acceptodds
Under review as a conference paper at ICLR 2027

EvoLab: Feedback-Driven Agents for Budgeted Wet-Lab Decisions

Abstract

Scientific agents increasingly use feedback to choose experiments, but a finite measurement budget makes each choice consequential twice: it yields discoveries now and determines which evidence can inform later decisions. We present EvoLab, which couples budgeted measurement selection with the interpretation and use of selectively acquired evidence. After each batch, it reviews what the observations change about the remaining decision problem, foregrounds provenance-linked evidence while retaining a retrievable record, and investigates feasible next allocations with numerical and retrieval tools. Its decision context can evolve while language-model weights remain fixed; outcomes of unmeasured alternatives remain hidden. We evaluate \method by sequential replay in molecular hit discovery, protein library design, and two drug-combination screening settings. Across these four reported settings, it improves mean final discovery yield over task-adapted ReAct by approximately 5–17% and reduces token consumption by 12–33%. Trajectories show delayed gains after prerequisite measurements in Hit and GFP and more productive revisits in Drug C2; a full-information Hit oracle quantifies remaining allocation headroom. These results support designing what to measure and how acquired evidence shapes the next decision together.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Active Learning, AI for scientific discovery, Agents

Reviews

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

The paper argues that the acquired evidence in sequential experiment selection influences not just the immediate discovery yield but also the follow-up batches in subsequent rounds. The proposed pipeline comprises an evidence-use decision step which highlights the guiding question for the next batch and selects relevant evidence from the whole record to substantiate it. The agent then compares the potential follow-ups based on cost, pre-requisites etc. and commits a batch for measurement. The method is evaluated on molecular hit discovery, protein library design, and two drug-combination settings, against numerical baselines and two agentic baselines based on the DeepSeek-V4.1-Flash backbone. EvoLab reports higher mean final discoveries than all baselines with lower token consumption.

Critical strengths

Following are the critical strengths of the paper:

  • Well formulated problem statement: Realistic to have different costs for different actions with ordering dependence between potential actions. Not every past outcome is necessary to observe at each round, and thus focused evidence becomes important.
  • The full information retrospective oracle analysis is interesting and sheds light on how the agent performance changes as rounds progress, highlighting issues with exploration/exploitation.
Critical weaknesses
  • The paper presentation is very confusing:
  1. Terminology related to experimental tasks (panels, grids etc.) is used extensively in the method section without being introduced.
  2. The paper talks about selecting relevant evidence to pass into the context but section 3.2 mentions all acquired raw records are presented.
  3. Towards the end of the manuscript, authors differentiate two versions of the method, single loop and explicit-review, without talking about it in the text before.
  • Issues with experiments:
  1. The results are only provided with DeepSeek-V4.1-Flash, where the questions of robustness to LLM choice remains unanswered.
  2. Section 4.3 digs deeper into breaking down the allocation trajectory without providing actual evidence of the hypotheses or the observations specified.
  3. Several captions specify the methods worked on different cohorts and model identifiers, which breaks the true paired comparison between methods and raises concerns about the conclusions drawn.
Questions for authors

Clarification Questions:

  1. Do the BioDiscoveryAgent and ReAct baselines receive exactly the same amount of information in the prompt as EvoLab?
  2. Table 1 is interesting as it shows how the agent performance compares to a full information oracle. Is there a reason why round 5 has such a high loss? Also, the table caption says the numbers are the lowest loss per round. Are the numbers across rounds corresponding to a single run for each method? Or do they correspond to different runs, wherever the lowest loss happened?
  3. In general, Figure 3 and Table 1 both mention that some numbers are with different ‘cohorts’. Please clarify what does it mean and why is it not exactly paired with the same seeds and model identifiers? If different seeds, could you provide results with the same seeds across multiple runs?
  4. Why are ablations done on the explicit-review variant and not the single-loop variant used in the main comparison? How do the two variants differ?

Follow-up Experiments:

  1. Is there any supporting evidence of the decompositions explaining agent behavior mentioned in Section 4.3?
  2. There is a lack of clarity about what actual prompts look like. Including a single round of LLM trace would have helped. For example, how is the evidence provided in the prompt, how many past records are selected at max, how many follow-ups are compared due to token constraints, is there 1 follow-up per question?
  3. Results on more LLMs will solidify the generality of the claims. Please provide results on at least 1 more LLM.
Reviewer_ngQpOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an agentetic method EvoLab for experimental design. The general framework is depicted in Fig. 2, where 4 components or stages of the agent in each loop are provided. The design of the agent (i.e. prompts) are provided in Appendix C, where in Appendix C1 a 'contract' is given to EvoLab and baselines. EvoLab can be run in 2 configurations (i) single-loop: The first prompts box in Appendix C2 or (ii) explicit: the third prompts box in Appendix C2.

The experiments are more or less convincing, and show improvements when compared to both numerical and agentetic methods, in terms of discovery success and token consumption.

Critical strengths

The empirical improvements are noticeable and significant, and on 3 tasks

  • molecular hit discovery
  • protein library design
  • drug combination screening
Critical weaknesses
  • The paper is really hard to read. Up until Sec. 3 (which the method is introduced), the emphasis is on delayed reward/gain, and little description about the big picture and the distinctions of the proposed method.

  • The 2 configurations of the proposed method are not consistently applied to datasets.

  • The first 3 pages only reiterate some challenges of experimental design, which is shared among almost all of previous methods.

For example

choice consequential twice: it yields discoveries now and determines which evidence can inform later decisions.

This is the classical exploration/exploitation trade-off, considered by almost all experimtal design methods.

These advances sharpen a practical question: when only a limited number of experiments can be performed, how should an agent use evidence created by its ow earlier choices?

This is the problem definition of experimental design, and not a gap to be picked and solved by the paper.

We introduce EVOLAB to couple two decisions: which measurements to acquire and how their outcomes shape the next allocation.

But any experimental design method is expected to do that.

Unmeasured alternatives remain unknown rather than becoming presumed failures.

This is also common between experimental design method.

Questions for authors
  • To understand the method, I went through Appendix C where the prompts are described. There doesn't seem to be any 1-to-1 correspondance between the boxes in Fig. 2 and the prompts provided in Appendix C? If so, please explain the correspondance explicitly.

  • What are the distinctions of the proposed method compared to baseline agents? For example, the task-specific interpretations (the 2nd box in Appendix C2) could be the distinctions?

  • More about the task-specific interpretaitons, the baseline agents are general and without those explicit ad hoc interpretations. That might make the comparision to baselines unfair.

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

This paper presents EvoLab, an LLM-based agent framework for selecting scientific experiments under a limited measurement budget, where each choice affects both immediate discoveries and the evidence available for later decisions. The framework reviews observed outcomes, organizes relevant evidence with links to supporting records, and uses retrieval and numerical tools to plan follow-up measurements while keeping the model’s weights fixed. Across four sequential-replay settings using existing experimental data for molecular hit discovery, protein library design, and drug-combination screening, the paper reports approximately 5–17% higher mean final discovery yield and 12–33% lower token consumption than task-adapted ReAct.

Critical strengths

The paper explicitly incorporates measurement costs, experimental prerequisites, and selectively observed outcomes, capturing the fact that a measurement can enable future discoveries without immediately producing one. The evaluation settings cover distinct decision structures, including costly confirmation assays, shared prerequisites, and follow-up measurements on unresolved candidates. This provides a meaningful setting for studying how agents use acquired evidence to allocate a limited experimental budget.

Critical weaknesses
  1. The method contribution is not sufficiently distinguished from existing agent frameworks. In the single-loop configuration used for the main Hit and GFP experiments, EvoLab primarily relies on task-specific instructions, evidence organization, and standard reasoning (tool interactions, without a mandatory review or separate planning stage). These mechanisms overlap with existing approaches such as ReAct and context engineering a lot.

    More importantly, the evaluation lacks a controlled ReAct baseline equipped with the same task-specific guidance and compact evidence representation. So it is unclear whether the reported improvements stem from the proposed coupling of evidence use and measurement selection or from better prompting and context management. This uncertainty directly weakens the paper's method contribution.

  2. The experimental evidence does not adequately isolate or validate the proposed mechanisms. First, the Hit comparisons use different experimental cohorts, including different seeds and recorded model identifiers (Figure 3 and Table 1), making the reported improvements difficult to attribute solely to the proposed method. Second, the ablations in Section 4.6 evaluate an explicit-review configuration, whereas the main Hit and GFP experiments use a different single-loop configuration without mandatory review. Thus, the reported benefits of review and selective evidence presentation do not directly establish their contributions to the main results. Additionally, persistent notes improve performance on Drug C2 but hurt performance on Hit and Drug C1. Taken together, these limitations leave the mechanisms responsible for EvoLab's improvements unclear.

Questions for authors
  1. Could the authors provide more details about the numerical prediction and acquisition tools used by EvoLab in each domain, including how they are trained or updated? In particular, how is the information boundary enforced to ensure that predictions and preprocessing do not use outcomes from unpurchased measurements?

  2. Appendix C mentions that the agent carries forward the actual executed purchase, including any fallback. How are invalid or infeasible actions handled, and how frequently do such fallbacks occur for EvoLab and the baselines? Could differences in fallback frequency affect the reported discovery yields?

  3. Since budget-aware decision-making is central to the paper, how sensitive are EvoLab's improvements to the measurement budget and relative assay costs? For example, do the advantages persist under smaller budgets or different cost ratios between prerequisite and confirmation measurements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.