acceptodds
Under review as a conference paper at ICLR 2027

When New Experiments Arrive: An Evidence-adjudicated Self-correction Framework for Virtual Cells

Abstract

Virtual-cell models increasingly need to support both accurate perturbation prediction and interpretable biological reasoning. However, numerical models provide precise continuous responses but only implicit biological knowledge, whereas LLM/knowledge-graph-based models offer explicit mechanistic reasoning but coarse categorical predictions. Jointly maintaining these capabilities raises a further challenge: as new experimental evidence arrives, which states should be corrected and which should be preserved? To address these challenges, we propose **VCrectify**, an evidence-adjudicated self-correction framework for virtual cells. It maintains a numerical predictive state for continuous perturbation prediction and an explicit knowledge state for knowledge-grounded mechanistic reasoning. Cross-state disagreement between the two states guides the active evidence acquisition of new perturbation experiments. Once outcomes are revealed, an evidence adjudicator determines which states receive correction. The updated states then guide subsequent experiment selection, forming a continual loop of evidence acquisition and self-correction. Experiments across multiple virtual-cell backbones and perturbation datasets show that VCrectify improves numerical prediction and knowledge-based directional reasoning, generalizes better to unseen cellular contexts, and produces higher-quality mechanistic rationales.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell, Self-Correcting, Continual Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KmknOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VCrectify maintains a numerical perturbation predictor alongside an explicit knowledge-based reasoning state. Disagreement between them guides the acquisition of experimental evidence, and observed outcomes determine which state is updated. Numerical updates use replay; knowledge updates adjust the reliability of cited information. Experiments span several prediction and reasoning backbones, perturbation datasets, and transfer settings. Acquisition and update ablations test the workflow, while a blinded expert comparison assesses its rationales. The results support the combined prediction approach, although correction of individual knowledge items and retention over successive updates are less directly evaluated.

Critical strengths
  1. The two-state design addresses a useful problem. New measurements need not correct a numerical predictor and an explicit knowledge state in the same way. Connecting disagreement-based acquisition with selective updates gives the framework a clear role in adapting virtual-cell models as experiments arrive.

  2. The gains extend across several backbone combinations (Tables 1–2), with update-all, random-correction, and random-acquisition controls. The blinded rationale assessment also uses the same graph and experimental memory in both conditions (Appendix C.4). These comparisons make the evidence stronger than a collection of prediction scores or selected explanations alone.

Critical weaknesses
  1. The sample-efficiency result needs stronger attribution. Random acquisition uses matched experimental budgets, and standalone baselines receive the full training split. However, numerical-uncertainty and diversity-based acquisition strategies are not compared. Total gradient steps and optimization schedules are also insufficiently specified for the sequential-versus-joint comparison. Thus, the benefits of the disagreement rule are not fully separated from selection strategy and training exposure.

  2. Prediction correction is better established than knowledge correction. Equation 7 updates the reliability of all cited knowledge from the final prediction outcome, without identifying which item caused the error. Expert source checks support rationale quality, but Appendix C.4 compares final learned weights with unit weights. That comparison combines initial calibration and continual correction, so it does not isolate the latter. The authors disclaim causal validation; the remaining issue is attribution to individual knowledge items and the continual update stage.

  3. The retention analysis mainly links the adaptation coefficient to final prediction error (PDF p. 9). It does not show whether earlier queries remain accurate as new experiments arrive. Replay is a reasonable mechanism, but the claimed retention benefit needs a direct assessment of forgetting along the update stream.

Questions for authors
  1. What batch sizes, acquisition rounds, and total gradient steps are used in the full, random-acquisition, and joint-training variants? Is a matched comparison with numerical-uncertainty or diversity-based acquisition available? The experimental budgets are already reported as matched; the question is whether selection and optimization exposure explain the gains.

  2. Can final reliabilities be compared with calibrated initial reliabilities held fixed, using the same final graph and experimental memory? The current unit-weight control includes both calibration and continual correction. For item-level examples, how do reliability changes relate to the source checks performed by the experts?

  3. Can saved checkpoints be evaluated on a fixed set of earlier queries after each update round? This would show directly whether replay preserves previous performance during the existing update streams.

Reviewer_nq8eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VCrectify, a framework that maintains a numerical perturbation predictor and an LLM/KG-based reasoning state. Disagreement between the two branches guides experiment acquisition. Once experimental outcomes become available, separate error thresholds determine whether each branch should be updated. The numerical branch is updated using supervised learning with replay, while the knowledge branch adjusts reliability weights associated with cited knowledge items. Experiments cover held-out perturbations, cross-context transfer, component ablations, and preference-based evaluation of mechanistic rationales.

Critical strengths
  1. The paper addresses a meaningful adaptation problem. New perturbation measurements may challenge a numerical predictor, a knowledge-based prediction, or both. Explicitly separating these update decisions is a useful perspective for maintaining virtual-cell systems as evidence accumulates.

  2. The framework is evaluated across multiple backbone combinations and settings. The reported improvements in continuous and directional prediction, together with the update-selection ablations, provide evidence that the overall learning procedure is useful. Its compatibility with several existing backbones also supports its practical relevance.

Critical weaknesses
  1. Prediction errors are attributed to all cited knowledge without identifying the cause. Under Eqs. (1) and (7), knowledge items cited in the same prediction receive the same correctness signal. If the prediction fails because one item is wrong, or because the LLM misuses correct evidence, valid items also receive negative feedback. A low reliability score therefore does not distinguish incorrect knowledge from reasoning errors or context mismatch. The experiments do not establish that these updates selectively identify unreliable knowledge.

  2. The baselines do not establish why the proposed mechanism is needed. The paper compares learned weights with unit weights, but does not test whether shuffling the weights across knowledge items preserves the gains. It also lacks a numerical acquisition baseline using the same replay mechanism and training budget. The improvements could therefore arise from changes in evidence selection and training order, without accurately identifying knowledge reliability or requiring both branches.

Questions for authors
  1. Can the method distinguish incorrect knowledge from correct knowledge that the LLM misuses? A controlled evaluation with known incorrect items, together with shuffled weight assignments, would help establish whether the learned scores identify unreliable knowledge.

  2. How does VCrectify compare with acquisition based only on numerical model uncertainty, using the same evidence budget, replay mechanism, and optimization budget? This comparison would clarify whether knowledge predictions provide additional acquisition value.

  3. Does one model maintain shared reliability weights across cell lines, or separate weights for each context? How well do weights learned in one cell line transfer when frozen in another? The numerical transfer results do not directly answer this question.

  4. Since low weights reduce retrieval priority and uncited items remain unchanged, what allows an erroneously downweighted item to regain influence? Has recovery been measured, particularly after switching cellular contexts? This is a potential failure mode requiring verification, rather than an established empirical failure.

Reviewer_q5TdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

VCrectify maintains a continuous perturbation response predictor and an LLM/KG reasoner as new experiments arrive. Their disagreement guides experiment selection. Once outcomes are observed, separate error thresholds determine whether to fine-tune the numerical model with replay, reweight the reasoner's cited evidence, update both, or update neither. The reasoning model's language model parameters remain fixed during correction. The paper reports gains across several numerical / reasoning pairings, context-transfer settings, and evaluations.

Critical strengths
  1. This is a practical problem and a modular experimental-feedback loop, as learning from newly available perturbation measurements is relevant to virtual cell. VCrectify accommodates different backbones and uses observed experiments to supervise correction, avoiding reliance on agreement between two models as a substitute for experimental evidence.

  2. Encouraging results across multiple settings. Table 1 reports improvements across 9 numerical / reasoning pairings, with standalone numerical baselines. The paper also compares update policies and random acquisition, examines context transfer, and evaluates rationales through blinded expert and LLM assessments.

Critical weaknesses
  1. The added value of maintaining two separate models is not sufficiently established. Combining a specialized numerical predictor with a knowledge-based reasoner is reasonable, but the paper needs to show what their interaction contributes beyond continual adaptation of either branch. The numerical branch is not directly trained from the reasoner's mechanisms, so its benefit depends largely on the experiments selected through disagreement. PertReasonLM also makes a reasoning-centered approach relevant to the directional and explanatory claims. It is not a replacement for continuous expression prediction, but the comparison raises an important design question: where does maintaining two states help more than adapting a reasoning model or numerical predictor alone?

  2. It might be hard to apply the golden ground truth of biological mechanism to the prediction task.

2.1. Answer correctness is an ambiguous basis for correcting knowledge. A valid fact used in the wrong context can lose weight, while an irrelevant citation accompanying a correct answer can gain weight (based on Equations 1 and 7). Appendix E acknowledges that the weights do not establish causal validity, which is appropriate. The remaining concern is whether they reliably improve evidence use, and when an answer error instead causes sound knowledge to be suppressed.

2.2. Cross-state disagreement does not necessarily equal to a conflicting biological prediction. Equation 2 compares the sign of a continuous mean effect with a 3-way classification label, in which 0 denotes no differential expression. The score also emphasizes large predicted effects and reduces to response-magnitude selection if the reasoner predicts 0 throughout.

Questions for authors
  1. Which reported result most directly shows that the numerical predictor and reasoner provide complementary learning signals, beyond adapting either branch alone? For directional prediction and explanation, how does the design relate to a reasoning-centered model such as PertReasonLM? The continuous-output requirement motivates a numerical predictor, but does not by itself establish the benefit of coupling their updates.

  2. What should an item's reliability weight represent when all cited items share final-answer correctness credit? How does the method distinguish unreliable evidence from valid evidence used incorrectly by the reasoner?

  3. How do the current results show about whether the error thresholds select beneficial updates, rather than simply a different learning schedule? Because correction changes the models and their later acquisitions, end-to-end policy gains might not reflect correction quality.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.