acceptodds
Under review as a conference paper at ICLR 2027

BioCausalX: An Environment for Causal Discovery in Biological Systems

Abstract

Large language model (LLM) agents combine extensive prior knowledge with limited experimental evidence when investigating biological systems. Evaluating and training these agents requires repeatable cycles of intervention, observation, and hypothesis revision. Biological experiments are often slow and costly; simulated environments offer repeatable feedback but also require scientific descriptions and interventions tied to documented mechanisms. We introduce BioCausalX, an environment for causal discovery across 43 executable biological systems and 168 source-supported intervention axes. Variable descriptions and permitted operations are reconciled with publications and database records. Each operation maps to an executable mechanism hidden from the agent, with supporting evidence retained for audit. Across evaluated LLMs, biological context supports recovery, while alignment with the underlying system shapes the benefits of experimentation. Within tested budgets, feedback partly compensates for missing descriptions but does not reliably overcome misleading ones. Larger experimental allowances can remain underused without improving recovery, identifying limitations in both resource allocation and evidence integration. GRPO training improves recovery on held-out systems with and without biological context. Together, these results demonstrate that BioCausalX supports reproducible evaluation and training of scientific agents that combine biological knowledge with experimental feedback.

Primary area:
datasets and benchmarks
Keywords:
Biological discovery, Interactive scientific discovery, Large Language Model

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_cro6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BioCausalX turns 43 published executable biological models into budgeted discovery tasks with biological descriptions and 168 source-supported intervention axes. Agents receive descriptions and simulated observations while equations, exact action bindings, and reference graphs remain hidden. The recovery target is the original model’s directed dynamical dependencies among declared states, excluding self-edges. The paper evaluates classical and LLM inference, context changes, experimental acquisition, and GRPO training on held-out systems. This is a model-grounded environment; its reference graphs describe the implemented models rather than complete biological networks.

Critical strengths
  • The environment combines heterogeneous published mechanisms with a precise model-level target. Linking declared biological states, executable operations, and direct-dependency scoring provides a useful basis for reproducible evaluation. The explicit limits of the graph target and operation mappings help distinguish recovery of the implemented dynamics from broader biological knowledge.
Critical weaknesses
  • The empirical consequence of the added grounding layer is not established relative to prior environments. SciGym already supports discovery using published biological models and perturbations, and BioCausalX builds on the CausalDynamics/causalx codebase. The combination of source-supported descriptions, action mappings, and a common target is concrete, but I need a comparison showing what the added layers change in task construction or agent behavior. An example explains the workflow; it does not establish that comparative contribution.

  • The results do not sufficiently separate structural learning from prior knowledge and submission validity. Biological-context recovery can include prior or source-specific knowledge, and observation-versus-intervention comparisons do not replace a context-only, zero-trajectory control. In the GRPO study, Directed F1 rises from 0.235 to 0.448 with context while validity rises from 50.0% to 97.7%; without context, F1 rises from 0.092 to 0.212 while validity rises from 34.1% to 100%. Invalid-as-zero scoring is a meaningful end-to-end measure, but these changes do not isolate improvement in graph recovery among valid submissions. Both separations are needed before interpreting the results as learning mechanisms from newly acquired evidence.

Questions for authors
  • Can you provide a same-source comparison or ablation that isolates the source-supported operation and mapping layers, and quantify changes in executable action coverage, task behavior, or agent decisions? A complete, inspectable operation record would also clarify mapping correctness : )
  • How does recovery compare with a biological-context-only, zero-trajectory baseline? Can you identify a setting where the acquired trajectories distinguish mechanisms that the supplied prior information alone cannot resolve?
  • I encourage the authors to deepen the existing SciGym [1] comparison with one shared-source example, explaining how the allowed operations, hidden execution mappings, and recovery targets differ. This would help readers see what task behavior the added grounding layer enables beyond an existing biological dry lab.
  • Please report Directed F1 conditional on valid submissions, with per-system coverage at the initial and final GRPO checkpoints, or use a common formatting control. How much of the gain remains after separating submission validity from structural accuracy?

[1] Duan, H., Lu, S., Harrigan, C.F., Desai, N., Lu, J., Koziarski, M., Cotta, L. and Maddison, C., 2026. Measuring Scientific Capabilities of Language Models with a Systems Biology Dry Lab. Advances in Neural Information Processing Systems, 38.

4 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for their careful and constructive feedback. Below, we address each concern and question in turn, using W to denote a critical weakness and Q to denote the corresponding question raised by the reviewer.

W1 The empirical consequence of the added grounding layer is not established relative to prior environments. SciGym already supports discovery using published biological models and perturbations, and BioCausalX builds on the CausalDynamics/causalx codebase. The combination of source-supported descriptions, action mappings, and a common target is concrete, but I need a comparison showing what the added layers change in task construction or agent behavior. An example explains the workflow; it does not establish that comparative contribution.

Q1 Can you provide a same-source comparison or ablation that isolates the source-supported operation and mapping layers, and quantify changes in executable action coverage, task behavior, or agent decisions? A complete, inspectable operation record would also clarify mapping correctness : )

Response:

BioCausalX differs from SciGym in both interface and target: it provides source-grounded node-level context and reviewed operations, supports sequential intervention–observation interaction with hidden executable bindings, and evaluates the complete directed dependency graph. Directed F1 and nSHD therefore measure structural causal recovery directly, consistent with graph-recovery targets in biological network-inference benchmarks [3,4], whereas SciGym’s NTS, RMS, and STE evaluate reaction/kinetic-model reconstruction [1]. The same interface also supports GRPO: the context-trained Qwen policy reaches 0.448 Directed F1 versus 0.212 without context on held-out systems, and 0.579 at step 200. This shows that the full grounded task interface is trainable; the GRPO comparison jointly varies biological descriptions and action naming, while the reviewed operation mappings are audited separately in Appendix D.1.1.

W2 The results do not sufficiently separate structural learning from prior knowledge and submission validity. Biological-context recovery can include prior or source-specific knowledge, and observation-versus-intervention comparisons do not replace a context-only, zero-trajectory control. In the GRPO study, Directed F1 rises from 0.235 to 0.448 with context while validity rises from 50.0% to 97.7%; without context, F1 rises from 0.092 to 0.212 while validity rises from 34.1% to 100%. Invalid-as-zero scoring is a meaningful end-to-end measure, but these changes do not isolate improvement in graph recovery among valid submissions. Both separations are needed before interpreting the results as learning mechanisms from newly acquired evidence.

Response: The complete 0–200 training curve gives the structural check requested:

Training checkpointStrict Directed F1ValidF1 conditional on validity
Step 00.23550.0%0.450
Step 1000.44897.7%0.451
Step 2000.579100.0%0.579

The 0–100 gain is mainly a validity gain: F1 conditional on validity changes by only 0.001. Across the complete 0–200 curve, strict Directed F1 then rises from 0.448 at step 100 to 0.579 at step 200 (+0.131), while validity changes by only 2.3 percentage points. The complete curve in Appendix C, Figure 10, shows validity at 93.2% by step 50 and 97.7% by step 70, with Directed F1 continuing to improve through step 200. This supports an improvement in causal discovery beyond formatting.

AuthorsOpenReview ↗

Response 2/2

Q2 How does recovery compare with a biological-context-only, zero-trajectory baseline? Can you identify a setting where the acquired trajectories distinguish mechanisms that the supplied prior information alone cannot resolve?

Response:

Context-only versus agentic evidence use. The zero-trajectory control uses the same biological context and operation catalog on 43 systems × 3 trials, with every trajectory request rejected:

ModelContext-onlyAgentic interactive
Terra0.7580.765
Claude0.6420.546
Gemini0.5280.551

Context-only recovery is already strong, while the benefit of interaction depends on the model and context. A biological-network review notes that current interaction maps remain incomplete and noisy, and that interactions vary across tissues and conditions [2]. We therefore use Anonymous as a controlled stress test of missing or misaligned prior information. In this condition, biological identities and mechanisms are removed but trajectories remain; Terra improves from 0.160 observational to 0.355 interactive Directed F1 (+0.195), with a +0.117 gain on PC12. The interactive gain shows that trajectories provide additional structural information when the supplied context is incomplete. Agentic RL further raises Qwen's interactive Directed F1 to 0.579 at step 200, above its context-only zero-trajectory baseline of 0.464 (Appendix D.1.1, the context-only versus agentic interactive recovery table).

Q3 I encourage the authors to deepen the existing SciGym [1] comparison with one shared-source example, explaining how the allowed operations, hidden execution mappings, and recovery targets differ. This would help readers see what task behavior the added grounding layer enables beyond an existing biological dry lab.

Response:

SciGym provides a partial SBML model and prompt-level biological framing, then lets the agent write code and request initial-concentration perturbations before submitting a reaction/kinetic model scored by NTS, RMS, and STE. BioCausalX provides source-grounded node-level context and reviewed operations, supports sequential intervention--observation interaction with hidden executable bindings, and evaluates a complete directed graph over all declared states. This state-level causal target directly measures which biological states influence one another.

Q4 Please report Directed F1 conditional on valid submissions, with per-system coverage at the initial and final GRPO checkpoints, or use a common formatting control. How much of the gain remains after separating submission validity from structural accuracy?

Response:

The complete 0–200 context-policy training curve rises from 0.448 Directed F1 at step 100 to 0.579 at step 200 (+0.131), while validity changes from 97.7% to 100.0%. The gain after validity nears saturation shows structural improvement beyond submission format.

[1] Duan, H., Lu, S., Harrigan, C.F., Desai, N., Lu, J., Koziarski, M., Cotta, L. and Maddison, C., 2025. Measuring Scientific Capabilities of Language Models with a Systems Biology Dry Lab. Advances in Neural Information Processing Systems, 38.

[2] Barabási, A.-L., Gulbahce, N. and Loscalzo, J., 2011. Network medicine: a network-based approach to human disease. Nature Reviews Genetics, 12, 56–68.

[3] Marbach, D., Costello, J.C., Küffner, R., Vega, N.M., Prill, R.J., Camacho, D.M., Allison, K.R., DREAM5 Consortium, Kellis, M., Collins, J.J. and Stolovitzky, G., 2012. Wisdom of crowds for robust gene network inference. Nature Methods, 9, 796–804.

[4] Chevalley, M., Roohani, Y.H., Mehrjou, A., Leskovec, J. and Schwab, P., 2025. A Large-Scale Benchmark for Network Inference from Single-Cell Perturbation Data. Communications Biology, 8, 412.

In light of these responses, we hope we have addressed your concerns, and hope you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

AuthorsOpenReview ↗

Kind reminder

Thank you again for your thoughtful comments. With less than three days remaining in the discussion period, we would greatly appreciate any further feedback or concerns you may have. If our clarifications have addressed your concerns, we would sincerely appreciate your consideration of raising your score.

With the discussion deadline approaching, we wanted to briefly follow up. If any concerns remain or if there is anything we can clarify, we would be grateful for the chance to respond while the discussion is still open. If our clarifications resolve your remaining concerns, we hope you might consider raising your overall score.

Thank you again for your time and feedback.

Reviewer_Q9GtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BioCausalX is an interactive evaluation environment built from 43 published biological models and 168 intervention axes. Agents receive biological descriptions, select experiments, observe simulated trajectories, and submit directed graphs scored against dependencies extracted from the models’ ordinary differential equations. The paper compares eight numerical methods and four language-model agents, studies context, tools, output handling, and experimental budgets, and trains Qwen3.5-35B-A3B on 32 systems with evaluation on 11 held-out systems. The task is reconstruction of existing model structure, rather than discovery of unknown biological mechanisms.

Critical strengths

Known reference graphs make structural recovery objectively measurable. The controlled descriptions, interventions, and budgets also allow reproducible comparisons of how agents combine prior knowledge with experimental evidence.

Critical weaknesses

1.The main result is recovery of dependencies encoded in published models, not discovery of unknown biology. The authors acknowledge this, and a simulation benchmark does not need wet-lab experiments to be useful. My concern is what a better graph score establishes beyond this reconstruction task. D.5 provides limited support: it observes the first 10% of simulated responses before predicting later affected nodes, and correlations between structural and response F1 range from −0.014 to 0.275. Recovered graphs even outperform the reference graph under this response metric, which the authors explain through greater selectivity from sparsity. That explanation is reasonable, but it also shows that better structural recovery and better response prediction are not interchangeable. The biological interpretation of the scores should remain correspondingly narrow.

2.The training results do not adequately separate better graph recovery from more reliable submission. For context-trained Qwen, Directed F1 increases from 0.235 to 0.448 while valid submissions rise from 50% to 97.7%; invalid outputs receive zero F1. Some systems improve even when submission validity stays at 100%, so the gains cannot all be dismissed as formatting. Nevertheless, the aggregate results do not establish how much improvement comes from better experimental inference rather than successful output handling. This distinction matters for the claim that training improves structural recovery. (§5.4; Appendix C.)

3.The paper reports source checks and numerical validation, but does not present a complete worked example connecting a documented manipulation to the exact model changes and intervention strength. It also does not quantify review coverage or resulting corrections. Because these mappings are a distinctive part of the environment, the available description is insufficient to assess their biological fidelity. (Appendix A.2.)

4.The 32/11 split excludes evaluation systems from parameter updates, but does not establish separation between underlying mechanism families. Public source models may also have been encountered during pretraining, a possibility the authors acknowledge. The results support transfer to systems excluded from this optimization, not demonstrated generalization to previously unseen biological mechanisms. (Appendices C.1 and E.5.)

Questions for authors

1.SciGym already includes simulated interventions, topology scoring, context variations, and knockout experiments. Which reported finding best demonstrates BioCausalX’s additional evaluation value, beyond providing another collection of reconstruction tasks?

2.Can the existing pre- and post-training Qwen outputs be evaluated under the same format-recovery procedure, with submission validity reported separately, to clarify how much improvement concerns graph content?

3.Could you provide one complete intervention record, including its source location, exact model modifications, and strength justification, together with the coverage and correction statistics of the completed review?

4.How were “source families” defined for the held-out split, and which evaluation systems share pathways, reaction mechanisms, or equation forms with training systems?

5.Could you report absolute response F1 and per-system results for D.5? Given that this endpoint can favor graphs missing true edges, what downstream interpretation of higher structural F1 remains supported?

4 comments
AuthorsOpenReview ↗

Response 1/2

We are truly grateful for the time you have taken to review our paper and your insightful review. Below, we address each concern and question in turn, using W to denote a critical weakness and Q to denote the corresponding question raised by the reviewer.

W1 The main result is recovery of dependencies encoded in published models, not discovery of unknown biology. The authors acknowledge this, and a simulation benchmark does not need wet-lab experiments to be useful. My concern is what a better graph score establishes beyond this reconstruction task. D.5 provides limited support: it observes the first 10% of simulated responses before predicting later affected nodes, and correlations between structural and response F1 range from −0.014 to 0.275. Recovered graphs even outperform the reference graph under this response metric, which the authors explain through greater selectivity from sparsity. That explanation is reasonable, but it also shows that better structural recovery and better response prediction are not interchangeable. The biological interpretation of the scores should remain correspondingly narrow.

Response:

We use Directed F1 and nSHD to evaluate recovery of the frozen model graph; both are standard metrics in causal discovery. We further study conditional response propagation in Appendix D.5. We thank the reviewer for this important comment and have narrowed the interpretation accordingly: our claims concern model-level recovery of direct dependencies in the frozen ODE system, while D.5 is a separate diagnostic and does not imply wet-lab mechanism discovery, response prediction, or effect-size prediction.

W2 The training results do not adequately separate better graph recovery from more reliable submission. For context-trained Qwen, Directed F1 increases from 0.235 to 0.448 while valid submissions rise from 50% to 97.7%; invalid outputs receive zero F1. Some systems improve even when submission validity stays at 100%, so the gains cannot all be dismissed as formatting. Nevertheless, the aggregate results do not establish how much improvement comes from better experimental inference rather than successful output handling. This distinction matters for the claim that training improves structural recovery. (§5.4; Appendix C.)

Q2 Can the existing pre- and post-training Qwen outputs be evaluated under the same format-recovery procedure, with submission validity reported separately, to clarify how much improvement concerns graph content?

Response:

The original 0–100 evaluation separates strict recovery from validity:

Training checkpointStrict Directed F1ValidF1 conditional on validity
Step 00.23550.0%0.450
Step 1000.44897.7%0.451
Step 2000.579100.0%0.579

The strict 0–100 gain therefore combines submission reliability with graph recovery: F1 conditional on validity changes by only 0.001. Across the complete 0–200 curve, strict Directed F1 increases from 0.448 at step 100 to 0.579 at step 200 (+0.131), while validity changes from 97.7% to 100%; at step 200, validity-conditional F1 is therefore also 0.579. The extended curve is shown in Appendix C, Figure 10.

W3 The paper reports source checks and numerical validation, but does not present a complete worked example connecting a documented manipulation to the exact model changes and intervention strength. It also does not quantify review coverage or resulting corrections. Because these mappings are a distinctive part of the environment, the available description is insufficient to assess their biological fidelity. (Appendix A.2.)

Q3 Could you provide one complete intervention record, including its source location, exact model modifications, and strength justification, together with the coverage and correction statistics of the completed review?

Response:

For PC12 (BIOMD0000000033), the five admitted mappings are:

OperationPrivate model modification
Omit EGFEGF 10,002,000→0 at t=0; EGFBindingReaction
Omit NGFNGF 456,000→0 at t=0; NGFBindingReaction
LY294002PI3KbyEGFRActivationReaction and PI3KbyRasActivationReaction→0
DN Rap1BRafByRap1ActivationReaction→0
DN RasRaf1ByRasActivationReaction and PI3KbyRasActivationReaction→0

The frozen audit has 5/5 admitted axes, 45/45 evidence claims, 5/5 compile-and-solve checks, 5/5 changed trajectories, native ODE PASS, symbolic graph-audit PASS (N=28, E=62, Q_ground=0.96875), and response non-redundancy PASS; two candidates were withheld. The complete record is now provided in Appendix D.1.1. The frozen BioCausalX contains 43 systems and 168 admitted axes.

AuthorsOpenReview ↗

Response 2/2

W4 The 32/11 split excludes evaluation systems from parameter updates, but does not establish separation between underlying mechanism families. Public source models may also have been encountered during pretraining, a possibility the authors acknowledge. The results support transfer to systems excluded from this optimization, not demonstrated generalization to previously unseen biological mechanisms. (Appendices C.1 and E.5.)

Q4 How were “source families” defined for the held-out split, and which evaluation systems share pathways, reaction mechanisms, or equation forms with training systems?

Response:

We split 32 training and 11 held-out systems by canonical source-model/publication lineage; the resulting evaluation measures transfer across held-out source systems rather than mechanism-disjoint generalization.

Q1 SciGym already includes simulated interventions, topology scoring, context variations, and knockout experiments. Which reported finding best demonstrates BioCausalX’s additional evaluation value, beyond providing another collection of reconstruction tasks?

Response:

Our design differs substantially from SciGym. BioCausalX provides source-grounded node-level context and reviewed operations, giving the task a more concrete context design than interventions without biological context. Its Directed F1 and nSHD metrics evaluate structural causal recovery directly, consistent with graph-recovery targets in biological network-inference benchmarks [3,4], whereas SciGym evaluates reaction/kinetic-model reconstruction. The same interface also supports agentic RL, as shown in Section 5.4.

Q5 Could you report absolute response F1 and per-system results for D.5? Given that this endpoint can favor graphs missing true edges, what downstream interpretation of higher structural F1 remains supported?

Response:

ModelObservationalAgent-selected
GPT-5.6 Terra0.4320.372
Claude Sonnet 4.60.5060.434
DeepSeek V4.1 Flash0.4470.396
Gemini 2.5 Flash Thinking0.4830.448

D.5 observes the first 10% of a held-out response and predicts the later affected-node footprint by graph reachability. The table reports absolute response-propagation F1; complete per-system results are provided in the appendix table “Per-system conditional response propagation F1” (86 system×mode cells; 73 have eligible held-out actions).

We hope that our responses have adequately addressed the concerns, and we would appreciate it if you could reconsider your score in light of these clarifications.

AuthorsOpenReview ↗

Kind reminder

Thank you again for your thoughtful comments. With less than three days remaining in the discussion period, we would greatly appreciate any further feedback or concerns you may have. If our clarifications have addressed your concerns, we would sincerely appreciate your consideration of raising your score.

As the discussion period draws to a close, we wanted to follow up on our response to your review. If you have any remaining questions or concerns, we would appreciate the opportunity to address them before the deadline. If you feel that our responses have addressed your concerns, we would be grateful if you would consider raising your overall score in light of the discussion.

Thank you for your time and for reviewing our submission.

Reviewer_ru7jOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

BioCausalX provides 43 executable biological systems with equation-derived reference graphs, literature-grounded interventions, and controlled interaction budgets. The paper compares numerical methods and LLM agents, studies context and tool effects, and reports reinforcement-learning results on 11 systems excluded from optimization.

Critical strengths
  1. The environment connects mechanistic models, documented interventions, and explicit structural targets. This makes scientific-agent experiments more interpretable and reproducible than evaluations based only on open-ended answers.

  2. The experiments examine several distinct failure modes through context controls, fixed-evidence tool comparisons, acquisition policies, and held-out training evaluation. These analyses provide useful evidence about when biological context and experimental interaction help.

Critical weaknesses
  1. The contribution of experimental evidence remains insufficiently separated from prior biological knowledge. Deidentification retains biological descriptions, and no description-only baseline is reported. Moreover, the Permuted condition changes both descriptions and action references, so its degradation cannot isolate failure to revise misleading beliefs.

  2. The RL gains combine structural recovery with improved submission validity. With context, validity rises from 50.0% to 97.7% while Directed F1 rises from 0.235 to 0.448 (Table 8); invalid outputs receive zero. One training run per condition and no matched format-repair comparison leave the gain in causal inference itself uncertain.

Questions for authors
  1. How does a description-only agent perform without trajectories, and does the Permuted effect persist when action references remain correctly grounded?
5 comments

We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! Below, we address each concern and question in turn, using W to denote a critical weakness and Q to denote the corresponding question raised by the reviewer.

W1 The contribution of experimental evidence remains insufficiently separated from prior biological knowledge. Deidentification retains biological descriptions, and no description-only baseline is reported. Moreover, the Permuted condition changes both descriptions and action references, so its degradation cannot isolate failure to revise misleading beliefs.

Q1 How does a description-only agent perform without trajectories, and does the Permuted effect persist when action references remain correctly grounded?

Response:

ModelContext-onlyAgentic observationalAgentic interactive
Terra0.7580.7460.765
Claude0.6420.5300.546
Gemini0.5280.6060.551

The context-only agents still achieve 0.758, 0.642, and 0.528 Directed F1 for Terra, Claude, and Gemini, respectively, showing that the supplied biological context alone supports causal discovery.

For the permutation experiments, only the biological prose is permuted; node IDs, operation targets, and action references remain correct. On Terra, Directed F1 decreases from 0.733 with the correct prose to 0.296 with permuted prose. The permutation effect therefore persists when action references are fixed.

W2 The RL gains combine structural recovery with improved submission validity. With context, validity rises from 50.0% to 97.7% while Directed F1 rises from 0.235 to 0.448 (Table 8); invalid outputs receive zero. One training run per condition and no matched format-repair comparison leave the gain in causal inference itself uncertain.

Response:

The complete 0–200 training curve gives the requested structural check:

Training checkpointStrict Directed F1ValidF1 conditional on validity
Step 00.23550.0%0.450
Step 1000.44897.7%0.451
Step 2000.579100.0%0.579

The 0–100 gain is mainly a validity gain: F1 conditional on validity changes by only 0.001. Across the complete 0–200 curve, Directed F1 rises from 0.448 at step 100 to 0.579 at step 200 (+0.131), while validity changes from 97.7% to 100.0%. This supports an improvement in causal discovery beyond formatting. The complete curve is shown in Appendix C, Figure 10.

Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions.

AuthorsOpenReview ↗

Kind reminder

Thank you again for your thoughtful comments. With less than three days remaining in the discussion period, we would greatly appreciate any further feedback or concerns you may have.

Thank you for the detailed response and additional experiments. The context-only baseline and clarified permutation setup help address my first concern, while the extended RL training results provide useful evidence of improvements beyond output validity. Overall, my concerns have been largely addressed, and I will maintain my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Feedback and Rebuttal Evidence

Dear Area Chair,

Thank you for overseeing the review process of BioCausalX. To facilitate your final assessment, we summarize the complementary strengths recognized by the reviewers and the additional evidence provided during the rebuttal to address their key concerns.

The reviewers recognized several complementary strengths of our work. ru7j highlighted its interpretability and the experiments examining when context and interaction help. Q9Gt appreciated how reference graphs and controlled experimental budgets enable objective, reproducible evaluation. cro6 valued the integration of diverse published biological models, executable interventions, and well-defined causal discovery targets.

During the rebuttal, we provided additional evidence addressing two recurring concerns:

  • RL gains beyond submission validity (ru7j, Q9Gt, cro6). We clarified that early improvements were primarily driven by submission validity. However, after validity reached 97.7% in the context-trained run, continued training improved valid-only Directed F1 from 0.451 to 0.579, demonstrating improvements in causal discovery beyond formatting (Figure 10, Tables 8 and 10).

  • Intervention fidelity and distinction from SciGym (Q9Gt, cro6). We provided an auditable PC12 example with explicit intervention mappings (Appendix D.1.1, Table 15). We also clarified that BioCausalX differs from SciGym in both interface and evaluation target: it provides source-grounded node-level context and reviewed operations, enables sequential intervention–observation interaction through hidden executable bindings, and evaluates complete causal graph recovery rather than reaction and kinetic-model reconstruction.

We are encouraged that ru7j explicitly acknowledged that their concerns were “largely addressed” and maintained their positive rating, citing the context-only baseline, permutation control, and extended RL results. Although we repeatedly invited Q9Gt and cro6 to provide further feedback during the author discussion, we did not receive any responses. We believe that our additional experiments and detailed clarifications have substantively addressed their original concerns.

We respectfully ask the AC to consider the full discussion record, including the additional experimental evidence, detailed clarifications, and reviewer feedback, in the final assessment of BioCausalX.

Thank you again for your time and careful consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.