acceptodds
Under review as a conference paper at ICLR 2027

Search or Experiment? Budget-Aware Metareasoning for LLM-Driven Scientific Discovery

Abstract

Large language models (LLMs) are increasingly used to automate scientific discovery. Yet, existing approaches typically optimize hypothesis search over fixed evidence or active experimentation over newly acquired evidence in isolation. These operations have sharply asymmetric costs, since hypothesis refinement consumes computation, whereas experiments require substantial resources. Moreover, the marginal value of each action changes throughout discovery and must be inferred from the evolving discovery state. The central challenge is therefore to continually determine whether the next unit of budget is better spent refining what is already known or acquiring evidence that can change what is known. We introduce CoREAgent (Cost-optimized Refinement and Experimentation Agent), a budget-aware framework that casts scientific discovery as sequential allocation between SEARCH and EXPERIMENT. At each step, a learned metareasoner estimates the cost-adjusted downstream value of each action from the discovery state and chooses between refining hypotheses and acquiring evidence, while a separate acquisition policy decides where to experiment. CoREAgent learns these values online through prequential credit assignment, scoring each decision on evidence acquired after it was made. Across four interactive discovery benchmarks, CoREAgent improves exact accuracy by 8 − 19 percent over the strongest baseline while reducing experimental usage by up to 31%. Our analyses isolate the contribution of each component and show that efficient discovery requires optimizing not only what to search for or experiment on, but when each action is worth its cost.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Large Language Models, Scientific Discovery, Budget-aware Decision Making, Active Experimental Design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5uH7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for balancing Search, which consumes computational resources, and Experiment, which consumes experimental resources, during scientific discovery. It evaluates the framework on NewtonBench and other benchmarks.

Critical strengths
  1. The figures are well designed.
  2. The research question is important to building automated laboratories.
Critical weaknesses

Limited novelty: The method relies on common components, including a meta-controller. Despite the name “CoREAgent,” the paper does not demonstrate clear algorithmic advances in agent design. Its limited action space and lack of explicit task planning also raise questions about whether it qualifies as an agent.

Questions for authors
  1. Could the authors compare against LLM-assisted Bayesian optimization methods such as LLAMBO, LGBO, LLINBO, and LABO?
  2. Could the evaluation benchmarks have appeared in the LLM’s training data? Potential contamination makes it difficult to assess the LLM’s contribution to scientific discovery.
  3. Could the authors provide a systematic component-wise ablation of the full framework to justify its design?
  4. Could the authors conduct more evaluations in high-dimensional settings, which are common in scientific applications?
  5. Could the authors better explain the motivation behind the methodological choices?
  6. Could the authors provide theoretical analysis establishing bounds or guarantees for the method?
  7. Could the authors evaluate different LLMs, prompt sensitivity, and the effect of reasoning capabilities? GPT-4o-mini is a relatively old and less capable model. With more advanced but expensive models, how would API costs affect resource allocation and overall performance?
9 comments
AuthorsOpenReview ↗

Response to Reviewer 5uH7 (1/3)

Thank you for dedicating your time and expertise to review our submission. Please find our responses below.

Limited novelty

We thank the reviewer for this concern. Our novelty lies not in introducing a meta-controller, but in what the controller learns to decide and how it learns from experimental evidence.

ParadigmHypothesis searchExperiment designLearned meta-controllerCost-aware search vs. acquisition
Static discovery✓✗✗✗
Active discovery✓✓✗✗
Budget-aware LLM routing✗✗✓✗
CoREAgent✓✓✓✓
  • vs. Active Discovery: These methods refine hypotheses and select informative experiments, largely neglecting acquisition costs. However, these methods do not learn when further hypothesis refinement is more valuable than acquiring new evidence; their search–experiment allocation follows fixed schedules or rules.

  • vs. Budget-Aware Methods: These methods allocate computational resources or select models for a given task. CoREAgent addresses a distinct challenge: experiments generate new evidence that can invalidate existing hypotheses and change the value of subsequent search. Thus, the allocation decision depends not only on computation, but also on how acquiring new evidence reshapes the discovery process. CoREAgent formalizes this interdependent decision as a sequential allocation problem (Eq. 4) and introduces online prequential credit assignment (Eq. 10), learning the value of both actions from predictive improvements on subsequently acquired observations.

  • CoREAgent builds on discovery agents such as MDA and DP-Agent, where planning takes the form of hypothesis generation and experiment design, and adds a learned decision on top about when to search and when to experiment.

Empirical Evidence. With the shared hypothesis-search engine (Sec. 5.2, Table 2), CoREAgent achieves comparable recovery on NewtonBench while using only 40.3% of the search and 69.5% of the experimental budget. On ODEBench, it improves recovery from 80.7% to 84.4% with fewer resources. Under matched search, acquisition, and stopping (App. C.4, Table 8), it additionally achieves comparable recovery with 21.6% fewer LLM calls than a fixed schedule.

Thus, our central contribution is learning how to orchestrate hypothesis refinement and evidence acquisition as scientific understanding evolves, rather than introducing novelty in every underlying component or claiming a new general-purpose planning architecture.

Comparison with LLM-assisted Bayesian optimization methods

We thank the reviewer for highlighting these relevant works. While they share our motivation, their objective is optimizing a scalar function, whereas CoREAgent aims to identify the governing scientific mechanism. This distinction changes both the role of the LLM and the decisions being optimized.

  • LLAMBO uses LLMs for warm-start initialization, surrogate modeling, and candidate sampling.

  • LGBO incorporates LLM-generated preferences into the Bayesian surrogate to guide optimization.

  • LLINBO combines LLM-proposed candidates with statistical surrogate guidance, determining when to trust LLM suggestions versus conventional BO.

  • LABO adaptively selects between inexpensive LLM-based evaluations and costly experiments using an evolving uncertainty-based gating criterion.

The key distinction from LABO is the decision being learned. LABO selects between evaluation fidelities to improve objective optimization. CoREAgent instead decides between refining competing scientific hypotheses using existing evidence and acquiring new observations to distinguish them. Hypothesis refinement does not itself produce new experimental observations; its value is assessed through subsequent evidence using online prequential credit assignment (Eq. 10).

Our evaluation also includes BED, LLM+BED, and MDA, which directly target Bayesian experiment selection and law recovery (Sec. 4; Sec. 5.1, Fig. 2). We have now added LLM-assisted Bayesian optimization and its distinction from mechanism discovery in Sec. 6.

Component-wise ablation

Fig. 4 (Sec. 5.4) ablates each component, now over three seeds. CoREAgent outperforms every ablation (using the fewest experiments (67.1%) and search calls (49.7%)): removing the metareasoner, guided acquisition, or feedback drops accuracy from 88.9% to 82.2%, and removing diversity drops it to 80.0%.

We added controls to understand the benefits of the learned allocation (see response to reviewer yjqQ and WgsQ):

  • With the same search engine and stopping rule, CoREAgent reaches comparable recovery with 21.6% fewer LLM calls (Table 8).
  • Against simple switching rules, CoREAgent achieves the highest recovery and uses 32.6% fewer LLM calls than the next best (Tables 9–10).

These results show that every component is therefore necessary, and together they achieve the highest accuracy at the lowest cost.

AuthorsOpenReview ↗

Response to Reviewer 5uH7 (2/3)

Potential benchmark contamination makes it difficult to assess the LLM’s contribution.

We agree that contamination is important and cannot rule out pretraining exposure. However, our benchmarks incorporate safeguards that reduce opportunities for direct recall.

  • Benchmark design. NewtonBench employs counterfactual physical laws; ForceBench requires inferring unknown interaction laws and experimental control roles; ActiveSciBench-Chem conceals the relevant variables, functional forms, and parameters; and ODEBench removes semantic descriptions. Every task begins with a single seed experiment and requires further evidence acquisition. These safeguards mitigate direct recall without guaranteeing contamination-free evaluation.

  • Empirical evidence. All methods use the same backbone and task descriptions. If recall drove performance, the LLM-only baseline would already solve these tasks; instead, it recovers only 63.0% on NewtonBench and 36.7% on ODEBench. Under identical budgets, CoREAgent raises this to 88.0% and 84.4% (+25.0 and +47.7 points; Sec. 5.2, Table 1), indicating that the gains come from the framework built around the LLM, which generates, tests, and refines hypotheses using experimental evidence.

We have included this clarification in Appendix A.1 (Datasets), acknowledging potential pretraining exposure and explaining the scope of our evaluation.

Evaluations in high-dimensional settings

We agree that high-dimensional evaluation is important and an active area of research. However, realistic, high-dimensional benchmarks supporting interactive scientific discovery and mechanism recovery remain limited. We thus evaluate on four established benchmarks spanning diverse scientific domains, including seven-input enzyme kinetics (ActiveSciBench-Chem) and multidimensional dynamical systems (ODEBench) and physics laws. These evaluations establish performance across heterogeneous experimental interfaces. We acknowledge this limitation in Sec. 7.

Importantly, CoREAgent is not inherently restricted to low-dimensional problems, its metareasoner operates on discovery-state features rather than a fixed-dimensional experimental input, and its hypothesis-search and acquisition modules can accommodate different experimental interfaces. Nevertheless, scalability of these underlying modules in substantially higher dimensions remains to be established. We acknowledge this and identify broader high-dimensional evaluation as future work.

Motivation behind the methodological choices

CoREAgent addresses a central challenge in scientific discovery: further hypothesis refinement cannot resolve ambiguities caused by insufficient evidence, while premature experimentation wastes scarce experimental resources. This motivates three complementary components.

  • Diverse hypothesis search (Sec. 3.1): Exploration–exploitation populations, LLM-guided proposals, symbolic fitting, and cross-population exchange maintain structurally distinct candidate mechanisms, preventing premature convergence.

  • Hypothesis-guided acquisition (Sec. 3.2): Belief-weighted information gain targets experiments that discriminate among competing mechanisms rather than merely reducing predictive uncertainty.

  • Online metareasoning (Sec. 3.3): Bayesian linear value models and Thompson sampling support online allocation decisions. Prequential credit assignment evaluates earlier decisions using subsequently acquired evidence, providing a measurable signal for learning when additional refinement or experimentation is more useful.

Together, these components enable the framework to learn not only which hypotheses to investigate and where to experiment, but when to allocate resources between the two. The ablations (Fig. 4) support their complementary roles, while the cost-sensitivity study (Fig. 3) demonstrates adaptive allocation behavior.

Evaluate different LLMs, prompt sensitivity, and the effect of reasoning capabilities. How would API costs affect resource allocation and overall performance?

We thank the reviewer for these suggestions.

Different LLMs. We evaluated three backbones (GPT-4o-mini, Qwen3-32B, and Claude Haiku 4.5) in App. C.3, Table 7. Recovery varies with the backbone, but experiment usage stays similar, indicating that the metareasoner's allocation behavior is stable across models.

API costs and allocation. We vary the search cost up to 30× and the experiment cost up to 40× (Sec. 5.3, Fig. 3). CoREAgent shifts from whichever action becomes more expensive while maintaining recovery, compared with 78% for LLM-AutoSciLab (see our response to WgsQ). Learned allocation also directly reduces LLM usage: with the same search engine and stopping rule, it reaches comparable recovery with 21.6% fewer LLM calls (Table 8; see our response to yjqQ).

Prompt sensitivity. We could not complete the experiments on these within the rebuttal period and will include them in the camera-ready version.

AuthorsOpenReview ↗

Response to Reviewer 5uH7 (3/3)

Theoretical analysis establishing bounds or guarantees for the method

We agree that theoretical guarantees would strengthen the work. Our contribution is a sequential decision formulation (Eq. 4) with an online learning procedure, not a regret or optimality guarantee; only budget feasibility is guaranteed, by construction. Guarantees are hard to derive because search changes the set of hypotheses, experiments change the evidence, and feedback arrives late. For the same reason, the prequential credit (Eq. 10) is a practical, short-horizon proxy for an action's value, not an unbiased estimate. We therefore support the method empirically; allocation responds to changing costs (Fig. 3), and learned allocation improves on the ablations, matched controls, and simpler switching rules (Fig. 4, Tables 8–10). Regret bounds for this setting remain an open direction.

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

The paper argues that LLM-driven discovery methods optimize either hypothesis search over fixed data or experiment selection, but never decide between the two. These actions have asymmetric costs: search consumes computation, while experiments consume scarce resources. The authors cast discovery as sequential allocation between SEARCH and EXPERIMENT under separate budgets. They propose CoREAgent, in which a metareasoner estimates the cost-adjusted value of each action from the discovery state and learns these values online from evidence acquired after each decision.

The contributions are as follows:

  1. A budget-aware framework that casts scientific discovery as sequential allocation between search (refining hypotheses on existing evidence) and experimentation (acquiring new evidence), under separate, asymmetric budgets.
  2. CoREAgent, which combines three components: (a) two-population LLM + PySR hypothesis search; (b) information-gain experiment selection over a BIC-weighted belief; (c) a Bayesian linear metareasoner trained online with prequential credit, where each decision is scored on the log-loss improvement on the next experiment’s observation, which it has not seen.
  3. Experiments showing higher accuracy with fewer experiments across four benchmarks, and allocation that shifts with relative cost.

Validation:

  1. Main comparison (Figure 2). On NewtonBench, ActiveSciBench-Chem, ForceBench and ODEBench with gpt-4o-mini, CoREAgent reports 88/58/83/84% exact accuracy, versus 78/39/67/76% for the strongest baseline. It uses 13.9, 19.8, 15.8 and 7.8 experiments on average.
  2. Equal budgets (Table 1). On NewtonBench and ODEBench, CoREAgent uses 40.3-43.3% of the search budget and 69.5-78.7% of the experiment budget, while LLM-only and LLM+BED exhaust both.
  3. Cost sensitivity (Figure 3). On 15 NewtonBench tasks, raising search cost 30× lowers searches per experiment by 33%. Raising experiment cost raises searches per experiment by up to 49%. Accuracy is roughly preserved in both cases.
Critical strengths
  1. The paper makes an under-explored trade-off explicit: when an experiment is worth more than further search. Framing this as value of computation versus value of information under separate budgets connects LLM discovery to classical metareasoning and Bayesian design. The decision will matter more as these agents move to settings where experiments are expensive. This matters to my assessment because the framing is useful even if the specific metareasoner turns out not to be what drives the gains.
  2. The credit assignment is principled. Each decision is scored only on observations it has not seen, following prequential evaluation. Search is therefore not rewarded for fitting data it already used, and an experiment is not rewarded for predicting its own observation. The authors acknowledge that this credit is a myopic, biased proxy, and they document the full set of design rules and constants in Appendix B. This matters because it gives a workable learning signal for a value that cannot be observed directly.
Critical weaknesses
  1. The experiments do not fully isolate the contribution of learned allocation (Table 1, Figure 4, Appendix B).
    • CoREAgent ends runs with a rule-based convergence test (Appendix B.3), while LLM-only, LLM+BED and the w/o Metareasoner ablation run until both budgets are exhausted (Sections 5.2, 5.4). The reduced resource usage may therefore come partly from early stopping rather than allocation.
    • As described in Appendix A.2, LLM-only and LLM+BED use single-population LLM proposals with constant fitting, with no mention of the two-population search or PySR seeding used by CoREAgent (Section 3.1). If so, Table 1 cannot separate allocation from a stronger search engine.
    • The introduction argues that naive switching policies are inadequate (Section 1), but no state-dependent switching heuristic is compared. The baselines and the w/o Metareasoner ablation use fixed schedules.
  2. The statistical support appears thin. I could not find the number of runs, seeds, or confidence intervals reported, although each run involves several sources of randomness (LLM sampling at temperatures 0.7–1.0, Thompson-sampled action selection, randomly drawn candidate conditions, PySR). Several key differences correspond to one or two tasks: ForceBench has 6 tasks (Table 2), so 83% vs 67% is a single task, and the ablations and cost sweeps use 15 NewtonBench tasks (Sections 5.3–5.4), so 80.0% vs 66.7% in Figure 4 is 12/15 vs 10/15. If results are averaged over multiple runs, reporting the spread would address this concern.
Questions for authors
  1. Matched stopping and heuristic controls. Can you add two controls?
    • A fixed search–experiment schedule with the same search engine and the same convergence stopping rule.
    • A state-dependent heuristic, for example experimenting after k search steps without improvement in the best score.
  2. Seeds. Can you report at least three seeds per method with confidence intervals, paired across tasks, at least for NewtonBench and the Figure 4 ablations?
  3. Metareasoner behaviour. Per Appendix B.3, every run begins with 20 Search steps, the first three experiments are fixed by rule, and Search blocks are capped at 20 steps. What fraction of decisions are made by Thompson sampling rather than forced by these rules? How does CoREAgent compare with the same system whose value model is frozen at its prior, so that actions are sampled without learning?
5 comments
AuthorsOpenReview ↗

Response to Reviewer yjqQ (1/2)

Thank you for dedicating your time and expertise to review our submission. Please find our responses below.

CoREAgent uses rule-based convergence stopping, whereas LLM-only, LLM+BED, and w/o Metareasoner exhaust both budgets. Thus, resource savings may partly reflect early stopping rather than adaptive allocation.

We agree that the original comparisons combined differences in hypothesis search, adaptive allocation, and early stopping. We therefore matched the control with the same hypothesis-search engine, acquisition policy, budgets, and convergence criterion, differing only in its fixed SEARCH–EXPERIMENT schedule (Appendix C.4, Table 8).

MethodAcc.LLM callsExps.
LLM-only + stop55.678.080.5
LLM+BED + stop62.275.377.0
w/o Metareasoner + stop84.471.368.0
CoREAgent88.949.767.1

Compared to the fixed SEARCH–EXPERIMENT schedule (w/o Metareasoner), CoREAgent recovers 89% laws versus 84% with fewer LLM calls and comparable experiments. These results demonstrate efficiency gains from learned allocation beyond stronger search and early stopping.

We also acknowledge that improved search contributes significantly to the original accuracy gains: search-matched fixed schedules attain 88.9–89.8% recovery versus 88.0% for CoREAgent on NewtonBench (Table 2), but consume their full budgets.

Together, these comparisons establish that early stopping and improved search contribute to performance, while learned allocation provides an additional computational-efficiency benefit relative to the evaluated fixed schedule. We have added this controlled comparison to Appendix C.4 and summarized the key finding in Section 5.2.

Table 1 cannot separate allocation from a stronger search engine.

We thank the reviewer for highlighting this important distinction. We agree that Table 1 alone cannot isolate the contribution of adaptive allocation from a stronger search engine. We have directly addressed this in the updated Sec. 5.2 (Table 2).

  • We introduce CoRE-Search and CoRE-Search+BED, which use CoREAgent's identical two-population search and PySR refinement, but follow fixed schedules with random or BED-guided acquisition. On NewtonBench, these controls achieve 88.9% and 89.8% accuracy versus CoREAgent's 88.0%, but exhaust both budgets, whereas CoREAgent uses only 40.3% of search and 69.5% of experiments. On ODEBench, CoREAgent achieves higher accuracy (84.4% vs. 80.7%) with substantially fewer resources.

Thus, Table 2 isolates the contribution of the search engine, which accounts for the accuracy gain. Table 8 separates learned allocation from early stopping, and shows that allocation reduces LLM calls at comparable recovery.

Seeds. Can you report at least three seeds per method with confidence intervals, paired across tasks, at least for NewtonBench and the Figure 4 ablations?

We thank the reviewer for emphasizing statistical robustness. We now evaluate three seeds on the same 15 NewtonBench tasks (five per difficulty level; 45 runs per setting) and report 95% bootstrap confidence intervals, resampling tasks while keeping each task's three seeds together.

  • NewtonBench ablations (Sec. 5.4, Fig. 4):
MethodAccuracy (%) [95% CI]Experiments (%)LLM calls (%)
CoREAgent88.9 [80.0, 97.8]67.149.7
w/o Metareasoner82.2 [71.1, 93.3]94.0100.0
w/o Diversity80.0 [68.9, 91.1]69.753.6
w/o Guided Acquisition82.2 [71.1, 93.3]75.252.5
w/o Feedback82.2 [71.1, 93.3]67.452.2

CoREAgent achieves the highest mean recovery (88.9% vs. 80.0–82.2%) while consuming the fewest resources. Although confidence intervals overlap, the results are consistent with the original ablation trends, with smaller gaps providing a more conservative assessment. Our separate stopping-matched comparison (Table 8) isolates the contribution of learned allocation from early stopping.

  • ForceBench (App. C.1, Table 6): Recognizing its small size (six tasks), we now evaluate three runs per method, reporting per-world successes and mean ± standard deviation. CoREAgent achieves 83.3 ± 0.0 accuracy, comparable to MDA's 77.8 ± 9.6, rather than drawing conclusions from single-task differences.

  • Search-matched controls (Sec. 5.2, Table 2). On NewtonBench, CoREAgent achieves comparable recovery (88.0% vs. 88.9–89.8%) while using only 40.3% of the search budget and 69.5% of the experimental budget, versus 100% for both controls. This reinforces our central finding: adaptive allocation achieves comparable discovery performance with substantially fewer resources.

  • Cost-sensitivity results and confidence intervals are continued in Response (2/2).

AuthorsOpenReview ↗

Response to Reviewer yjqQ (2/2)

Seeds (contd.)

Cost sensitivity (Sec. 5.3, Fig. 3). Results across the same 3 seeds × 15 tasks, with 95% task-bootstrap confidence interval. S/E = searches per experiment; E/10S = experiments per 10 searches.

CostAccuracy (%)S/EE/10S
Default88.9 [80.0,97.8]8.13 [6.90,9.59]1.23 [1.04,1.45]
S×1084.4 [73.3,93.3]7.73 [6.54,9.18]1.29 [1.09,1.53]
S×2093.3 [84.4,100]7.48 [6.51,8.74]1.34 [1.14,1.54]
S×3088.9 [80.0,97.8]6.98 [5.98,8.24]1.43 [1.21,1.67]
E×1082.2 [71.1,93.3]8.88 [7.89,9.99]1.13 [1.00,1.27]
E×2086.7 [75.6,95.6]9.41 [8.46,10.47]1.06 [0.96,1.18]
E×4084.4 [73.3,93.3]10.80 [9.55,12.22]0.93 [0.82,1.05]

Allocation responds monotonically to costs: S/E decreases 14% as search cost rises to 30× and increases 33% as experiment cost reaches 40×. Mean recovery remains 82.2–93.3%. Although CIs overlap, these trends support cost-responsive allocation, with more conservative effects than the original single-run evaluation.

No state-dependent switching heuristic is compared; Add:

  1. A fixed search–experiment schedule with the same search engine and the same convergence stopping rule.
  2. A state-dependent heuristic, for example experimenting after k search steps without improvement in the best score.

We thank the reviewer for suggesting these additional controls. To address this concern, we evaluate (1) a fixed search–experiment schedule, which performs an experiment after a fixed number of SEARCH steps (4 or 10), and (2) a state-dependent heuristic, which triggers an experiment after (k) consecutive SEARCH steps without improvement in the best hypothesis score (k=2 or 6). Both controls use the same hypothesis-search engine, acquisition policy, budgets, and convergence criterion as CoREAgent, differing only in the allocation rule.

We evaluate on the same 30 tasks used in our metareasoner-behavior analysis: 15 from NewtonBench and 15 from ActiveSciBench-Chem, stratified across three difficulty levels (5 tasks per level per benchmark).

Switching ruleAccuracy (%) ↑LLM calls (%) ↓Experiments (%) ↓
CoREAgent70.058.085.0
Fixed (4)50.044.389.0
Fixed (10)66.786.085.5
No-improvement (k=2)56.731.586.5
No-improvement (k=6)53.367.383.5

CoREAgent achieves the highest recovery (70.0%) among the evaluated controls. Compared with the fixed schedule (10), it improves recovery by 3.3 percentage points while using 32.6% fewer LLM calls. Against the state-dependent heuristic (k=6), it improves recovery by 16.7 points with nearly identical experimental usage. These findings provide additional evidence that learned allocation offers a favorable balance between recovery and resource consumption relative to the evaluated fixed and heuristic policies.

Metareasoner behaviour. Given the forced initial 20 Search steps, first three experiments, and 20-step Search cap, what fraction of decisions use Thompson sampling? How does CoREAgent compare with a frozen value model that samples actions without learning?

We analyze the 30 tasks (15 NewtonBench and 15 ActiveSciBench-Chem), one run per task and we assign every SEARCH/EXPERIMENT decision to the rule in App. B.3 that made it:

  • Thompson sampling: both actions are feasible, and the learned values decide.
  • Calibration: until an action has two credits, it is taken whenever feasible, so that the value model has data before Thompson sampling begins. This covers the opening Search steps and the second and third experiments.
  • Feasibility: only one action is allowed. The cause is the 20-step cap on Search blocks, the requirement of minimum competitive hypotheses before an experiment, or the budget.

The two Search steps required after each experiment are fixed by construction, so we exclude them below. Of the remaining decisions, Thompson sampling makes 73.9%, calibration 19.8%, and feasibility 6.3%. After the third experiment, Thompson sampling makes 97.3% and calibration none. The rules therefore mainly initialize the controller, and learned values govern allocation for the rest of the run.

Frozen-prior comparison. We froze the value model at its prior, so actions are sampled without learning; search, experiment design, the feasibility rules, and stopping are unchanged. We also disabled calibration, because it only gathers the first credits for learning; with a frozen model, it would choose SEARCH up to the cap in every block, turning the control into a fixed schedule. With learning, CoREAgent solves 70% tasks versus 60% (at 85% vs. 83.5% of the experimental budget) by investing additional search where its value model predicts gains, using 58% LLM calls versus only 29% for the frozen variant. Because a frozen model never uses credits, its allocation also cannot respond to action costs.

AuthorsOpenReview ↗

Request to Review Our Rebuttal

Dear Reviewer yjqQ,

We are writing to kindly follow up on our rebuttal. If there were any remaining concerns, we would sincerely appreciate the opportunity to clarify them before the discussion period for the authors ends. We believe our responses have addressed in detail the full set of questions/concerns you raised.

Should there be no additional concerns, we kindly ask you to consider revising your score.

Thank you for your time and thoughtful feedback!

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

The paper presents CoREAgent, a framework for symbolic regression-based scientific discovery, which explicit learns a policy that interleaves (a) hypothesis (equation) generation to explain known, collected data (which the paper refers to as "search"), and (b) data collection to expand the set of points that need to be explained by current and subsequent hypotheses (which the paper refers to as "experiment"). The framework is evaluated on 4 SR benchmarks and shows improvement over non-LLM and LLM-based baselines. Ablations are also provided showing which components contribute to the final performance of the proposed system.

Critical strengths
  1. I think the question being tackled in this paper is interesting, and (beyond SR) is one that shows up often in science as when to commission resources to collect additional data vs. spend those resources in refining and testing hypotheses (though, in this SR instantiation, the hypothesis refinement and testing is cheap, which may make the studied design less generalizable).
  2. Empirical results presented support the idea of an LLM-based SR system that appears competitive with alternative methods (including other LLM methods).
Critical weaknesses
  1. As alluded to in "strengths", my main concern is in the framing of the paper as tackling scientific discovery broadly. In my view, and in the evidence presented, the framework applies well to symbolic equation discovery. Whether it applicable to broader scientific discovery settings that may involve much more expensive (in both time and resources) experimentation is unclear. It would be helpful to hear any points that the authors may offer related to this.
  2. Given that learning a metareasoner is a core aspect of the proposed framework, the experiment setup in its ablation appears weak. Currently, the no-metareasoner uses a fixed schedule that is continued until the budget is exhausted. However, simpler, quite intuitive, heuristics should be compared against instead, e.g., running the experiment step only after the search steps don't show improvements for a fixed number of steps, or using hypothesis disagreements and a threshold over them to decide when to run experiment.
Questions for authors
  1. Please see weaknesses (both points).
  2. In Fig. 3, how did performance of the best-found solutions vary when varying the relative costs compared to baselines?
3 comments
AuthorsOpenReview ↗

Response to Reviewer WgsQ (1/2)

Thank you for dedicating your time and expertise to review our submission. Please find our responses below.

Applicability to broader scientific discovery settings: Beyond symbolic equation discovery More expensive (in both time and resources) experimentation is unclear. Discussion on broader scientific discovery

We thank the reviewer for this thoughtful point. CoREAgent targets interactive mechanism discovery and its metareasoner acts on candidate hypotheses, experimental feedback, and action costs, and is not tied to equations. Our existing benchmarks span physics (NewtonBench), chemical kinetics (ActiveSciBench-Chem), force laws (ForceBench), and dynamical systems (ODEBench). We address both parts of the concern below.

Beyond symbolic equations. To evaluate generalization beyond symbolic equations, we additionally study 45 ActiveSciBench-GRN tasks spanning five network motifs, three difficulty levels, and three variants (App. C.6, Table 11). Unlike equation discovery, hypotheses are signed directed graphs, and experiments involve controlled gene perturbations. CoREAgent uses GPT-4o-mini with a maximum of 20 experiments similar to the design by ActiveSciBench-GRN.

MethodAccuracy (%)Edge F1 (%)
LLM-only0.050.4
LLM-AutoSciLab31.172.5
CoREAgent77.897.0

CoREAgent recovers 35/45 regulatory graphs, compared with 14/45 for LLM-AutoSciLab and 0/45 for LLM-only. These results provide direct evidence that the framework can accommodate discrete graph-structured mechanisms, rather than being restricted to symbolic equations.

More expensive experimentation. We agree this is central to broader applicability. To mimic settings where experiments are far costlier than computation, our cost-sensitivity study (RQ3; Sec. 5.3, Fig. 3) raises the cost of each experiment up to 40× relative to search. CoREAgent adapts as intended: it extracts more from existing evidence (up to 33% more searches per experiment) and runs experiments up to 25% less often, while maintaining a recovery above the strongest baseline (78%).

Broader scientific discovery. We acknowledge that our evaluations do not capture the full complexity of real-world experimentation. Equation and graph discovery offer known ground truth, controllable experiments, and objective evaluation, which makes them rigorous, verifiable testbeds for the discovery process. In real science, mechanisms are unknown, experiments are costly and hard to reproduce, and validation is much harder. We view these evaluations as an important step toward broader scientific discovery, while recognizing that validating CoREAgent in real-world experimental settings remains a natural and important direction for future work and will add it to the conclusion.

In Fig. 3, how did performance of the best-found solutions vary when varying the relative costs compared to baselines?

We thank the reviewer for this question. We additionally compare the accuracy of best-found solutions against existing baselines across different cost settings. We evaluate 15 NewtonBench tasks selected that cover three difficulty levels, using GPT-4o-mini.

Method / Cost SettingExact Accuracy (%)
LLM-only62
LLM+BED64
LLM-AutoSciLab78
CoREAgent (Default)89
Search 10×84
Search 20×93
Search 30×89
Experiment 10×82
Experiment 20×87
Experiment 40×84

Comparison with baselines. Across the evaluated cost settings, CoREAgent achieves 82–93% exact recovery, compared with 78% for the strongest baseline (LLM-AutoSciLab). Fixed baseline trajectories remain unchanged when only reporting-cost weights are varied, whereas CoREAgent adapts its SEARCH–EXPERIMENT decisions to the modeled costs.

Cost-sensitive allocation. As search cost increases, searches per experiment decrease by up to 14%, while the experimentation rate increases by up to 16%. Conversely, increasing experiment cost raises searches per experiment by up to 33% and lowers the experimentation rate by up to 25% (Sec. 5.3, Fig. 3). Thus, the learned controller responds systematically to relative action costs while largely preserving recovery.

AuthorsOpenReview ↗

Response to Reviewer WgsQ (2/2)

Ablation appears weak. Implement fixed control where the experiment step is chosen only after the search steps don't show improvements for a fixed number of steps, or using hypothesis disagreements and a threshold over them to decide when to run experiment.

We thank the reviewer for this suggestion and have implemented both heuristics, along with fixed schedules (App. C.5, Tables 9–10). Every rule shares CoREAgent's search engine, experiment selection, Search-step constraints, budgets, and convergence-based stopping, so only the decision of when to experiment differs. Unlike the original ablation, no rule runs until its budget is exhausted. We evaluate 30 tasks (15 NewtonBench and 15 ActiveSciBench-Chem, five per difficulty level; one run per task).

  • No-improvement: experiment after k = 2 or 6 Search steps without improvement in the best hypothesis.
  • Disagreement: experiment when the maximum expected information gain exceeds τ = 0.5 or 1.0 bits.
  • Fixed: experiment every N = 4 or 10 Search steps.
Switching ruleAccuracy (%) ↑LLM calls (%) ↓Experiments (%) ↓
Fixed (N=4)50.044.389.0
Fixed (N=10)66.786.085.5
No-improvement (k=2)56.731.586.5
No-improvement (k=6)53.367.383.5
Disagreement (τ=0.5)56.767.084.0
Disagreement (τ=1.0)56.780.579.5
CoREAgent70.058.085.0

CoREAgent achieves the highest recovery of all six rules. Against the heuristics the reviewer suggested, it recovers 13.3–16.7 points more and uses fewer LLM calls than all of them except the cheapest no-improvement rule (k = 2), which recovers 13.3 points less. The only rule close in accuracy, Fixed (N = 10), needs far more computation: CoREAgent uses 28.0% fewer LLM calls. Experiment usage is similar across all rules (79.5–89.0%).

These results suggest that neither search stagnation nor hypothesis disagreement alone is a sufficient signal for when to experiment. CoREAgent's value model uses both, along with budget and diversity features, and learns their cost-adjusted value from subsequently acquired evidence.

AuthorsOpenReview ↗

Request to Review Our Rebuttal

Dear Reviewer WgsQ,

We are writing to kindly follow up on our rebuttal. If there were any remaining concerns, we would sincerely appreciate the opportunity to clarify them before the discussion period for the authors ends. We believe our responses have addressed in detail the full set of questions/concerns you raised.

Should there be no additional concerns, we kindly ask you to consider revising your score.

Thank you for your time and thoughtful feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful and constructive feedback. We are encouraged that reviewers found:

  • The important and under-explored question, when is a costly experiment worth more than further search? (yjqQ, WgsQ, 5uH7).
  • Our prequential credit assignment principled (yjqQ).
  • CoREAgent competitive with LLM and non-LLM baselines (WgsQ).
  • The paper clear (WgsQ, yjqQ), with well-designed figures (5uH7).

Across benchmarks, CoREAgent improves accuracy over the strongest baseline on NewtonBench, ActiveSciBench-Chem and ODEBench using fewer experiments, and matches MDA on ForceBench.

While reviewers were largely positive about the problem and framework, their main shared concern was the benefit of the learned metareasoner itself, which we have addressed using the new experiments (specifically Table 2 and Table 8). Beyond our individual responses, we have incorporated the following changes suggested by the reviewers into the paper, marked in blue for readability:

  • Search-matched controls (Sec. 5.2, Table 2; yjqQ): fixed-schedule variants of LLM-only and LLM+BED that use our exact search engine.

  • Stopping-matched control (App. C.4, Table 8; yjqQ): with search and stopping matched, CoREAgent reaches comparable recovery with 21.6% fewer LLM calls.

  • Alternative simple switching rules (App. C.5, Tables 9–10; WgsQ, yjqQ): CoREAgent achieves the highest recovery.

  • Multiple seeds for the ablations and cost sweeps (Figs. 3–4), and three runs per method on ForceBench (Fig. 2, Table 6; yjqQ).

  • Evaluation beyond symbolic equations (App. C.6, Table 11; WgsQ): on gene regulatory network recovery, CoREAgent recovers 35/45 graphs exactly, vs. 14/45 for the strongest baseline.

  • LLM-assisted Bayesian optimization in Related Work and Table 3 (5uH7), to clarify the distinction from prior methods.

  • Training-data contamination discussion and safeguards (App. A.1; 5uH7).

We thank the reviewers again for their valuable suggestions, which have substantially strengthened the paper. We would be grateful if the reviewers would consider these results in their final assessment, and we are happy to address any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.