acceptodds
Under review as a conference paper at ICLR 2027

When to Trust Your AI Co-Scientist? Learning Prior Reliability in Closed-Loop Bayesian Optimization

Abstract

AI “co-scientists” promise to accelerate discovery by proposing what to test next from knowledge distilled across millions of publications. Their reliability, however, depends on literature coverage: on under-studied targets they fabricate plausible candidates and fail silently, so a practitioner cannot tell a useful prior from a harmful one before committing costly experiments. Larger models mitigate but cannot eliminate failures where the literature is thin. We propose HARIBO, which learns a prior’s reliability from the experiments themselves. HARIBO adds the prior as a single feature of the Bayesian optimization (BO) surrogate. Its coefficient, fit to the sparse observations already being collected, provides an interpretable measure of trust that exploits, ignores, or inverts the prior as the data dictate. Across gene perturbation, protein variant effects, and antibody developability, and across language-model agents from 7B to frontier scale and protein foundation models, HARIBO improves on BO wherever the prior carries signal and matches it where the prior is noise. Under a systematically wrong prior, every scheme that fixes the prior’s weight in advance, by a constant, schedule, or model confidence, falls below BO, while HARIBO stays above it. The model is consulted before the first experiment and never again, which is enough: re-querying it every round instead, at 20–32× the requests and 124–234× the tokens, helps the resulting closed-loop agent on one of three domains and leaves it below knowledge-free BO on the other two. By learning trust from data rather than from model confidence, HARIBO lets AI co-scientists enter discovery loops safely.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI Co-Scientists, Bayesian Optimization

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_an8oOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses least-squares fitting to adaptively learn the trust of LLM-derived priors on Bayesian optimization. The methods are evaluated on six gene perturbation screens, 184 ProteinGym protein fitness assays, and 43 FLAb antibody developability assays.

Critical strengths

The figures and tables are visually polished and clearly presented.

Critical weaknesses
  1. Limited novelty. Previous work, including LLINBO: Trustworthy LLM-in-the-Loop Bayesian Optimization(https://arxiv.org/abs/2505.14756) and LABO (ICML 2026), has explored trustworthy use of LLMs in BO. The paper lacks both methodological and empirical comparisons with these approaches.
  2. Incremental contribution. Beyond updating the trust coefficient, could the prior itself evolve with experimental feedback?
Questions for authors
  1. Could the authors evaluate generalization to other domains, such as AutoML?
  2. How does the method perform in high-dimensional optimization settings?
  3. Can prior reliability be estimated accurately from very few observations? I remain unconvinced and would appreciate further evidence.
12 comments
AuthorsOpenReview ↗

Response to Reviewer an8o (1/3)

We thank Reviewer an8o for the review, for finding our figures and tables "visually polished and clearly presented", and for the pointers to LLINBO and LABO. We address each point with new experiments, all run in the paper's loop with the same features, seeds and paired Wilcoxon tests over datasets.


[W1. Comparison with LLINBO and LABO]

Thank you for these pointers. Both are relevant, and we now discuss them in §2 and add empirical comparisons using adaptations to our setting. Our contribution is a specific mechanism for incorporating an unreliable prior: fitting its signed coefficient jointly with candidate features, allowing the surrogate to exploit, ignore, or invert the prior without further LLM queries during the discovery loop.

What they do.

  • LLINBO (Chang et al., 2025) wraps an LLM optimizer with a GP. At every iteration the LLM proposes a candidate, and one of three rules decides whether to use it:
    • Transient follows the LLM with a probability that decays as pt=min⁡(t2/T,1)p_t = \min(t^2/T, 1) grows;
    • Justify accepts the LLM's point only if its UCB is within ψt\psi_t of the GP's best;
    • Constrained conditions the GP on the LLM's point being an improvement.
  • LABO (Chen et al., ICML 2026) treats the LLM as a low-fidelity simulator in a Kennedy–O'Hagan model, fR(x)=ρ fL(x)+δ(x)f_R(x) = \rho\, f_L(x) + \delta(x). Here ρ\rho is fitted by least squares on the real observations, and a gate decides, per candidate, whether to run a real experiment or only query the LLM.

How HARIBO differs.

  1. Cost. Both methods query the LLM inside the loop: LLINBO at every iteration, LABO at its initial design and at the candidates it selects. HARIBO uses one initial elicitation round, before the first experiment (5.9k tokens per gene screen; Table 5).
  2. Setting. Both target continuous search spaces with GP surrogates. Ours are finite pools of 10210^2–1.5×1041.5\times10^4 discrete candidates (genes, mutants, antibodies), often with sparse priors such as a 120-gene list.
  3. How the LLM's signal is weighted. LLINBO decides whether to follow the LLM. LABO and HARIBO both learn a signed scale from labels, but LABO fits it on the prior alone, while HARIBO fits it jointly with the features (Prop. 1(a)).

Empirical comparison. Both methods need an LLM inside the loop, so we adapted their decision rules to our one-shot prior, with the LLM's "suggestion" at each round being its top-ranked unmeasured candidates:

  • LLINBO-Transient and LLINBO-Justify (ψt=sd(yobs)/t\psi_t = \mathrm{sd}(y_{\text{obs}})/t);
  • LABO-style: ρ\rho fitted by least squares of yy on [1,π][1, \pi], then a ridge discrepancy on ϕ\phi. Its gate is moot here, because π\pi is available for every candidate.

Table R1. LLINBO vs. HARIBO, Recall@R10 (protein: 184 assays, PCA-50; gene: 6 screens; 10 seeds). All methods use the same cached prior; these comparisons evaluate adapted decision rules rather than the full original LLM-in-the-loop systems.

methodprotein +ESM2protein −ESM2protein, Qwen priorgene Sonnet 5gene Sonnet 5, flipped
pure BO0.5590.5590.5590.1160.116
LLINBO-Transient (adapted)0.6010.5160.5620.1460.112
LLINBO-Justify (adapted)0.6530.5070.5780.1370.104
HARIBO0.7260.7050.5940.1390.136

What the table shows. LLINBO's rules guarantee no-regret asymptotically, but in this short-horizon, batch regime they are conservative in both directions.

  • Good prior: on protein they recover only +0.04 to +0.09 of HARIBO's +0.17.
  • Wrong prior: both rules fall below BO (protein −0.04-0.04 and −0.05-0.05, gene −0.004-0.004 and −0.012-0.012), because they only filter the LLM's suggestions and never invert them. HARIBO learns a negative weight and stays far above BO (0.705 vs. 0.559; 0.136 vs. 0.116).
AuthorsOpenReview ↗

Response to Reviewer an8o (2/3)

Comparison with a LABO-style adaptation.

LABO is the closest method to ours: like wπw_\pi, its KOH scale ρ\rho is fitted to labels and free in sign. The difference is how that scale is fitted. LABO regresses yy on the prior alone, whereas HARIBO fits wπw_\pi jointly with the features (Prop. 1(a)).

Table R2. LABO-style vs. HARIBO with LLM priors, Recall@R10 (10 seeds).

protein (protein_matched), LLM priorfeaturesLABO-styleHARIBOpp
Qwen2.5-32B (212)PCA-500.5800.5920.02
Qwen2.5-32B (212)full one-hot0.6040.6160.15
Claude Sonnet 5 (184)PCA-500.6300.6340.14
Claude Sonnet 5 (184)full one-hot0.6530.6634×10−44\times10^{-4}

HARIBO is ahead in every cell. With ESM2 and on gene the two agree within 0.006 (n.s.), and both survive the flipped prior, unlike LLINBO.

Cost. This comparison removes LABO's in-loop LLM. As published, LABO uses the LLM as a low-fidelity simulator and evaluates it at its initial design and at the candidates it selects. HARIBO needs one initial elicitation round before the first experiment (5.9k tokens per gene screen) and a closed-form ridge fit on a CPU (39 ms for ten rounds).

The key distinction is where adaptation occurs: LLINBO filters LLM proposals, and LABO calibrates a low-fidelity signal before fitting a discrepancy. HARIBO fits the prior coefficient jointly with the candidate features, estimating the prior's value beyond them, so even a fixed prior can be exploited, ignored, or inverted as evidence accumulates, without further LLM queries.

Added in Sec. 2 and "LLM-in-the-loop BO baselines" (new Appendix E, blue).


[W2. Can the prior itself evolve with experimental feedback?]

Thank you for this suggestion. Yes, and we see three ways, two of which we tested.

  1. Re-elicit the prior after measurement. Qwen2.5-32B sees its first measured batch, with or without tools (gene search, pathways, the BO surrogate), and returns a revised prior. The revision improves prior quality on proteins, antibodies and an AutoML benchmark (LCBench), but lowers it on genes (hit rate 0.088 → 0.056). HARIBO's ten-round recall moves by at most 0.02 either way (gene 0.131 → 0.129, protein 0.595 → 0.598, antibody 0.615 → 0.633, LCBench 0.622 → 0.619 without tools; tools do not add to this), while the revision costs 113k–122k extra tokens against one-shot elicitation's 5.9k.
  2. Re-query every round with measured feedback. This is the closed-loop co-scientist we already compare against (§4.4; 20–32× the requests). From round two its prompt contains the measured score and hit status of every candidate it proposed. Yet the feedback buys little:
    • Qwen2.5-32B: after ten rounds of feedback the agent is below one query plus HARIBO in all three domains (gene 0.109 vs. 0.131, protein 0.505 vs. 0.590, antibody 0.464 vs. 0.599), and below knowledge-free BO on protein and antibody.
    • Claude Sonnet 5: the agent leads on gene (0.169), but one wider call plus HARIBO, with no feedback to the LLM, reaches 0.163 (App. C.8), so the remaining recall gap is 0.006. Its gene gain sits on the three best-covered screens (App. C.6). On protein, the difference from one-round HARIBO is not statistically significant, and on antibody the agent ends below BO.
    • BioDiscoveryAgent's critic and tools every round do not help either: 0.095 with tools vs. 0.098 without, against 0.128 for HARIBO (3 gene screens).
  3. Let the evolving prior enter as a new column. HARIBO's construction already accommodates this, because kk prior columns get kk learned coefficients (§3). Its weight is again learned from the labels, so an updated prior is not trusted blindly either. We now state this as an extension for in-the-loop agents.

In short, letting the prior evolve can improve the prior, but the tested prior revisions produced modest changes in downstream recall while multiplying the cost. One-shot elicitation stays the default; the evolving variants are reported as options.

Added in "Prior revision with data and tools" (new Appendix E, blue).


[Q1. Generalization to other domains such as AutoML]

We added an AutoML benchmark, LCBench (Zimmer et al., 2021; MLP hyperparameters on OpenML datasets). The prior comes from one LLM request per dataset, which returns a recommended range and importance for each hyperparameter, converted into a score per configuration.

The pattern is the same as in biology. The prior is weak on its own (0.177 vs. 0.514 for BO), yet HARIBO turns it into a gain over BO at both budgets: 0.605 vs. 0.514 at 10% of the pool, and 0.388 vs. 0.275 at 5% (both p<10−5p < 10^{-5}). With the prior's sign flipped, the fixed-weight schemes collapse (fixed blend 0.157, π\piBO-style 0.141 vs. BO 0.514 at 10%). HARIBO stays near BO at 10% (0.502) and above it at 5% (0.317 vs. 0.275).

Added in Sec. 4 and the appendix (blue).

AuthorsOpenReview ↗

Response to Reviewer an8o (3/3)

[Q2. High-dimensional settings]

We evaluate high-dimensional candidate representations within finite pools. The paper compresses features with PCA-50; we reran without compression.

  • Protein: the full one-hot encoding of each mutant (position and residue indicators, 75–707 dimensions, median 450), on 212 assays.
  • Gene: raw STRING-800 + scGPT-512 + DepMap profiles concatenated, 2,120 dimensions, with no PCA. The pool is restricted to the genes all three sources cover.

Table R3. Recall@R10, compressed vs. uncompressed features.

domain, priorfeatures (dim)pure BOHARIBOHARIBO − BO
protein, +ESM2PCA (50)0.5530.700+0.147
one-hot (~450)0.5940.717+0.124
protein, −ESM2PCA (50)0.5530.684+0.131
one-hot (~450)0.5940.706+0.112
protein, Sonnet 5PCA (50)0.5590.634+0.076
one-hot (~450)0.5990.663+0.065
gene, Sonnet 5STRING-PCA (50), same pool0.1910.215+0.024
raw (2,120)0.2510.291+0.041
gene, Qwen2.5-32BSTRING-PCA (50), same pool0.1910.209+0.017
raw (2,120)0.2510.279+0.028

HARIBO stays well above BO in high dimensions. On gene the richer 2,120-dimensional features lift BO from 0.191 to 0.251, and HARIBO's margin grows rather than shrinks (+0.024 → +0.041 with Sonnet 5, where it also beats the fixed blend, 0.291 vs. 0.272). On protein the margin stays large, including under the adversarial prior, where the fixed blend and π\piBO-style fall to 0.421, 0.17 below BO. (Gene numbers are on the ~9k genes all three sources cover, so ten batches of 128 screen ~14% of the pool; all arms share that pool.)

The construction scales with dimension because the prior is a single extra column. Its coefficient is left unpenalized, while the dd feature coordinates share a ridge penalty chosen by leave-one-out cross-validation, which matters most when dd approaches or exceeds the number of labels (App. A.2). The fit stays closed-form: for nn labels and dd features it costs O(min⁡(n,d)2max⁡(n,d))O(\min(n,d)^2 \max(n,d)) per round, so it runs on a CPU at either dimensionality.

Added in "Budget and feature strength" (new Appendix E, blue).


[Q3. Can prior reliability be estimated from very few observations?]

Thank you for raising this. It is central to the method, because round one has no labels and the trust must be learned on the fly. We tested it directly by shrinking the batch and recording, each round, whether the sign of w^π\hat w_\pi matches the population coefficient γ⋆\gamma^\star computed from all labels. Protein: 184 assays, 10 seeds. "Labels" is the number of measurements behind the estimate.

Table R4. Recall@R10 and sign accuracy of w^π\hat w_\pi with tiny batches.

priorbatchlabels at 1st / 2nd / 3rd estimatesign of w^π\hat w_\pi correctBOfixed blendπ\piBO-styleHARIBO
+ESM224 / 6 / 881% / 84% / 85%0.0650.0990.1070.110
+ESM244 / 8 / 1271% / 84% / 89%0.1240.2050.2060.210
+ESM288 / 16 / 2477% / 88% / 93%0.2400.3530.3770.379
−ESM224 / 6 / 878% / 80% / 82%0.0650.0290.0300.102
−ESM244 / 8 / 1279% / 80% / 84%0.1240.0700.0610.189
−ESM288 / 16 / 2484% / 86% / 89%0.2400.1520.1410.359

With only four labels, the learned coefficient already has the right sign on 71–81% of runs, and the accuracy keeps rising as labels accrue. That is enough to matter: with two measurements per round (20 in total), HARIBO under an adversarial prior reaches 0.102, against 0.065 for BO and 0.03 for the fixed weights. A weight fixed in advance cannot recover from a wrong sign at any budget.

On genes (6 screens), HARIBO beats BO on all six screens at every batch size from 8 to 64 for Sonnet 5 (p=0.031p = 0.031, the minimum attainable with six screens), e.g. 0.064 vs. 0.026 at batch 32. At batch 8 the sign is close to chance (57%), because round one contains just 4 nominated genes.

These results provide empirical evidence that useful prior weighting can emerge from very few observations in the tested settings. Equation 3 motivates collecting contrasting prior values in the initial batch under its stated OLS assumptions; it is not an exact variance guarantee for the regularized estimator in the few-label regime. Reliability remains task-dependent, as the gene result at batch 8 shows.

Added in "Few observations" (new Appendix E, blue).

We hope these clarifications and experiments address your concerns, and we would be glad to clarify anything further.

Reviewer_575UOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces HARIBO, a method that learns how much to trust an LLM prior inside Bayesian optimization. The LLM is queried once to score candidates and this score is added as a feature of a ridge surrogate. Its coefficient, refitted from experimental outcomes each round, acts as a signed "trust" that can exploit, ignore or invert the prior. The idea is motivated by an evaluation showing that LLMs hallucinate more on niche tasks, depending on how well a specific problem is covered in the literature, and that they cannot flag their own fabrications. Validation is retrospective on 6 CRISPR screens, 184 ProteinGym and 43 FLAb assays. It measures the fraction of top candidates found after 10 rounds, against plain BO, prior-weighting baselines (fixed, decaying, confidence-scaled and learned blends) and an LLM agent re-queried every round.

Critical strengths

A well motivated and important problem.

Deciding how much to trust LLM priors within an optimization strategy is very important, and the motivation in Fig. 1 and Table 10 is valuable. LLMs underperform on niche problems not covered by their prior, which is exactly the setting of new discovery pipelines. Combining LLM priors with Bayesian optimization "safely", achieving at least the result of either and preferably the better one, is a promising direction for the field.

Cheap and practical.

The method requires no fine-tuning and works with a single query and a closed-form fit. It can be easily plugged in with proprietary models, as the authors show. The paper also asks many relevant questions and is honest about what did not work.

Critical weaknesses

The robustness evidence covers only a globally inverted prior.

The central claim is that HARIBO stays safe under a misleading prior where fixed-weight schemes fail (Table 4). The sign-flip experiment demonstrates the value of adapting the prior's influence from observations rather than committing to a fixed positive weighting. Methods that learn the prior's effect from data recover almost fully: the learned-λ blend loses 0.015 relative to the good prior (0.688 vs 0.703, Table 4), the GP variant loses 0.014 (0.754 vs 0.768, Table 11, n = 104), and HARIBO loses 0.020 (0.705 vs 0.725). No single fixed weight performs well under both the original and the inverted prior (Table 12).

However, this is the only deliberately misleading prior with a full comparison. The misapplied gene prior keeps a positive weight (Fig. 7), so a harmful gene prior is never evaluated. Realistic failures would involve a mixed prior that's informative in some regions and misleading in others, which a single global coefficient can only average. This is not tested, and the local-trust variants in A.3 use globally good or bad priors and an older protocol.

In the released code, the baselines differ only in the strength of a positive prior weight. πBO and BOPrO follow the same decay up to a factor of 10, and LGBO is a constant 0.7. Their losses under the inverted prior follow that weight strength. These are reasonable weighting schedules, but the comparison shows an advantage over these schedules rather than over the cited methods as published. The §4.2 claims that πBO's trade-off is especially sharp and that LGBO fails "because self-report is no substitute for labels" reflect the chosen constants, not those methods' mechanisms.

The evaluation setup makes the reported gains hard to judge.

On proteins and antibodies, half of each pool is measured (n/20 × 10 rounds), so random selection should already reach ~ 0.50 recall, and pure BO, greedy ridge on PCA-compressed one-hot features, reaches only 0.560 on proteins. On genes, the authors themselves describe the STRING features as weakly informative (C.7), and pure BO (0.111) is close to the random rate (~0.085). Since HARIBO only weights the part of the prior not explained by the features, the gains may shrink with stronger standard features.

Fig. 8 appears to show the re-queried Sonnet agent outperforming HARIBO on proteins, whereas §4.4 reports HARIBO winning over the agent. The caption also claims that the agent loses. This would change the abstract's conclusion that re-querying helps on only one of three domains.

Questions for authors

Clarifications:

  • Results under realistic wrong or mixed priors: a prior corrupted in a controlled fraction of the candidate space (recall vs fraction corrupted), a gene list nominated for the opposite phenotype direction, and the naturally negative antibody assays reported separately.
  • A realistic budget (5–10% of the pool), and results with stronger standard features that are independent of the prior (e.g. uncompressed one-hot as in Hsu et al., or ESM2 embeddings when the prior comes from the LLM)
  • Fig. 8: which protein result is correct for HARIBO, and with which prior? Does the re-queried agent beat HARIBO and BO on proteins?
  • How is λ fitted in the learned blend, and why does it score below λ = ½ under +ESM (0.703 vs 0.711)?
  • How do you define the “uninformative” prior (Figs. 4, 7).

Suggestions:

  • Shared initialization. Methods start from different first batches: random for BO, 50/50 for HARIBO, all-prior for the weighting baselines. Rerunning all methods from a shared first batch (e.g. HARIBO's 50/50) would separate the effect of initialization from that of the weighting rule. This matters particularly on genes, where HARIBO's first batch contains about half the nominated genes and BO's almost none. Fig. 8 suggests this matters: on genes, HARIBO's lead over BO is established in rounds 1–2, while the nominated genes are being measured, and stays roughly constant afterwards.
  • One-shot vs in-the-loop use of the LLM. The paper presents a one-shot prior and a re-queried agent as alternatives. However, HARIBO’s fallback is one-sided: when the features are weak, it can only rely on a frozen prior, which on genes is exhausted after one or two rounds, while Fig. 8 suggests a frontier model adds value by learning from results during the loop. A discussion of this limitation, or a small experiment that queries the LLM again for an updated prior (e.g. when progress stalls) while keeping the learned weight, would clarify whether HARIBO replaces such agents or can serve as a safety layer for them.
  • Positioning. While looking into related literature, I found Hsu et al. (Nat. Biotechnol. 2022), whose “augmented” ridge model combines one-hot features with a zero-shot score whose coefficient is effectively unpenalized, while the other features are regularized with cross-validated penalties. More broadly, combining zero-shot priors with learned regression models appears to be used in protein fitness prediction. I understand that HARIBO extends beyond this setting by interpreting the learned coefficient as signed trust in LLM priors within optimization pipelines, but could the authors position their work relative to broader related literature and clarify the main methodological contribution?

[1] Hsu, Chloe, et al. "Learning protein fitness models from evolutionary and assay-labeled data." Nature biotechnology 40.7 (2022): 1114-1122.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 575U (1/4)

We thank Reviewer 575U for an unusually careful review: it describes the problem as "well motivated and important", the method as "cheap and practical", and the paper as "honest about what did not work". Your reading of the released code is accurate, and it led to concrete improvements. Below we say what we changed and report the experiments you asked for. All new runs use the paper's loop, features, seeds and statistics (paired Wilcoxon over datasets) unless stated otherwise.


[W1 & Q1. Robustness beyond a globally inverted prior]

We ran the three realistic failure modes you proposed.

(a) Mixed prior: right in some regions, wrong in others. On ProteinGym (protein_matched, n=212n=212 assays, 10 seeds) we flip the sign of the ESM2 prior on a contiguous block of residue positions that covers a fraction ff of the pool. The prior is then informative on one part of the protein and misleading on the rest.

Table R1. Recall@R10 against the corrupted fraction ff (pure BO = 0.553 throughout).

methodf=0f=00.250.50.751.0
fixed blend (λ=12\lambda=\frac12)0.6960.6190.5560.4820.409
π\piBO-style0.7070.6270.5520.4730.398
BOPrO-style0.6210.5800.5340.5020.481
LGBO-style0.6810.6090.5470.4920.428
learned-λ\lambda blend0.6830.5950.5440.5860.670
HARIBO0.7000.6150.5630.5980.684
HARIBO, 4 local coefficients0.6940.6110.5650.5940.673

HARIBO stays at or above BO at every corruption level. At f=0.5f=0.5 the prior's net signal is about zero; HARIBO's global coefficient then lands near zero and the loop returns to BO (+0.010, p=0.16p=0.16). All tested fixed or decaying schedules fall below BO at f=0.75f=0.75 and f=1f=1, losing up to 0.155.

Your point about averaging is also borne out. At f=0.25f=0.25 a single coefficient does no better than the fixed weights (0.615 vs. 0.619–0.627), so it does not fully exploit the clean 75% of the pool. Giving HARIBO one coefficient per region (four kk-means clusters of ϕ\phi, each with its own unpenalized column π⋅1[ck]\pi\cdot\mathbb{1}[c_k]) does not help either, consistent with App. C.3. We will therefore describe the global coefficient as a global mechanism for mitigating prior harm, rather than an oracle for local reliability, and state this as a limitation.

(b) A harmful gene prior. Sign-flipping the nomination indicator turns each list into "everything except what the model named". Gene, 6 screens, 20 seeds:

priorpure BOfixed blendπ\piBO-styleHARIBO
Qwen2.5-32B, flipped0.1110.1080.1080.117
Claude Sonnet 5, flipped0.1110.1070.1070.126

Both fixed weights fall below BO, while HARIBO reads the sign off the labels and stays above it. The opposite-direction list you suggested (tau-up nomination on the tau-down screen and vice versa, 3 models) turns out not to be harmful, because the benchmark scores |LFC|. HARIBO still ends above BO on average (0.109 vs. 0.104).

(c) Naturally negative antibody assays, reported separately. ESM2 naturalness is negatively correlated with fitness on 15 of the 43 FLAb assays.

subsetBOprior armfixed blendπ\piBO-styleBOPrO-styleLGBO-styleHARIBO
corr < 0 (n=15n=15)0.6990.2560.5380.4750.6840.6160.672
corr < −0.2 (n=7n=7)0.8870.1710.5900.4670.8350.7160.916

On these naturally adversarial assays the fixed blend and π\piBO lose 0.16–0.42 relative to BO. Across all 15, HARIBO is numerically below BO by 0.027, although the difference is not statistically significant (p=0.28p=0.28). On the seven strongly negative assays it inverts the prior and finishes above BO (0.916 vs. 0.887), and significantly above the blend and π\piBO (p=0.016p=0.016).

Added in App. E.13, "Mixed and naturally adversarial priors" (Tables 20–22, blue).

AuthorsOpenReview ↗

Response to Reviewer 575U (2/4)

[W1 (cont.). The baselines are weighting schedules]

You are right, and we accept the framing. All four baselines run in our loop on the same ridge surrogate, so they differ only in the weight placed on a positive prior:

  • π\piBO-style and BOPrO-style decay that weight with the round and with the number of labels;
  • LGBO-style is a constant c=0.7c=0.7.

We made this choice deliberately, to isolate the weighting rule, but the arms are schedules, not reimplementations. In the revision:

  1. We rename them "π\piBO-style", "BOPrO-style" and "LGBO-style (constant confidence)".
  2. We remove the two §4.2 sentences you flag. "The trade-off is sharpest for π\piBO" becomes a statement about the schedule with the largest early weight. "LGBO fails because self-report is no substitute for labels" is deleted, because our arm never varies cc.
  3. We add three faithful-mechanism checks:
    • π\piBO as published. π\pi is turned into a density p(x)p(x) (uniform over the top 10%, Zipf over the ranking, or softmax). The first batch is sampled from pp, and acquisition is α(x) p(x)β/n\alpha(x)\,p(x)^{\beta/n}. Across all 12 variants, every one falls below BO under the flipped prior (0.384–0.516 vs. 0.560), while HARIBO holds 0.705 (App. E.16; our reply to nJEx, Table R1).
    • LLINBO-style (Chang et al., 2025; transient and justify rules).
    • LABO-style (Chen et al., ICML 2026; a KOH scale ρ\rho fitted by least squares).

The last two are detailed in App. E.17 and our reply to an8o. In brief, the two LLINBO rules fall below BO under the flipped ESM2 prior (0.516 and 0.507 vs. 0.559). LABO's learned scale ρ\rho is the only one that survives, because it is also a label-fitted, sign-free weight. It performs similarly to HARIBO on ESM2; with LLM priors, HARIBO is numerically higher in all four comparisons, significantly in two. The ingredient that separates safe methods from unsafe ones is a weight fitted to labels and free in sign, which is the paper's thesis.


[W2 & Q2. Budget and feature strength]

(a) A realistic budget. On gene the budget is already realistic: ten batches of 128 out of ~15k genes is about 8.5% of the pool. On protein we reran at a total budget of 10% and 5% of each pool (batch n/100n/100 and n/200n/200, 184 assays, 10 seeds). Random selection reaches about 0.10 and 0.05.

Table R2. Protein recall@R10 at 10% / 5% of the pool.

priorpure BOfixed blendπ\piBO-styleLGBO-stylelearned-λ\lambdaHARIBO
+ESM20.118 / 0.0650.194 / 0.0990.195 / 0.1070.186 / 0.0990.174 / 0.0890.201 / 0.110
−ESM20.118 / 0.0650.064 / 0.0290.059 / 0.0300.071 / 0.0330.171 / 0.0870.180 / 0.102
Qwen2.5-32B0.118 / 0.0650.123 / 0.0640.124 / 0.0680.122 / 0.0630.110 / 0.0510.124 / 0.064

At these budgets BO is close to random, as you predicted. HARIBO improves on it by 70% under a good prior, and it is the best method in both columns under an adversarial one, where all tested fixed-weight schedules substantially reduce recall relative to BO (0.059–0.071 vs. 0.118). The weak Qwen prior gives no gain at these budgets, and HARIBO does not lose either.

(b) Stronger features, independent of the prior. We replaced PCA-50 with the uncompressed one-hot encoding of Hsu et al. (position and mutant-residue indicators, median 450 dimensions) on protein_matched.

Table R3. Recall@R10, PCA-50 → full one-hot. "Margin" is HARIBO minus BO on the same features.

priorpure BOHARIBOmarginfixed blend / π\piBO-style
+ESM20.553 → 0.5940.700 → 0.717+0.147 → +0.1240.696 / 0.707 → 0.720 / 0.725
−ESM20.553 → 0.5940.684 → 0.706+0.131 → +0.1120.409 / 0.398 → 0.421 / 0.421
Claude Sonnet 50.559 → 0.5990.634 → 0.663+0.076 → +0.0650.643 / 0.648 → 0.666 / 0.672
Qwen2.5-32B0.553 → 0.5940.592 → 0.616+0.039 → +0.0220.583 / 0.579 → 0.602 / 0.591

Stronger features raise BO by 0.04. HARIBO's margin shrinks by 0.01–0.02 but stays significant in every row (p≤0.004p\le 0.004). This is what Prop. 1 predicts, since wπw_\pi weights only the part of π\pi the features do not explain. The robustness advantage persists with stronger features: under −ESM2 the fixed weights remain 0.17 below BO, while HARIBO is 0.112 above it.

On gene, stronger features independent of the prior (raw STRING-800 + scGPT-512 + DepMap profiles, 2,120 dimensions, no PCA) lift BO from 0.191 to 0.251 on the shared pool. HARIBO's margin grows there rather than shrinks, from +0.024 to +0.041 with Sonnet 5 and from +0.017 to +0.028 with Qwen2.5-32B (details in our reply to an8o, Q2).

Added in App. E.14, "Budget and feature strength" (Tables 23–24, blue).

AuthorsOpenReview ↗

Response to Reviewer 575U (3/4)

[W3 & Q3. Fig. 8 contradicts §4.4]

Thank you for the detailed comments. We identified both a plotting error and an incorrect attribution of the prior used for the result reported in §4.4. We have corrected the comparison and are revising the associated conclusions.

What went wrong. The protein and antibody panels of Fig. 8 drew the HARIBO curve from the Qwen2.5-32B prior, whatever agent was plotted. The caption ("Sonnet 5 as both the agent and the source of our prior") was therefore false. Separately, the 0.703 that §4.4 reports for protein is HARIBO given the ESM2 prior, not the Sonnet prior.

Corrected comparison. Every arm is scored on one pool per panel. Protein uses the 184 matched-pool assays that have both a Sonnet prior and a Sonnet agent run, so its BO differs slightly from other tables.

Table R4. Recall@R10, Claude Sonnet 5 as both the re-queried agent and the source of our prior.

domainpure BOagent (10 rounds)HARIBO, Sonnet prior (1 round)HARIBO, ESM2 prioragent − HARIBO (Sonnet)
gene (6)0.1110.1690.139–+0.030, p=0.16p=0.16
protein (184)0.5610.6570.6370.713+0.020, p=0.14p=0.14
antibody (36)0.6040.4490.5740.597−0.125, p=0.003p=0.003

On protein, the Sonnet agent achieves numerically higher recall than HARIBO with the Sonnet prior (0.657 vs. 0.637), but the difference is not statistically significant (p=0.14p=0.14). Both outperform BO, and HARIBO is ahead once given the ESM2 foundation-model prior (0.713 vs. 0.657). On antibody, where the Sonnet prior carries no signal, the agent falls 0.155 below BO. HARIBO substantially outperforms the agent, but remains 0.030 below BO (p=0.033p=0.033), illustrating that learned weighting does not eliminate all costs of an uninformative prior.

We revised the abstract, §4.4 and Fig. 8 (redrawn, App. E.6), all in blue. The abstract now reads: "Re-querying it every round instead, at 20–32× the requests and 124–234× the tokens, helps a frontier closed-loop agent on gene, gives numerically higher but not significantly different recall on protein, and leaves it below knowledge-free BO on antibody."


[Q4. How λ is fitted in the learned blend, and why it is below λ=1/2]

Each round, lam_learned computes out-of-fold ridge predictions μ^oof\hat\mu_{\text{oof}} on the observed rows (5-fold). It then fits y≈c0+w1z(μ^oof)+w2z(π)y \approx c_0 + w_1 z(\hat\mu_{\text{oof}}) + w_2 z(\pi) by ordinary least squares, with w2w_2 free in sign, and scores candidates with w1z(μ^)+w2z(π)w_1 z(\hat\mu) + w_2 z(\pi), where μ^\hat\mu is the full-data fit.

It scores below λ=12\lambda=\frac12 under +ESM2 (0.703 vs. 0.711). Two plausible contributors are:

  1. Estimation noise, and 12\frac12 is nearly optimal. The best fixed λ\lambda under +ESM2 is 0.6 (Table 13), so the blend at λ=12\lambda=\frac12 sits close to an oracle that is chosen in hindsight and pays no estimation cost. The learned blend has to estimate two weights from 20–180 labels.
  2. Attenuation of w1w_1. Out-of-fold predictions from a small training set are noisy, which attenuates w1w_1 (errors-in-variables) and tilts the ratio toward the prior.

HARIBO avoids the separate stacking stage by jointly fitting the prior coefficient and feature coefficients (Prop. 1(b)), although it remains subject to finite-sample estimation error. Under −ESM2, where 12\frac12 is the wrong value, the ordering reverses (0.688 vs. 0.410).


[Q5. The "uninformative" prior in Figs. 4 and 7]

It is a random prior with the same marginal distribution as the real one:

  • gene: a uniformly random set of genes of the same size as the model's nomination list;
  • protein: the ESM2 score randomly permuted across candidates.

It therefore has the same sparsity or scale as the real prior and zero expected signal. We now state this in both captions.

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

In a variety of settings black box optimization settings, there is some existing system that can provide some useful score pi(x) indicating the suitability of a candidate point x for having high f(x). For example, we can ask an agentic system to rate candidate x or we can use a specialist model like ESM for proteins. Then, the question becomes how to incorporate pi(x) into the optimization. This paper argues that pi(x) shouldn't be treated as a formal prior over solutions, since pi(x) may be unreliable/uninformative for the task at hand. Instead, the paper argues that pi(x) should be treated as an additional feature used in Bayesian optimization when characterizing x. Then, standard Bayesian regression machinery will adaptively decide how much to trust pi(x).

Critical strengths

The paper addresses a practical question: how do we merge traditional optimization approaches such as Bayesian Optimization with the prior knowledge available in foundation models?

Critical weaknesses

The proposed algorithm is very simple. Instead of using a feature vector phi(x) to characterize a point in a problem's search space, it uses [pi(x); phi(x)], where pi(x) is a scalar feature from a 'prior' system. Once these are concatenated, the approach uses standard machinery for Bayesian linear regression.

The notion of a 'prior' is imprecise. What is this a distribution over? Is pi(x) the prior probability that x is the solution? that x has non-zero f(x), etc? Also, pi(x) is not treated as a prior in any formal sense in the proposed method. It's just another feature. I found it very confusing to have an off-the-shelf exposition about Bayesian linear regression that refers to one of the coordinates in the feature vector as the 'prior', yet it receives no treatment as a prior probability distribution.

I found the introductory content about AI co-scientists too removed from the content of the paper. The paper offers a precise black-box optimization approach, and seems only tangentially related to co-scientists. The main connection is that they use scientific background knowledge from language models, but lots of modern work uses language models!

Questions for authors

Please address my critique above about the novelty of the approach, the use of 'prior', etc.

The results in Fig 3 definitely need error bars across multiple replicates. It's hard to know if any difference is significant.

Please comment more on the baseline approaches (piBo, etc). Are these doing something more formal where pi(x) is actually treated as a prior distribution? How does that work for the particular applications you consider, where there was no argument otherwise that pi(x) is a proper probability distribution.

Please comment on the relationship between your paper and Hsu 2022 "Learning protein fitness models from evolutionary and assay-labeled data".

5 comments
AuthorsOpenReview ↗

Response to Reviewer nJEx (1/4)

We thank Reviewer nJEx for the review and for recognizing that the paper "addresses a practical question: how do we merge traditional optimization approaches such as Bayesian Optimization with the prior knowledge available in foundation models.". We address each point below, with new experiments where they help.


[W1. The method is simple]

We agree that the estimator is simple. This simplicity is intentional, and we view it as a practical strength. HARIBO requires neither fine-tuning nor access to model weights or logits, making it easy to integrate with AI co-scientists that provide candidate scores or nominations, including proprietary chat-only models. Its signed coefficient provides an interpretable estimate of the prior’s predictive contribution beyond the candidate features, updated as experimental evidence accumulates. As Reviewer dvfR notes, the method “is simple, but experiments show that prior knowledge from LLMs is being used effectively.” We also acknowledge that the model class is not new: ridge regression on candidate features augmented with an effectively unpenalized zero-shot score is the “augmented” model of Hsu et al. (2022), which we now explicitly cite and discuss (Q3).

Our contribution is what this estimator does when it drives sequential experiment selection under a prior of unknown quality and sign:

  • Characterization (Prop. 1). The coefficient converges to the prior's contribution beyond the features, γ⋆\gamma^\star. Any fixed weight λ\lambda incurs excess risk (λ−γ⋆)2 Var(π~)(\lambda-\gamma^\star)^2\,\mathrm{Var}(\tilde\pi), which is largest exactly where the prior is wrong.
  • Identification (Eq. 3). Round one has no labels. We show that the first batch must contain both prior-chosen and other candidates for wπw_\pi to be estimable, and derive the even split. This question does not arise when fitting once on a fixed labelled set.
  • Safety in practice (Sec. 4.2). Under wrong or inverted priors, every method that fixes the prior's weight in advance falls below BO, and HARIBO does not.

New experiments for the other reviewers extend this to mixed priors that are right in some regions and wrong in others, to realistic budgets of 5–10% of the pool, and to uncompressed features. In every new setting HARIBO stays at or above BO, while the fixed weights do not.

In practice, HARIBO is a closed-form fit that runs in milliseconds, needs only scalar scores from a chat-only model queried once, and exposes a signed coefficient that a scientist can read each round.

Updated in Sec. 2 ("Score-augmented regression and LLM-in-the-loop BO"), Sec. 3 ("Interpreting wπw_\pi") and the appendix ("Design and scope", "Relationship to score-augmented regression"), blue.


[W2. The notion of "prior"]

We thank the reviewer and agree that we used the term loosely. We will fix it.

What π(x)\pi(x) is. π\pi is not a distribution over candidates. It does not sum to one, and it is not the probability that xx is the optimum or a hit. It is a scalar score from an external knowledge source that is believed to increase with yy:

  • a nomination indicator (gene);
  • the ESM2 log-likelihood ratio (protein);
  • the mean per-residue log-likelihood (antibody);
  • an LLM's 0–10 rating.

We assume only that this belief may be right, irrelevant, or wrong in sign. In the revision we call π\pi a prior score.

In what sense it is a prior. π\pi is a prior over the function ff, not over xx. Eq. 1 is exactly the posterior mean of the Bayesian linear model

\qquad w_\phi\sim\mathcal{N}\left(0,\frac{\sigma^2}{\alpha}I\right), \qquad p(b,w_\pi)\propto 1.$$ Leaving $\pi$ unpenalized *is* the flat prior on $w_\pi$. Equivalently, $\pi$ is an explicit basis function of the prior mean of $f$ (Rasmussen & Williams, 2006, §2.7). $\pi$ states the shape $f$ is expected to have, and the data decide how strongly, and in which direction, to believe it. $\pi$BO, by contrast, treats $\pi$ as a density over the location of the optimum.
AuthorsOpenReview ↗

Response to Reviewer nJEx (2/4)

[W2 (cont.). What happens if π\pi is treated as a formal prior]

We tested this directly on protein (184 assays, 20 seeds, the paper's loop); results are in Table R1 below.

  1. Normalizing π\pi into a distribution changes nothing for HARIBO. HARIBO is invariant to affine re-encodings of π\pi. Feeding it log⁡p(x)\log p(x) with p(x)∝exp⁡(π(x))p(x) \propto \exp(\pi(x)) reproduces HARIBO exactly. On the gene screens, every distribution built from the nomination set (uniform, rank-based, softmax) does too.
  2. Using π\pi as a proper distribution, as π\piBO does, is not robust. We converted π\pi into p(x)p(x) in three ways (uniform over the top 10%, Zipf over the ranking, softmax). We then ran π\piBO as published: the initial batch sampled from pp, then acquisition α(x) p(x)β/n\alpha(x)\,p(x)^{\beta/n} with β∈{1,10}\beta \in \{1, 10\} and two base acquisitions, for 12 variants in total. Under the sign-flipped prior, all 12 fall below no-prior BO, and the variants that do best on the good prior collapse hardest.
  3. A proper prior on the trust itself trades one regime for the other. We replaced the flat prior on wπw_\pi with N(μ0,τ2)\mathcal{N}(\mu_0, \tau^2):
    • a credulous prior (μ0>0\mu_0 > 0) loses 0.23 under the adversarial prior;
    • a sceptical prior (μ0=0\mu_0 = 0) loses 0.07 even under the good prior.

HARIBO achieves higher recall than the tested sceptical Gaussian prior in both regimes, while avoiding the large adversarial degradation of the credulous prior.

Table R1. Protein (n=184n = 184), recall@R10, 20 seeds.

how π\pi is usedgood (+ESM2)adversarial (−ESM2)
pure BO (no π\pi)0.5600.560
π\piBO, π\pi as a proper distribution (12 variants)0.611 – 0.7310.384 – 0.516
Gaussian prior on wπw_\pi, credulous (μ0=+0.3\mu_0 = +0.3, τ=0.03\tau=0.03)0.7330.470
Gaussian prior on wπw_\pi, sceptical (μ0=0\mu_0 = 0, τ=0.03\tau=0.03)0.6580.607
HARIBO fed log⁡p(x)\log p(x), p∝exp⁡(π)p \propto \exp(\pi)0.725 (identical)0.705 (identical)
HARIBO (flat prior on wπw_\pi)0.7250.705

We will rename π\pi to "prior score", add the Bayesian reading to §3 and Table R1 to the appendix.


[W3. Connection to AI co-scientists]

The co-scientist is the object of study. HARIBO addresses the question every scientist faces when an AI co-scientist proposes candidates: how far should I follow it? Reviewer dvfR describes the failure mode we start from as "an interesting failure mode of LLM based co-scientists, which is of interest in many domains". Reviewer 575U notes that LLMs underperform on niche problems, "which is exactly the setting of new discovery pipelines". Our experiments compare against co-scientists run as closed-loop agents. We strengthened this connection in three ways.

  1. wπw_\pi as a diagnostic. We now present the learned wπw_\pi as a report for the scientist. Its sign and uncertainty say, round by round, whether the co-scientist's suggestions are helping, uninformative or misleading. Fitted on two priors at once, the weights track each prior's actual quality (Spearman +0.86 for ESM2, +0.72 for the agent; §4.1).
  2. The co-scientist baselines, specified fully. These are a frozen prior and BioDiscoveryAgent re-queried with experimental feedback. We now state what each one sees and may do.
  3. New experiments with other co-scientist architectures.
    • Multi-agent debate/critic. A PI agent proposes genes, an immunologist and a molecular geneticist critique and revise the list, and the votes are aggregated (Claude Opus 5, 3 calls, 2 repeats, 6 screens). On its own this prior barely beats BO (0.113 vs. 0.111). Fed to HARIBO it reaches 0.144, above BO on all 6 screens (p=0.031p=0.031), about the same as a single BDA-style call (0.148). The extra critic calls buy nothing that the learned weight does not already provide.
    • BioDiscoveryAgent's own critic and tools, re-queried every round (Qwen2.5-32B; 3 screens; AI critic, literature search, gene search). The re-queried agent reaches 0.077 with the critic only, 0.095 with all tools and 0.098 with none. One-shot HARIBO reaches 0.128, and BO 0.118.
    • Prior revision. The co-scientist revises its prior after one measured batch, with or without tools. This improves prior quality in three of four domains, but HARIBO's ten-round recall moves by at most 0.02 either way (gene 0.131 → 0.129, protein 0.595 → 0.598, antibody 0.615 → 0.633, LCBench 0.622 → 0.619 without tools; tools do not add to this).

The practical message: one query plus a learned weight gives most of the benefit at 20–32× fewer requests, and protects the budget when the co-scientist is wrong, which it cannot tell us itself (Fig. 1, Table 1).

Updated in Sec. 3 ("Interpreting wπw_\pi"), Sec. 4 ("Co-scientist-only arms") and the appendix ("Re-queried agent: what it sees from round two", "Prior revision with data and tools", "The re-queried agent with tools"), blue.

AuthorsOpenReview ↗

Response to Reviewer nJEx (3/4)

[Q1. Error bars for Fig. 3]

Thank you for noting this. Every curve in Fig. 3 is already a mean over 20 seeds per dataset, but the figure did not show the spread. We have redrawn it with ±1 s.e. bands over datasets (the replication unit throughout the paper) and now report paired Wilcoxon tests per round. The values, with Qwen2.5-32B as both the agent and the source of the prior:

Table R2. Cumulative recall, mean ± s.e. over datasets; pp is the paired Wilcoxon test of HARIBO against the agent.

domainroundpure BOagent (re-queried every round)HARIBO (1 query)pp
gene (n=6n=6)10.008 ± 0.0000.024 ± 0.0070.019 ± 0.0060.031
50.053 ± 0.0040.066 ± 0.0090.072 ± 0.0110.094
100.111 ± 0.0090.109 ± 0.0130.131 ± 0.0160.031
protein (n=212n=212)10.048 ± 0.0010.049 ± 0.0040.051 ± 0.0030.32
50.286 ± 0.0050.249 ± 0.0080.301 ± 0.0076×10−66\times10^{-6}
100.554 ± 0.0070.505 ± 0.0080.590 ± 0.0093×10−113\times10^{-11}
antibody (n=36n=36)10.051 ± 0.0020.084 ± 0.0370.051 ± 0.0070.18
50.365 ± 0.0360.281 ± 0.0410.336 ± 0.0360.20
100.604 ± 0.0370.464 ± 0.0460.599 ± 0.0360.036

The differences are well outside the error bars. By round 10, one query plus HARIBO is significantly ahead of the agent that is re-queried every round, in all three domains. On antibody, where the prior carries no signal (corr ≈ 0), HARIBO tracks BO (0.599 vs. 0.604), while the agent falls 0.14 below it. On protein, HARIBO is also ahead of BO (p=2×10−6p=2\times10^{-6}).

On gene, six screens cap the attainable significance at p=0.031p=0.031, the smallest value a paired Wilcoxon test can reach with six datasets. We therefore also report replicate-level error bars within each screen. Over 20 seeds, HARIBO's round-10 gain over BO exceeds two standard errors on 4 of 6 screens with Qwen2.5-32B and on 5 of 6 with Claude Sonnet 5. The remaining screens are ties (|Δ| ≤ 0.003), never losses.

Fig. 3 and Fig. 8 redrawn with s.e. bands (blue).


[Q2. Do the baselines treat π\pi as a proper prior?].

In their original papers, π\piBO and BOPrO do, but their prior is over something different from ours:

  • π\piBO (Hvarfner et al., 2022) takes a user-specified density over the location of the optimum and multiplies the acquisition by π(x)β/n\pi(x)^{\beta/n}, so the prior's influence decays with the iteration count nn.
  • BOPrO (Souza et al., 2021) combines the same kind of density with a model-based density, and the model's share grows with the data.
  • LGBO (Yuan et al., 2026) does not use a density. It shifts the surrogate by the LLM's preference, scaled by the model's self-reported confidence.

In our applications no such density exists: π\pi is a nomination indicator, a log-likelihood ratio or a 0–10 rating (W2). Any density-based method must therefore first turn π\pi into a distribution.

Reviewer_dvfROpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LLM based co-scientists are being used for scientific discovery. The authors note a drawback of using these co-scientists - mainly in areas where the literature does not have a lot of information the LLMs either hallucinate or fail to provide the right level of uncertainty. The authors motivate this with simple examples where LLMs hallucinate as coverage thins and their confidence scores do not separate real from fabricated genes. The authors propose instead of using the LLM to construct prior features that act as an input dimension in a BO surrogate model. These prior features can be candidates nominated by the LLM or scores reflection a desired property. These prior features are combined with data features to learn a ridge regression model on the output, the top K unmeasured candidates with the highest predicted output are selected for the next round. The authors justify this by showing theoretically that in regression the weight will be 0 if the LLM prior feature is not informative and negative if it is not helpful (using standard facts about regression). The experiments, carried out in 3 distinct closed loop discovery settings, show that the approach is able to use information from the LLM features if is correlated (or anti-correlated), and is able to ignore it if it is noisy.

Critical strengths

The authors identify an interesting failure mode of LLM based co-scientists, which is of interest in many domains (although underexplored). The proposed method is simple, but experiments show that prior knowledge from LLMs is being used effectively. The experiments themselves are convincing and comprehensive (at least against the pure BO side).

Critical weaknesses

The failure mode of the LLM based co-scientists is not explained or explored rigorously. The advantages over pure Bayes opt are quite clear in the paper, but the advantages against pure LLM co-scientists are not. For this to be clear, more details are needed on the agent arms run in the experiment.

Questions for authors
  • How do you ensure there is no dataset leakage to the LLMs in your studies? The prior knowledge could be useful because the LLMs have seen the datasets at some stage in training, and thus the outcomes are known.
  • Did you check if when provided data and tools, the LLM agents could not analyse the data to 'fix' their own prior?
  • Its not clear in the experiments what the co-scientist only methods are doing (agent arm in Table 3 and figure 4). Are they given the whole dataset and asked to provide a candidate? Are they allowed to use tools? If they are allowed to use Bayesian optimisation code, can they not already combine their own prior information with the outcome of the Bayesian optimisation? Could you be more clear in what the agents are provided with and what they are allowed to do?
4 comments
AuthorsOpenReview ↗

Response to Reviewer dvfR (1/2)

We thank Reviewer dvfR for the positive assessment, describing the failure mode as “interesting” and “of interest in many domains,” noting that the method is “simple, but experiments show that prior knowledge from LLMs is being used effectively,” and describing the experiments as “convincing and comprehensive.” We address each point below.


[W1. A more rigorous account of the co-scientist failure mode]
The failure mode of LLM co-scientists is indeed one of our key motivations for developing HARIBO, which is designed to remain robust in such settings. Reviewer 575U also found this motivation (Fig. 1, Table 10) “valuable.” We have expanded our analysis into a dedicated appendix presenting three empirical findings:

  • (i) Fabrication is driven by the batch demand, conditional on coverage (Table 10). When Qwen2.5-32B is asked only for genes it is confident about, it returns 7–15 symbols and invents almost none (0–6%). When it is asked for the 128 genes a closed-loop batch needs, it invents 56–66% on four of six screens. The mechanism is padding: it continues real gene families into names that do not exist (PPP2R5P, DCTNIP1, and PTPN314…PTPN373 when asked for 1,280 names).
  • (ii) The model cannot flag its own fabrications (Table 1). On the niche screen it rates 28% of the fake genes at 5/10 or higher. This is why weighting the prior by self-reported confidence is unsafe.
  • (iii) Scale removes fabrication, but not the dependence on coverage. Claude Sonnet 5 fabricates 0% at the same demand, yet how useful its list is still depends on how well studied the phenotype is:

Table R1. Hit enrichment of the round-one nominations over the panel base rate.

screenQwen2.5-7BQwen2.5-32BClaude Sonnet 5Claude Opus 5
IL2 (well-studied)5.5×4.4×10.2×11.8×
IFNG (well-studied)3.7×1.9×7.2×6.7×
Carnevale22 (adenosine)1.3×2.1×1.8×1.6×
Sanchez21 (tau, up)0.4×1.4×1.9×2.3×
Sanchez21 down (tau)1.3×2.2×1.1×3.8×

Which regime a new screen falls into cannot be known before measuring it. That is exactly the situation HARIBO is built for: the learned weight rises on well-covered screens and falls toward zero on the others (Fig. 7).


[W2. Advantage over pure LLM co-scientists]

The main advantage is cost, with no loss in recall in most settings. HARIBO queries the LLM once (5.9k tokens per gene screen). The closed-loop agent is re-queried every round, which takes 20–32× more requests and 124–234× more tokens (Table 5).

Table R2. Recall@R10, one query plus HARIBO vs. the agent re-queried for ten rounds.

domainQwen2.5-32B: HARIBO / agentClaude Sonnet 5: HARIBO / agent
gene0.131 / 0.109 (p=0.031p=0.031)0.139 / 0.169
protein0.590 / 0.505 (p=3×10−11p=3\times10^{-11})0.637 / 0.657 (p=0.14p=0.14)
antibody0.599 / 0.464 (p=0.036p=0.036)0.574 / 0.449 (p=0.003p=0.003)

The pattern follows the failure mode above.

  • Strong prior knowledge (AI > BO): HARIBO benefits from it.
  • Weak prior knowledge (BO > AI): HARIBO stays near BO. The agent, which has no mechanism to discount itself, ends below knowledge-free BO on protein and antibody (Qwen) and on antibody (Sonnet).
  • Where a frontier agent leads (gene): its gain is concentrated on the best-covered screens, +0.013 to +0.084 on the three most enriched and −0.007 to +0.003 on the rest (App. C.6). A wider single call recovers most of it (0.140 → 0.163; App. C.8).
AuthorsOpenReview ↗

Response to Reviewer dvfR (2/2)

[W3 & Q3. What the agent arms are given and allowed to do]

This was under-specified, and we now describe it in a dedicated appendix. There are two arms.

  • Prior arm ("agent arm" in Table 3): the model is queried once, and each round we measure its top-scored untested candidates.
  • Re-queried agent (Fig. 3, Fig. 8, §4.4): BioDiscoveryAgent's released loop, with its prompt unchanged (App. B.1), ported to proteins and antibodies.
re-queried agent
sees in round 1the same prompt as our prior (task and pool size; or the candidate list)
sees from round 2measured score and hit status of every candidate it already proposed (protein/antibody: last 120)
never seesunmeasured outcomes, the full dataset, features ϕ\phi, our surrogate, or code
toolsnone in the main comparison; tested separately (below)

It is therefore a closed-loop competitor that uses the same labels as our surrogate, with a language model where we put a ridge regression.

Could an agent with BO code combine its prior with the data itself? In principle, yes. HARIBO is a principled version of that combination: one prior column in the BO surrogate, with its weight learned from labels. In practice, BioDiscoveryAgent's own tools used in every round (AI critic, literature search, gene search) bring the Qwen2.5-32B agent to only 0.095 on three screens (0.098 without tools), against 0.128 for one-shot HARIBO.


[Q1. Could the LLMs have seen the datasets (leakage)?]

Most screens predate the training cutoffs, so the models may have seen the publications. What matters is whether they behave as if they had memorized the outcomes, and they do not.

Table R3. Round-one nominations, pooled over six gene screens (~120 names per screen).

proposerhits among namesprecisionshare of all screen hits named
Qwen2.5-32B63 / 7130.091.9%
Claude Sonnet 5109 / 7280.153.2%
Claude Opus 5129 / 7200.183.8%

A model that had memorized the published hit tables could fill its ~120 slots almost entirely with hits. Instead, enrichment tracks how much the literature says about a phenotype (Table R1), not whether the screen was published. The protein priors carry no such channel either: ESM2 never sees assay labels, and the LLM protein prior correlates with fitness at only +0.08 to +0.31.

Leakage would also not affect the central claim. A leaked prior is simply a strong prior, which HARIBO exploits like any other. Our robustness results concern priors that are wrong by construction (sign-flipped, misapplied, partly corrupted), where no leakage helps, and that is where HARIBO stays above BO while fixed weights fall below it.


[Q2. Can the agent fix its own prior when given data and tools?]

We added this experiment. Qwen2.5-32B gives its one-shot prior, sees one measured batch (with or without tools), and returns a revised prior. Revision improves prior quality on protein, antibody and an AutoML benchmark (LCBench), but lowers it on genes (hit rate 0.088 → 0.056). Tools add no consistent further gain. Plugged into HARIBO, the revised prior moves ten-round recall by at most 0.02 (e.g. gene 0.131 → 0.129, protein 0.595 → 0.598), and it costs 113k–122k tokens against 5.9k for the one-shot query. One-shot elicitation is therefore the cheapest configuration, and none of the costlier ones is significantly better.

Added in "Re-queried agent: what it sees", "Leakage check" and "Prior revision with data and tools" (new Appendix E), blue.

We thank the reviewer again and hope these additions clarify the failure mode and the comparison with co-scientist agents.

Reviewer_dvfROpenReview ↗

Response

I have read the review response and the other responses to other reviewers.

I think that this is an obvious way of using prior knowledge from LLMs that seems underexplored (from what I see in response to Reveiwer an8o) and performs well.

I am happy to increase my score to 4 based on the responses by the authors and the additional details presented.

AuthorsOpenReview ↗

Thank you for your response

Thank you for your thoughtful engagement and for raising your score. We appreciate your recognition that this is a simple yet underexplored way to use LLM priors, and we will reflect the discussion in the final version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their time and constructive feedback. We were encouraged that the reviewers found:

  • An important, well-motivated problem. It is "a well motivated and important problem", and safely combining LLM priors with BO "is a promising direction for the field" (575U). The paper "addresses a practical question" (nJEx).
  • An interesting failure mode of AI co-scientists. This is "an interesting failure mode of LLM based co-scientists, which is of interest in many domains" (dvfR), and "the motivation in Fig. 1 and Table 10 is valuable" (575U).
  • A simple, cheap and practical method. It "requires no fine-tuning and works with a single query and a closed-form fit" and "can be easily plugged in with proprietary models" (575U). It "is simple, but experiments show that prior knowledge from LLMs is being used effectively" (dvfR).
  • Convincing experiments, honestly reported. The experiments are "convincing and comprehensive" (dvfR), the paper "is honest about what did not work" (575U), and its figures and tables are "visually polished and clearly presented" (an8o).

Below we summarize the shared concerns and what we did about them. All new experiments use the paper's loop, features, seeds and paired Wilcoxon tests over datasets. All revisions are marked in blue.

1. Novelty and positioning (nJEx, an8o, 575U).

  • We now cite and discuss Hsu et al. (2022). Their augmented ridge model is the same model class.
  • We also discuss LLINBO and LABO (ICML 2026).
  • Here, our contribution is the use of this estimator for sequential experiment selection under a prior of unknown quality and sign. It has three parts: the signed coefficient as the learned trust, Prop. 1's excess-risk characterization of any fixed weight, and the round-one design for sparse priors (Eq. 3).
  • Empirically, adapted LLINBO rules still fall below BO under a flipped prior. A LABO-style learned scale, the closest mechanism to ours, survives it; with LLM priors HARIBO is numerically higher in all four comparisons (significant in two) and needs no LLM queries inside the loop.

2. What "prior" means, and the baselines (nJEx, 575U).

  • π\pi is a prior score: external knowledge available before any measurement, not a probability distribution. In a Bayesian reading, the probability prior sits on its coefficient, and Eq. 1 is the posterior mean under a flat prior on wπw_\pi.
  • Treating π\pi as a formal prior does not rescue the baselines. π\piBO run as published, with π\pi turned into a density (12 variants), falls below BO under a flipped prior in every variant (0.384–0.516 vs. 0.560), while HARIBO holds 0.705.
  • We rename the arms "π\piBO-style", "BOPrO-style" and "LGBO-style", and correct two §4.2 sentences.

3. Robustness beyond a globally inverted prior (575U).

  • Mixed priors: wrong on a fraction ff of the protein. HARIBO stays at or above BO for every ff, while every fixed or decaying weight falls below BO at f≥0.75f \ge 0.75.
  • Naturally adversarial antibody assays and flipped gene lists: fixed weights fall below BO, while HARIBO is never significantly below it.

4. Evaluation setup (575U, an8o).

  • Realistic 5–10% budgets: HARIBO is about +70% over BO under a good prior and best under an adversarial one.
  • Stronger features: uncompressed one-hot (Hsu et al.) on protein, and 2,120-dimensional raw features on gene. The gains persist, and on gene they grow.
  • Very few labels: the sign of wπw_\pi is correct on 71–81% of runs after just 4 labels.
  • AutoML (LCBench): HARIBO beats BO at both budgets (p<10−5p < 10^{-5}).

5. Comparison with co-scientist agents (dvfR, nJEx, 575U).

  • The agent arms are now specified fully: what the agent sees, what it may do, and its prompts.
  • Fig. 8 had a plotting error and §4.4 an attribution error, both fixed. The figure's protein and antibody panels used a Qwen prior for HARIBO, and the 0.703 in §4.4 used ESM2. At a matched prior the frontier agent is numerically above HARIBO on protein (0.657 vs. 0.637, p=0.14p=0.14), both above BO. We revise the abstract accordingly: re-querying helps on gene, is not significantly different on protein, and falls below BO on antibody.
  • Other co-scientist architectures were added: a multi-agent debate/critic proposer, BioDiscoveryAgent's own critic and tools re-queried every round, prior revision after measured feedback, and re-querying whenever progress stalls. None significantly beats one query plus HARIBO, and all cost more.
  • A leakage check shows the models name only 1.6–3.8% of each screen's hits.

6. Presentation (nJEx, 575U). Fig. 3 and Fig. 8 now carry ±1 s.e. bands and per-round tests. The "uninformative" prior is defined in the captions. The limitations now state that one global coefficient does not measure local reliability.

We hope the responses and the revised manuscript address the reviewers' concerns, and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.