acceptodds
Withdrawn from ICLR 2027

EviRead: AI Agents Learn to Read Biological Evidence through Recursive Improvement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI agents have made substantial progress in software engineering and mathematics, building on language models strengthened by data and model scaling. The same agent wave has also reached biology, with the hope of combining agents with bioinformatics tools to accomplish downstream tasks more intelligently and automatically. A gap remains between them, however: the outputs of biological tools, such as biological language models (BioLMs), do not automatically translate into better evidence an agent can read and reason over. We introduce EviRead, a two-level framework that turns this heterogeneous biological evidence into agent-readable evidence through recursive improvement (RI). We instantiate EviRead on TempoGO-HQ, a retrospective temporal benchmark of 1,389 proteins for Gene Ontology term prediction. The BioLM-only workflow EviRead_base runs five RI rounds and improves on both the training and validation splits, showing that the loop can diagnose and iteratively refine the workflow. The second, tool-augmented workflow EviRead_+tools integrates sequence alignment, structural search, and GO predictors, and its final workflow outperforms all twelve compared methods in ontology-macro Fmax. By making evidence construction and tool coordination learnable at the workflow level, EviRead offers a concrete route to connect specialized biological models and tools with agent-based inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI agents, protein function prediction, recursive improvement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WBmHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EviRead, a two-level framework that applies recursive improvement to biological evidence integration for protein function prediction. In EviRead, the inner workflow converts outputs from biological foundation models and bioinformatics tools into agent-readable evidence, which is then used to predict Gene Ontology (GO) terms. The outer RI loop iteratively modifies the evidence construction and prediction workflow based on evaluation feedback.

The authors evaluate two configurations, EviReadbase and EviRead+tools, on a retrospective temporal benchmark containing 1,389 proteins. They report improvements over five RI rounds in the model-only setting, while the final tool-augmented workflow outperforms twelve compared methods on the reported validation split.

Critical strengths
  1. The two-level framework is reasonably designed. The separation between biological evidence processing and recursive workflow improvement provides a structured approach to integrating heterogeneous biological tools into agent-based prediction.

  2. Applying RI to biological evidence integration is a promising direction. The framework combines automated workflow improvement with evidence provenance and domain-specific prediction constraints. The reported results suggest potential practical value for protein function prediction.

Critical weaknesses
  1. The outer RI loop largely follows existing agentic workflow optimization approaches. Although the paper introduces implementation details such as workflow verification and self-repair, it remains unclear what fundamentally new improvement mechanism is proposed.

The main potential contribution therefore appears to lie in the inner evidence-reading workflow. However, previous work has already explored LLM agents for biological evidence integration and protein function prediction. The paper does not sufficiently distinguish its evidence-reading mechanism from these related approaches or clarify what is learned beyond selecting, organizing, and combining existing biological evidence.

  1. Insufficient attribution of performance improvements to evidence construction and recursive improvement.

Although the paper demonstrates improvements in GO prediction performance, the experiments do not sufficiently disentangle improvements in evidence construction from task-specific optimization of downstream decision rules. In particular, the five reported RI rounds in EviReadbase primarily involve score fusion, normalization, and threshold adjustment. These changes may improve prediction performance without necessarily improving how biological evidence is selected, structured, or represented. While such optimization falls within the proposed framework, the experiments do not clearly establish the distinct contribution of evidence construction.

Furthermore, EviReadbase and EviRead+tools differ substantially in the available tools and evidence sources. Although the final tool-augmented workflow achieves strong results, the absence of an ablation comparing workflows with and without RI under the same tool and evidence configuration makes it difficult to isolate the contribution of recursive improvement from that of additional information and conventional evidence integration.

Questions for authors
  1. Can they further disentangle the contributions of evidence construction and downstream decision optimization? Such ablations would help establish whether RI improves evidence interpretation or primarily optimizes prediction rules.

  2. Why are the main results reported on only 50 training and 60 validation proteins, given that TempoGO-HQ contains 1,389 proteins? Can the authors provide independent test-set results?

  3. What specific methodological novelty does EviRead introduce beyond existing RI frameworks and biological LLM-agent approaches?

  4. How might the proposed evidence-reading workflow generalize to biological tasks beyond GO prediction?

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

The manuscript introduces EviRead, a framework designed to bridge the gap between biological foundation models and LLM agents. Because LLMs cannot natively interpret raw numerical embeddings or alignment scores, EviRead uses a host program to translate these signals into a "two-sided evidence object," compiling raw data into text-based candidate cards. An inner LLM agent then reads these cards to reason over conflicting biological evidence. The authors also propose an outer Recursive Improvement (RI) loop to optimize the underlying code managing this translation. The framework is evaluated on a custom benchmark (TempoGO-HQ) for Gene Ontology (GO) term prediction, comparing both a model-only and a tool-augmented workflow against standard bioinformatics baselines.

Critical strengths

Clear Motivation and Research Gap: The authors address a highly relevant problem by identifying that standard LLMs cannot natively process raw numerical embeddings from biological foundation models. Bridging this gap is a timely goal for the bioinformatics community.

Innovative Conceptual Design: The "two-sided evidence object" is a clever architectural approach. By separating secure, machine-readable data from natural-language candidate cards, the framework successfully balances technical auditability with LLM interpretability.

Critical weaknesses

Unclear Utilization of the Held-Out Test Set: While the authors construct the TempoGO-HQ benchmark with 1,389 protein targets, it is currently unclear how the remaining ~1,279 proteins are used in the final results. The manuscript explicitly notes in the Reproducibility Statement that "all primary results are reported on the training (50 proteins) and validation (60 proteins) splits of TempoGO-HQ", and the primary comparative metrics in Tables 1 and 5 appear to display columns only for these 110 targets. Because the validation set is actively used to select workflow modifications during the recursive loop, evaluating the final frozen system on a separate, untouched test set is critical for proving generalization. We could not locate explicit test set results for the remaining 1,279 targets, which leaves the framework's true out-of-sample performance unclear to us.

Internal Contradiction Regarding Validation Set Usage: The manuscript presents a direct internal contradiction regarding how the validation dataset is utilized during the recursive optimization process. In Section 3.2.2, the authors establish a strict methodological boundary, claiming that "the held-out validation and test labels are never used to select a revision". However, the experimental trace in Section 4.1 and Figure 4 directly contradicts this rule, as candidate workflow modifications are explicitly kept, rejected, or backtracked based on validation set performance. For instance, the agent rejects Round 4 because the validation score regressed to 0.072, and it selects Round 5 as the winning architecture due to a "large validation gain". Because validation performance actively drives candidate selection and search decisions, the validation set functions as an optimization feedback loop rather than an independent, untouched evaluation benchmark.

Insufficient Operational Detail in the Main Text: The most important operational mechanisms and technical details of the proposed framework are insufficiently explained in the main text, relying heavily on the appendix to clarify essential steps. Critical components—such as how dense model embeddings are converted into agent-readable text, how vector-based neighbor retrieval is computed, and how the host formats candidate cards for LLM reasoning—are treated only conceptually in Section 3. The main text must be self-explanatory and self-contained, reserving only granular supplementary configurations, extended schemas, or code snippets for the appendix. Currently, by reading only the main text, a reader is unable to understand what is operationally happening under the hood or reproduce the method's core translation workflow.

Evaluation Limited to a Single Downstream Task: The generalizability of the EviRead framework remains in question because it is evaluated on a single downstream biological task. While the authors propose EviRead as a broad solution for turning heterogeneous biological tool outputs into agent-readable evidence, all experiments are restricted exclusively to Gene Ontology (GO) term prediction. Evaluating the framework on additional diverse tasks would be necessary to demonstrate its versatility and robustness across different biological domains. Without testing beyond GO prediction, it is difficult to determine whether the framework's recursive improvement loop is universally effective or merely over-tailored to the specific hierarchical rules of GO annotations.

Absence of a Non-Agentic Ensemble Baseline: In Section 4.2, the tool-augmented workflow (EviRead_{+tools}) is shown to outperform individual biological tools such as BLASTP, Foldseek, and DeepGOPlus. However, the paper omits a critical comparison against a standard, non-agentic statistical ensemble (e.g., a logistic regression, random forest, or simple weighted average combining the scores of all these tools). Without this baseline, it is impossible to determine whether the performance gains stem genuinely from the agent's semantic reasoning capabilities or merely from aggregating multiple complementary evidence sources. Including a traditional tool-ensemble baseline is necessary to isolate and prove the exact value added by the LLM reasoning layer.

Confounding Variables Between the Two Workflows: The experimental design introduces a significant confounder between the model-only (EviRead_{base}) and tool-augmented (EviRead_{+tools}) workflows, making direct comparison difficult. Specifically, Section 4 states that EviRead_{base} is restricted from accessing the database's known annotations for the query protein, whereas EviRead_{+tools} receives "all T_0-time database information for the query protein" alongside the external tools. Because EviRead_{+tools} gains access to prior query metadata that the base workflow lacks, it is unclear how much of the performance jump is attributable to the tool layer and agent reasoning versus the sudden inclusion of the query protein's historical database context.

Unspecified LLM Architectures and Computational Overhead: The manuscript omits critical reproducibility details regarding the specific language models and computational resources utilized to power the inner workflow. Appendix C.3 broadly mentions relying on a "high-reasoning language model" for GO decisions and a "general-purpose model" for synthesis and review, but completely fails to identify the exact models, parameter counts, or API versions used. Furthermore, the paper provides no profiling of inference latency, token consumption, or financial cost per protein target. Given the multi-stage, multi-prompt nature of the framework, quantifying the computational overhead is essential for evaluating its practical viability compared to standard, fast bioinformatics pipelines.

Absence of Qualitative Outputs and Traceability Examples: A central motivation of the paper is that EviRead converts heterogeneous raw outputs into "agent-readable evidence" and produces "an auditable explanation" complete with qualitative confidence, alternatives, and citations. Despite this heavy emphasis on interpretability, the manuscript does not provide examples of these generated artifacts. The total absence of concrete examples of the textual candidate cards, the generated synthesis narratives, or the independent critic logs makes it impossible for reviewers to assess the actual coherence, biological accuracy, and citation validity of the agent's reasoning.

Questions for authors

The following are clarification questions:

Clarify Data Splits: Provide a clear, unambiguous definition of the dataset splits, explicitly detailing the exact sizes and roles of the training, validation, and test sets within the 1,389-protein TempoGO-HQ benchmark.

Address Validation Usage and Reporting: Explicitly clarify whether the validation set is used for optimization. If it is actively used to select and keep workflow revisions, acknowledge this and provide the out-of-sample results on an untouched test set, rather than reporting the optimized validation metrics as the primary measure of performance.

The following are suggestions for improvement:

Expand to Additional Downstream Tasks: Evaluate the framework on at least one other diverse biological task beyond Gene Ontology (GO) term prediction to demonstrate the method's generalizability and robustness across different domains.

Make the Method Description Self-Explanatory: Revise the main text to explicitly detail the core translation mechanism. A reader must be able to understand exactly how raw, agent-unreadable numerical representations (such as embeddings) are operationally converted into agent-readable text representations without needing to search through the appendix.

Reviewer_wk7dOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EviRead uses an outer program-editing loop to improve a protein-function prediction workflow built from frozen biological models and optional tools. It combines donor evidence, GO decisions, and traceable explanations under temporal resource cutoffs. Reported experiments cover a five-round base workflow and an augmented workflow evaluated on 50 training and 60 validation proteins.

Critical strengths
  1. The evidence representation links predicted GO terms to donors and recorded sources, making the workflow’s decisions inspectable.
  2. Versioned program edits, rollback, and temporal resource restrictions provide a useful framework for studying automated workflow improvement.
Critical weaknesses
  1. The evaluation boundary is inconsistent. Section 3.2.2 states that held-out validation labels never guide revision selection, but Section 4.1 says each round reads training and validation scores to plan edits. Main comparisons use only 50 training and 60 validation proteins, with no untouched test result. Adaptive validation feedback would invalidate treating these scores as independent evidence of generalization.

  2. The experiments do not isolate the benefit of recursive improvement. The augmented workflow adds both tools and query-specific T0 annotations, while the base trace mainly changes evidence fusion and thresholds. Comparisons against individual predictors, without a fixed ensemble or conventional search using the same evidence and budget, cannot establish the claimed advantage of recursive evidence reading.

Questions for authors
  1. How does recursive improvement compare with a fixed evidence-fusion baseline and conventional parameter search under identical evidence access and total search budgets?

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

Related papers

Discussion (0)

Sign in to comment.