acceptodds
Withdrawn from ICLR 2027

Verifier-Guided Skill Evolution for Autonomous Spatial Omics Analysis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial omics platforms now pair histopathology images with spatially resolved transcriptomes at scale, yet turning such data into biological conclusions still depends on expert judgment at nearly every step. Recent large language model agents can invoke bioinformatics tools from natural language, but treat each dataset as an isolated episode and discard whatever they learn. We argue the missing ingredient is not a better planner but a usable reward: without a signal separating a biologically sound output from a merely runnable one, an agent has nothing to learn from. We present BioSkill-Evolver, an agent that stores each analysis capability as an executable skill annotated with explicit preconditions and postconditions, moving composability checking from the language model into a symbolic test, and scores every execution with a multi-granularity verifier combining statistical, marker, pathway and spatial checks with a language-model judge. Because this verifier turns an unlabeled analysis into a scalar, it can drive an evolution loop that rewrites failing skills from their own error traces and splits high-variance skills into platform-specific variants under a conservative acceptance rule. Across end-to-end episodes on public Visium, Xenium and Stereo-seq sections, BioSkill-Evolver raises autonomous task success from to and the verification score from to relative to a stateless agent with an identical backbone, while the skill library grows from to skills through three platform specializations. Ablations localize the effect to the evolution loop and the verifier rather than to perception or planning, and the ranking is preserved for three of four alternative backbones. We report this as a single-run, local-benchmark study and state precisely which design mechanisms it exercised.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Continual Skill Learning, Spatial Transcriptomics, Verification-Guided Self-Improvement, Automated Scientific Analysis

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_rD6zOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents BioSkill-Evolver, an agent that improves reusable code skills for spatial omics analysis through execution feedback. Skills carry explicit preconditions and postconditions, allowing the system to check their compatibility when constructing analysis pipelines. A verifier combines statistical, marker, pathway, spatial, and language-model assessments to guide skill refinement and platform-specific specialization. Evaluation covers spatial domain identification, cell-type annotation, expression prediction, and autonomous analysis. In a single run of 38 end-to-end episodes, task success rises from 7.9% to 50.0% compared with a stateless agent using the same backbone.

Critical strengths
  1. The paper addresses a meaningful problem: successful code execution does not guarantee biologically sound analysis. Using biological feedback to improve reusable skills is a well-motivated direction.

  2. The experiments cover several spatial omics tasks and include comparisons with a fixed backbone and evolution ablations. The discussion of unsuccessful transfer and weak component gains is useful.

Critical weaknesses
  1. The writing is often awkward, repetitive, and difficult to follow. The manuscript needs substantial language revision, particularly to explain the central idea more directly.

  2. The methodological contribution is unclear. The framework combines many engineering components, but the paper does not convincingly establish what is new beyond their integration.

  3. The complexity is insufficiently justified. Removing feature fusion, skill-graph composition, or beam search barely changes performance, raising questions about whether these components are necessary.

  4. Many important decisions rely on fixed heuristics, including retrieval weights, evolution thresholds, and specialization rules. Their justification and sensitivity are not adequately examined.

  5. The related-work discussion insufficiently covers biomedical skill evolution. Relevant work includes executable skill libraries in SkillFoundry and GuideSkill, procedural skill memory in SkeMex, and evolving reasoning templates and tool libraries in STELLA [1–4]. Comparing their skill representations, feedback mechanisms, and update strategies would help clarify the novelty of this work.

  6. The biological verifier lacks independent validation. Agreement among its own scoring components does not establish biological correctness, especially when the pathway score conflicts with other signals. Expert assessment would provide stronger evidence that evolution improves scientific quality.

References

[1] Shen et al. SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources. arXiv:2604.03964, 2026. https://arxiv.org/abs/2604.03964

[2] Cao et al. GuideSkill: Evolving Executable LLM Agent Skills for Guideline-Grounded Clinical Reasoning. arXiv:2607.26160, 2026. https://arxiv.org/abs/2607.26160

[3] Sun et al. Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory. arXiv:2606.09365, 2026. https://arxiv.org/abs/2606.09365

[4] Jin et al. STELLA: Self-Evolving LLM Agent for Biomedical Research. arXiv:2507.02004, 2025. https://arxiv.org/abs/2507.02004

Questions for authors

see Critical Weaknesses.

Reviewer_5y4sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents BioSkill-Evolver, an LLM agent for spatial omics that keeps a library of executable analysis skills and revises it across episodes. The central argument is that self-improvement in this domain is limited by the lack of a reward, not by the planner. Pipelines nearly always run, so execution success says little about whether a result is biologically sound.

Each skill is stored as code plus pre- and postconditions from a closed vocabulary, so composability becomes a set-containment check. Plans come from beam search under precondition, goal-coverage and conflict constraints. Outputs are scored by BioAudit. It combines statistical, marker, pathway and spatial checks with an LLM judge (GLM-5.1, which is also the planning and coding backbone), and scales the result by the fraction of requested outputs actually produced. Skills whose running score is low or unstable are rewritten by the LLM from their own failure traces, and a rewrite is kept only if it beats the old version on validation episodes. Skills whose failures concentrate on one platform are split into platform-specific variants.

The evaluation covers domain identification, cell-type annotation, expression prediction from H&E (within-section 80/20 spot splits) and 38 end-to-end episodes on Visium, Xenium and sequencing-based data. On the task-level benchmarks the method has the best mean on almost every metric, with some per-dataset exceptions. In the end-to-end setting it is compared with a stateless agent on the same backbone. Success rises from 3/38 to 19/38 and V_final from 0.311 to 0.463, while the library only grows from 19 to 22 skills, all three additions being platform specializations. Ablations put nearly all of the gain on the evolution loop and the verifier. The appendix adds checkpoint curves, transfer and backbone studies, and an internal-consistency check of the verifier, in which the pathway signal is anti-correlated with the others. All results come from a single run per configuration, and there is no expert evaluation.

Critical strengths

S1. The paper puts its finger on the right bottleneck, and it shows this with evidence that does not depend on its own verifier. Every system in Table 3 executes 35 or 36 of the 38 episodes, so execution success hardly separates them. In Table 1 the three stateless agents produce exactly the default Scanpy/Leiden result. In this domain a pipeline that runs is not necessarily a pipeline that is right, and agent scaffolding alone does not change the analysis. This is the main reason I see value in the paper. For agents doing scientific analysis, verification seems to deserve more effort than planning, and that lesson is useful even if this particular system is not adopted.

S2. The reporting is unusually transparent. Table 5 marks which mechanisms actually ran; merging, the diversity regularizer and the curriculum did not. The negative results are stated plainly: the pathway signal is anti-correlated with the rest of the verifier, the zero-shot Xenium-to-Visium transfer fails completely, and transfer to the sequencing-based track never yields a passing episode. Domain and annotation metrics can be recomputed from per-unit assignment files without any model call. This matters for my assessment because it makes the claims checkable.

Critical weaknesses

W1. The end-to-end results are measured with the method’s own reward. Table 3 has no label-based metric, and the full model’s Domain ARI and Annotation F1 are blank in Table 4. The headline gain (3/38 to 19/38 successes, V_final 0.311 to 0.463) therefore rests on SR and V_final.

SR is not defined precisely. The paper describes it as passing biological checking, which I read as a threshold on BioAudit, the same signal EvoCycle optimizes. That signal has several problems:

It has not been compared with expert judgment. Its pathway component is anti-correlated with the others (r = -0.70). Its LLM-judge term comes from GLM-5.1, the model that also writes and rewrites the skills. It is multiplied by the coverage factor κ, so an episode’s score jumps when it produces a missing output even if the analysis is unchanged.

The largest improvement in Table 7 happens between checkpoints 66 and 88: V_final goes from 0.2008 to 0.5123 and successes from 7/22 to 19/22. Library size does not change over that interval, and nothing in the paper rules out that the gain is mostly in κ. The headline also covers less than it seems. By the paper’s own numbers (19/38 overall, 19/22 on Visium+Xenium), none of the 16 sequencing-based episodes succeed.

This weakness drives my soundness score. The central thesis is that BioAudit is a usable reward, and the only evidence that optimizing it improves the analyses is the reward itself.

W2. The results that do use external labels may not be clean either. Tables 1, 2 and 6 are the main evidence not affected by W1, but there are three problems.

First, V_stat computes ARI, NMI and macro-F1 whenever a quantitative target exists. Sec. 4.5 notes that evaluation and evolution draw on the same pool of sections, so labels from the evaluation sections may have entered the reward during evolution.

Second, the expression skill was developed on the evaluation cohorts, with target genes chosen by full-section variance, and the skill itself is never described. The fused features of Eq. 8 include a neighborhood average of expression. A held-out spot may therefore be predicted partly from the measured expression of adjacent training spots, which no baseline in Table 6 has access to.

Third, the agent baselines raise doubts. Three rows of Table 3 agree to four decimals on every column, including V_final with its LLM-judge term, and LLMiniST reproduces Scanpy (Leiden) exactly in Table 1.

Combined with W1, this leaves no comparison in the paper that I can take at face value. If the answers to Q3 to Q5 rule these issues out, my view of the task-level results would improve substantially.

Questions for authors

Q1. What exactly is the success criterion behind SR? If it is a threshold on V_final or on BioAudit components, which threshold? Does any part of it not depend on the verifier? (W1)

Q2. Please report κ separately from the quality part of V_final, for Table 3 and for checkpoints 66 and 88 of Table 7. Which skill was rewritten between those checkpoints, and what did the rewrite change? This should be possible from the verifier traces you say you release. (W1)

Q3. During evolution, did V_stat ever compute ARI, NMI or macro-F1 against the released labels of sections that appear in Tables 1 and 2? Which sections were used for evolution episodes, and which for evaluation? (W2)

Q4. What model does the expression-prediction skill use, and what are its inputs for a held-out spot? Are the Eq. 8 features computed on the whole section before the 80/20 split, and is the neighborhood expression term used? (W2)

Q5. In Table 3, were ReAct agent (GLM-5.1), LLMiniST and GLM-5.1 (ReAct, no skills) run separately, and how do the first and third differ? Agreement to four decimals is surprising when V_final includes an LLM-judge term. Is the LLMiniST row the original method? (W2)

Q6. The stomics to visium+xenium row of Table 8 has exactly the values of checkpoint 286 and the Final row of Table 7 (0.5634, 0.5663, 0.8636). Is it a separate run?

Q7. A.11 shows that unassigned units count as errors (39 of 173 in C1, 101 of 306 in D1), so abstention cannot raise accuracy. Where does the accuracy gain in Table 2 come from? Please also give per-section HER2 results. Table 2 reports 0.7325, while the two sections in Fig. 2 are at 0.387 and 0.360.

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

 

The authors introduce BIOSKILL-EVOLVER, an agent for spatial omics analysis. The main summary finding of the authors' empirical evaluation is that a usable reward is preferable to a better planner for the purpose of engineering self-improving language agents for spatial omics analysis. While the application domain is relevant, the benefits of the LLM-as-a-judge setup in BioAudit are somewhat unclear in the absence of a non LLM-based verification mechanism. Additionally, the authors do not appear to have released the code for the work, which raises concerns for the reproducibility of the work. Lastly, the empirical evaluation, by the authors' own admission, hinges on the results of a single run. If the authors can make a convincing case for the real-world utility of their LLM-as-judge setup for biological plausibility, tighten the empirical evaluation, and release the code, I will be inclined to increase my score.

 

Critical strengths

 

The main strength of the work is that the application domain is relevant. Agents that can autonomously perform spatial omics analysis would accelerate research pipelines in tissue biology, pathology, and personalized medicine.

 

Critical weaknesses

 

I demarcate between major and minor points below.

 

MAJOR POINTS

 

  1. In the reproducibility statement, the authors claim to have released the plans, generated code, outputs, and verifier traces. These do not appear to be present with the submission however?

  2. In my view the principle weakness of the work is that the utility of the authors' setup is unclear. Do the authors expect self-improvement to be achievable solely from their LLM-as-a-judge setup for biological plausibility? What is the pathway towards an autonomous spatial omics agent and how does the paper contribute to that goal? Benchmarking against human experts on omics analysis tasks would help to establish this.

  3. It is clear that some parts of the paper are written by an LLM and as such, certain sections of the paper are difficult to follow. I would propose that the authors perform a human audit to improve clarity.

  4. By the authors' own admission their results hinge on a single experiment run. Why did the authors not perform multiple runs and report standard deviations over the metrics?

 

MINOR POINTS

 

  1. There are some missing capitalizations in the references e.g., "GHIST", "LLM", "SCANPY", "BayesSpace", and "GraphST".

  2. There are some missing conference citations in the reference list e.g., [1] was accepted to TMLR, [2] was accepted to COLM 2024, and [3] was accepted to ICLR 2023.

  3. There is no bioRxiv reference in the citation for OmicsNavigator.

  4. in Appendix A.3 it would be worth redefining terms such as Vˉka\bar{V}_{k_a} so that the reader does not need to refer to the main text.

  5. Line 855, typo, "Eq. equation 3".

  6. The ARI, NMI, and PCC evaluation metrics should be defined before being introduced as acronyms e.g., Pearson's correlation coefficient (PCC).

 

REFERENCES

 

[1] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, & Anima Anandkumar (2024). Voyager: An Open-Ended Embodied Agent with Large Language Models. Transactions on Machine Learning Research.

[2] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, & Chi Wang (2024). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversations. In First Conference on Language Modeling.

[3] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, & Yuan Cao (2023). ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations .

 

Questions for authors

 

  1. In Section 4.1 the authors state that they report no seed variance, why? Why not peform multiple runs per configuration?

 

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

Related papers

Discussion (0)

Sign in to comment.