 78.4 for periodic current-proposal evaluation and 79.9 for random archive selection (Table 5):

Method (T=100,Q=10T=100, Q=10, 30 systems)Top-1Top-5Top-5 AUCEF-topAF-topUnmatched
Online EI (Archive AtA_t)128.3(92)85.1(53)48.3(39)1.400(114)5.94(68)x12/30
Q90-UCB (Archive AtA_t)108.1(85)66.7(63)36.1(40)1.231(108)4.72(77)x10/30
Original Control (Q=TQ=T)99.7(73)57.9(45)32.6(30)1.00 (ref)1.00 (ref)0/30
Random archive79.9(71)33.3(45)18.4(26)0.880(79)2.06(55)x2/30
Fixed surrogate (Q=0Q=0)78.7(115)40.0(70)26.7(51)0.864(130)4.17(81)x9/30
Periodic proposal (Current CtC_t)78.4(51)40.9(37)22.3(24)0.906(85)1.17(33)x0/30
AuthorsOpenReview ↗

Response to Reviewer WpnS 2/2

[Q3 & W1] Evidence of transfer beyond MADE and to surrogates other than tabular in-context models.

We thank the reviewer for asking about transfer. The results demonstrate that OSR works beyond MADE and beyond tabular foundation surrogates. The Gaussian-process Archive-BO variant achieves mean regret 0.0807 in the synthetic agent loop (Table 2). On the new BBOB benchmark, online context repair reduces mean normalized gap from 0.693 to 0.376 at N=32N=32 and from 0.541 to 0.439 at N=1000N=1000. The LQ + TabPFN combination reaches the lowest means, 0.304 and 0.405. BBOB covers 24 functions in 5D and 10D, three instances each, with shared initial contexts and a 32-query allowance (Table 3):

MethodNorm. Gap (N=32N=32)Log Gap (N=32N=32)Norm. Gap (N=1000N=1000)Log Gap (N=1000N=1000)
TuRBO-10.326(16)1.820(33)0.506(17)1.573(36)
DTS-CMA-ES (EI)0.348(15)1.774(63)0.556(20)1.413(82)
Native lq-CMA-ES0.363(16)1.401(26)0.537(14)1.071(29)
DTS-CMA-ES (variance)0.460(17)1.976(64)0.748(19)1.767(62)
CMA-ES0.568(22)2.304(28)0.856(16)1.959(15)
Fixed TabPFN + EI0.693(21)2.356(28)0.541(17)1.580(26)
LQ + TabPFN + online EI0.304(15)1.423(29)0.405(15)1.051(22)
GP + TabPFN + online EI0.333(18)1.476(28)0.433(18)1.073(24)
Online TabPFN + EI0.376(17)1.958(29)0.439(17)1.465(24)
Online TabPFN + Peak0.428(17)2.083(26)0.465(16)1.499(22)

[W3 & W4] Statistical significance, definitional consistencies, and MADE reproducibility. We thank the reviewer for requesting clearer statistical comparisons and definitions. Paired BBOB comparisons support the benefit of online repair at both context sizes.

  • [W3] Statistical Significance: Online EI's paired normalized-gap improvement over fixed TabPFN + EI is 0.317 [0.271, 0.363] at N=32N=32 and 0.102 [0.066, 0.137] at N=1000N=1000 (95% intervals; Appendix H). On the 20 paired synthetic worlds, mean regret falls from 0.0984 to 0.0552 versus Static EI, a 44% reduction, and under noisy feedback, it falls from 0.1124 to 0.0905 (Table 2).
  • [W4] Definitions & Reproducibility: We thank the reviewer for requesting these clarifications. Equation 1 measures the regret of the design selected by optimizing the repaired surrogate, as evaluated in Section 4.1. Table 2 instead reports regret based on the oracle-measured designs, including the initial context. Section 4.1 acquires repair labels sequentially, the Static EI baseline in Section 4.2 serves the comparison to Online EI and acquires its labels in a single batch and then freezes the surrogate, as specified in Appendix B. We have corrected the Table 1 description in Section 4.1 and revised the feedback regime terminology in the Abstract, Introduction, and Section 5. Appendix I (pp. 22–23) now provides the 68-column crystal featurization, abbreviated orchestration prompt, episode feedback template, and the complete list of 30 chemical systems.

I thank the authors for the detailed response and for the effort that went into the new experiments. I have read it, together with the other reviews and the discussion. I accept the clarification on Q2, but my main concerns remain, and they are close to Reviewer Chng's.

First, novelty. I still do not see what is new beyond putting known parts together: EI selection over a pool of candidates, in-context updates of a pretrained surrogate, and sparse oracle evaluations during a surrogate-guided search. Describing the method as a new combination is not enough on its own. I would need to see what idea or insight the combination adds, and the response does not show this. Even with the explanation, the method still reads to me way more like a standard Bayesian optimization loop in which an LLM supplies the candidates, and I did not quite take away a new insight from it.

Second, the headline result. The 6-7x savings still pair discovery counts obtained by auditing all 300 proposals with only the 32 search-time queries, and the response confirms that the audit is needed to obtain these counts. Counting the same oracle calls on both sides, the factor is below one if all calls are counted. If only search-time calls are counted, it is about 2x for the two LLM workflows and 0.99x for Chemeleon + MLIP (Table 9 of the submitted version). I am also not convinced by the new AF-top. In the response's own table, the Chemeleon + MLIP Vanilla baseline is basically on par with Control in Top-1 (124.8 vs. 124.9), yet it still gets an AF-top of 4.28x.

Third, generalization. BBOB has no agent, so it does not show that the method carries over to other discovery problems, and MADE is still the only real discovery benchmark.

I appreciate the additional work, but I think the core issues of the paper are not resolved, and I still hesitate to say it reaches the level of novelty and solidity expected at ICLR. I will keep my score and have no further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Meta Response

We thank the reviewers for reading our manuscript revision and new experiments. Online Surrogate Repair (OSR) is a framework for agentic scientific discovery that lets a small number of high-fidelity measurements guide a much longer search. We are not aware of an earlier agentic framework that selects unmeasured solutions from the agent’s proposal history and uses their outcomes to update the context of the surrogate guiding later episodes. Below, we summarize the changes and additions we made during rebuttal in response to the reviewers’ comments.

Results. On MADE, OSR ran for 300 episodes with 32 oracle queries. At that same query budget, the mean score of the best material verified during search was 126.0 for OSR versus 114.3 for the full-feedback control with DeepSeek, and 133.7 versus 93.9 with Luna (Table 4). We also added a separate matched-budget MADE ablation, where the mean best verified score was 128.3 when expected improvement (EI) selected from the proposal history, 78.4 when verifying the latest proposal, and 79.9 when sampling the history at random (Table 5). Across 20 synthetic worlds, spending the 32 queries during search reduced mean regret by 44% relative to spending them upfront (Table 2). On BBOB with 32 initial observations, online repair reduced the mean normalized gap from 0.693 to 0.376 with the EI acquisition rule and query budget fixed for baseline methods we tested (Table 3).

Reviewer Chng asked what this framework teaches us beyond established surrogate-assisted optimization. In the MADE ablation, we kept search length and oracle budget fixed and varied which proposals were measured. Applying standard expected improvement (EI) to the agent’s proposal history yielded a higher mean best verified score than either measuring its latest proposal or sampling its history at random. The choice of repair policy also depended on how we judged the updated surrogate. Across both pretrained surrogate families, selecting labels by global uncertainty reduced average prediction error more than EI, yet the surrogate updated with EI-selected labels recommended better designs (Table 1).

Reviewer fTTV asked whether the synthetic results might reflect overlap between our tasks and TabPFN’s pretraining prior. We added the BBOB experiments to test online repair on independently defined functions and included comparisons with TuRBO and the evolution(CMA-ES) control methods as discussed in the rebuttal. At both initial context sizes, updating the surrogate during search reduced the mean normalized gap relative to a fixed surrogate with the same EI rule and oracle budget (Table 3).

Reviewer WpnS asked what the agent has actually verified when its measurement budget runs out, and why the discovery count required post-hoc evaluation. We explained that the discovery metric is multi-optima as defined by MADE, and post-hoc evaluation prevents the discarding of most of the agent's valid proposals. Even counting only discoveries verified during search, the full-feedback controls needed about twice as many queries to match OSR for the two LLM agents. We now report the quality of the best materials verified within budget in the abstract and main results (Table 4). We also agree with the reviewer on clarifying the scope of our evidence, and we have added BBOB to isolate the repair mechanism, and expanded MADE results as our applied benchmark for the full agentic workflow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.