acceptodds
Under review as a conference paper at ICLR 2027

GGBound: Genome-Grounded Reasoning for Life-Boundary Phenotype Prediction

Abstract

Predicting microbial life-boundary phenotypes from genomic sequence offers a scalable alternative to costly experimental screening, but remains challenging under sparse genomic observations and diverse phenotype types. Existing approaches either rely on predefined genomic features or perform direct genome-to-language prediction, leaving complementary predictive signals and intermediate phenotype proposals underused. In this paper, we formulate the task as *proposal-conditioned phenotype refinement*, where genomic evidence is used to refine the proposals. We characterize proposal-conditioned genomic utility, showing that effective genomic representations should retain phenotype-relevant information while suppressing irrelevant variation. Guided by this analysis, we introduce GGBound, which learns phenotype-specific genomic representations and grounds them in a language decoder through phenotype-oriented reasoning supervision, yielding interpretable rationales alongside predictions. We additionally introduce Sufficiency-to-Decision GRPO to strengthen the use of genomic evidence beyond auxiliary proposals during refinement. We further construct a strain-centric benchmark spanning 17 microbial life-boundary phenotypes from IJSEM strains and BacDive annotations. Experiments show that GGBound consistently outperforms genome-based baselines across phenotype categories and even surpasses frontier LLMs using strain names to leverage organism-specific knowledge on most phenotype groups.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Microbial phenotype prediction, Genome-grounded LLM reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NXUmOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a novel method to predict life boundary phenotypes - i.e. survival temperatures, survival pH ranges, oxygen tolerance etc. - from genomic data directly. Their method is GGBound - which takes a genomic sequence in, and outputs a phenotypic prediction: first, a genomic encoder (Evo2) is adapted to produce phenotype predictions directly, while also producing a latent representation that continues phenotype-relevant information. That latent representation is then adapted, via a projector, into the latent space of an LLM, which is then trained via SFT to create its own phenotypic predictions after reasoning about the phenotype. Next, these two predictions (the one from Evo2, and the one from the trained LLM) - called auxiliary predictions - are combined with the aforementioned genomic representation and all passed into a final LLM step ("Proposal-Conditioned Genomic Reasoning"), where the model reasons over all of them and outputs a final phenotype prediction. This final step is trained via GRPO, w/ added objectives to encourage reasoning beyond the provided auxiliary predictions. This entire pipeline combined substantially outperforms frontier LLMs and simpler baselines, setting a new SOTA for life-boundary phenotype prediction.

Critical strengths
  • The paper's extremely complex pipeline is extremely well-justified. Each part of it is ablated, and it is thoroughly shown that the entire thing combined outperforms its component parts. Controls are offered in Table 2, w/ a detailed follow up in Table 5 showing how their method helps resolve conflicting predictions. Furthermore, the pipeline outperforms the simple baseline of frontier models.
  • The paper introduces an ingenious new GRPO mechanism to encourage the model to genuinely use the information provided in the encoding and not just the auxiliary predictions - they measure, holding the aux predictions constant, whether changing the genomic encoding to a plausible distractor, on average, makes the overall rollout less likely. If so, the model was relying on the genomic encoding - and they increase the reward given to the rollout. This is a very important innovation - moving the ability to reconcile conflicting aux predictions from 63.4 -> 72.3, and boosts F1 accuracy from 0.645 on the Metabolism category of phenotype (the gains in the other types are smaller, but still in the right direction).
Critical weaknesses

None

Questions for authors
  • You use Qwen3.5 9B and Qwen3.5 27B for the generation of the grounding reasoning traces, as I understand it. These models - especially Qwen3.5 9B - aren't particularly strong. Why not use a more powerful OSS model (GLM-5.2 or GLM-5.3, Kimi K3, etc.)? Cost reasons? Or does using a Qwen model help Qwen performance (as Qwen3.5-4B is your LLM decoder)?
  • You state "The refined claims are subjected to deterministic validation before being included in the training set. We remove claims with invalid or missing evidence references, unsupported biological assertions, supervision leakage, or conclusions inferred solely from missing annotations, ensuring that each retained statement remains traceable to the provided protein evidence." Is there an easy way to do this programmatically? Rejecting "unsupported biological assertions" seems difficult - but I don't have a ton of domain knowledge in the realm of phenotype prediction, so perhaps there is. If you use simpler heuristics, adding a section to the appendix on how 'unsupported biological assertions' are rejected would be helpful. The rest of the rejection criteria make more sense - as they can be programmatically checked. If you mean 'unsupported biological assertions' in the sense of claims without a support_level, state this explicitly - and emphasize you're applying format checking, not checking underlying biological validity.
  • Table 2 lacks p-values or confidence intervals - while it is clear your method is the best across phenotypes, whether or not the difference is statistically significant with runner-ups (i.e. E2-M or G/P-Qw.) isn't listed. Similarly, the grids in FIgure 5 could use annotations to indicate whether differences are statistically significant - as sometimes the gaps are quite close! Because your method clearly dominates all others in most cases, this isn't a huge concern - but it would be a nice addition, and one you could likely cheaply compute.
5 comments

We sincerely thank you for your valuable and positive comments. We hope our response below could address your concerns.

Q1. The choice of grounding reasoning traces generation models

A: We thank the reviewer for this question. Our choice was primarily driven by limited computational resources, favoring efficient, locally deployed models. We use Qwen3.5-9B for initial reasoning generation and Qwen3.5-27B for evidence-grounded refinement. Regarding potential same-family bias, prior work [1] suggests that stronger teachers do not necessarily yield better reasoning distillation, as teacher–student compatibility also matters, potentially favoring same-family models. We have acknowledged this potential influence and clarified our model choice in the revised Appendix B.3.

Q2: Clarification on Biological Claim Validation and Quality Control

A: We thank the reviewer for this helpful comment. We agree that automatically verifying biological claims is challenging, and our original wording was imprecise. Rather than relying on deterministic biological validation, we use Qwen3.5-27B to assess claims against the provided protein annotations, revising or rejecting those lacking sufficient evidence.

For example, a protein annotation in Flavobacterium pectinovorum describes “Catalyzes the anaerobic formation of alpha-ketobutyrate and ammonia from threonine in a two-step reaction.” This supports potential amino-acid nitrogen utilization, but not growth on threonine as the sole nitrogen source. Our refinement process is designed to retain the former inference while rejecting the latter unsupported conclusion, relying on Qwen3.5-27B's ability to assess evidential support rather than deterministic biological rules.

Subsequent deterministic checks only validate output structure, evidence references, and predefined failure patterns. Importantly, support_level is model-generated, not independently verified. We additionally conducted a GPT-based review of 136 trajectories (8 strains per task across 17 tasks), all of which passed. Nevertheless, these procedures do not guarantee biological correctness. We have clarified the validation process and its limitations in Appendix B.3.

Q3. Statistical Significance of Performance Improvements

A: We thank the reviewer for this valuable suggestion. We conducted 20,000 strain-level bootstrap resamples to estimate 95% confidence intervals (CIs) for performance differences, where Δ denotes the performance difference between GGBound and the corresponding baseline (GGBound − baseline).

As shown in the table below, for Table 2, GGBound's improvements have confidence intervals excluding zero on 12/17, 12/17, and 14/17 phenotypes compared with E2-M, G-Qwen, and P-Qwen, respectively. Furthermore, for Figure 5, all 15 comparisons between our full method and the three ablation variants have confidence intervals excluding zero, further supporting the contribution of each component. We have added these results to the Appendix D.1 and D.6 of revised manuscript.

GroupPhenotypeMetricΔ vs E2-MΔ vs G-QwenΔ vs P-Qwen
Phys. BoundaryTemperatureRMSE ↓\downarrow−0.197 [−0.418, -0.002]−1.727 [−2.320, −1.150]−1.867 [−2.913, −1.041]
pHRMSE ↓\downarrow−0.039 [−0.108, -0.010]−0.186 [−0.266, −0.106]−0.112 [−0.178, −0.047]
SalinityRMSE ↓\downarrow−0.053 [−0.092, −0.023]−0.805 [−1.071, −0.522]−0.579 [−0.831, −0.321]
Phys. OptimumTemperatureRMSE ↓\downarrow−0.449 [−0.672, −0.239]−1.588 [−2.882, −0.504]−1.370 [−2.692, −0.274]
pHRMSE ↓\downarrow−0.035 [−0.087, -0.008]−0.327 [−0.414, −0.242]−0.301 [−0.387, −0.217]
SalinityRMSE ↓\downarrow−0.084 [-0.108,-0.01]−0.674 [-0.87,-0.504]−0.461 [-0.68,-0.24]
MetabolismCarbon SourceF1 ↑\uparrow+0.008 [0.001, 0.017]+0.053 [0.028, 0.079]+0.168 [0.116, 0.221]
Electron AcceptorF1 ↑\uparrow+0.348 [0.120, 0.614]+0.228 [0.094, 0.378]+0.248 [0.114, 0.388]
Electron DonorF1 ↑\uparrow+0.351 [0.073, 0.631]+0.015 [0.00,0.056]+0.075 [0.017, 0.138]
Nitrogen SourceF1 ↑\uparrow+0.226 [0.087, 0.352]+0.339 [0.133, 0.555]+0.290 [0.080, 0.515]
CategoricalOxygen ToleranceAcc. (%) ↑\uparrow+1.04 [0.00, 2.59]+8.29 [3.11, 13.47]+7.77 [3.63, 12.44]
Energy TypeAcc. (%) ↑\uparrow+7.69 [0.00, 19.23]+3.85 [0.00, 11.54]+3.85 [0.00, 11.54]
MorphologyCell ShapeAcc. (%) ↑\uparrow+8.74 [0.97, 16.50]+11.65 [3.88, 19.42]+6.80 [0.00, 14.56]
MotilityAcc. (%) ↑\uparrow+1.50 [0.00, 3.50]+9.50 [4.50, 14.50]+27.00 [18.50, 35.00]
FlagellaAcc. (%) ↑\uparrow+4.60 [1.15, 9.20]+4.60 [0.00, 11.49]+1.15 [0.00, 3.45]
Spore FormationAcc. (%) ↑\uparrow+1.01 [0.00, 2.51]+2.01 [0.00, 4.52]+42.71 [35.18, 50.25]
Gram stainAcc. (%) ↑\uparrow+0.50 [0.00, 1.50]+1.00 [0.00, 2.50]+27.50 [21.50, 34.00]
GroupImprovement metricΔ vs GRPOΔ vs +UtilityΔ vs +CE
Physiological BoundariesNRMSE ↓\downarrow-0.006 [-0.011, -0.001]-0.005 [-0.011, -0.001]-0.004 [-0.01, -0.001]
Physiological OptimaNRMSE ↓\downarrow-0.009 [-0.015, -0.002]-0.005 [-0.01, -0.001]-0.008 [-0.013, -0.003]
MetabolismF1 ↑\uparrow+0.179 [0.105, 0.255]+0.183 [0.102, 0.269]+0.162 [0.093, 0.236]
Categorical PhysiologyAcc (%) ↑\uparrow+4.85 [0.52, 11.03]+1.04 [0.12, 2.36]+2.70 [0.25, 4.74]
MorphologyAcc (%) ↑\uparrow+3.65 [1.98, 5.42]+2.38 [0.78, 4.08]+3.22 [1.26, 5.30]

[1] Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment, ACL 2026

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

Thank you for the detailed response! As for Q1, I completely understand limited computational resources - that resolves my question!

As for Q2, thank you for clarifying this in the appendix. This resolves my concerns there!

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

This paper proposes GGBound, which learns phenotype-specific genomic representations and grounds them in a language decoder through phenotype-oriented reasoning supervision.GGBound is Validated on a strain-disjoint benchmark of 2,080 records from 745 strains, with crops fixed across methods.

Critical strengths

The token-level genomic utility term is a critical strength. The contrast is computed against proposal-matched alternatives, which hold both auxiliary predictions fixed while varying the underlying genome, so the signal isolates the decoder’s dependence on genomic evidence from its dependence on the proposals.

Critical weaknesses

Setting the framing aside, the paper reads as an integration of standard components around one new ingredient and a new dataset. Task-token conditioning is hard prefix tuning, the DNA reconstruction loss is a familiar anti-forgetting regularizer, and the projector-plus-soft-token interface follows prior genome-to-language work. The one component that is new is the token-level genomic utility, and the dataset is a substantial piece of curation.

The claim I find hardest to support is “proposal-conditioned genomic reasoning” as a task formulation. Placing two models’ predictions in the prompt and asking an LLM to arbitrate is, in substance, ensembling with an LLM combiner. More importantly, the paper’s own P-Qwen row suggests the formulation does not carry weight on its own.

I would add that P-Qwen may also be disadvantaged by construction, since the grounding prompt in Appendix C.3 contains no proposal fields and the model is therefore evaluated under a prompt format it never saw during training. A fairer comparison would train an SFT model with proposals in the prompt and then compare it against the same model after S2D-GRPO, which would separate the benefit of RL from the benefit of format exposure.

Questions for authors

In the sequence-only condition, many entries are identical across models that are otherwise unrelated. In Tables 10 and 11, DeepSeek-V4 and Kimi-K3 agree to three decimals on all nine physiological-boundary metrics. Could the authors confirm whether these reflect models emitting a single constant label for the whole task? What fraction of outputs per model per task were unique labels, and what fraction were malformed and mapped to a default?

5 comments

We sincerely thank you for your valuable comments. We hope our response below could address your concerns.

Q1. What is the methodological novelty beyond integrating existing components, token-level genomic utility, and dataset curation?

A: We thank the reviewer for this question. We agree that task tokens, DNA reconstruction, and soft-token projection are established techniques. Our methodological contribution lies in studying diverse microbial phenotype prediction through a proposal-conditioned information bottleneck, deriving two complementary optimization principles: phenotype-relevant genomic information retention and effective genomic evidence utilization.

  1. Phenotype-specific genomic specialization. Unlike prior approaches that directly apply a prediction head to general-purpose genomic representations, our formulation emphasizes retaining phenotype-relevant information while suppressing irrelevant variation. Table 2 shows that encoder adaptation (E2-M) outperforms frozen representations (E2-P) on 15 of 17 phenotypes. Table 3 further shows that our specialization improves categorical accuracy from 62.7% to 71.9% and morphology accuracy from 73.9% to 78.1%, while boundary prediction remains comparable, highlighting different information requirements across phenotypes.

  2. Genome-grounded proposal refinement. Our theory motivates effective utilization of retained genomic information beyond auxiliary proposals (Eq. 2 and Theorem 1). We operationalize this principle through S2D-GRPO, which introduces token-level genomic utility to guide policy optimization. As shown in Table 5, when exactly one proposal is correct, accuracy improves from 44.6% (P-Qwen) and 63.4% (GRPO) to 84.2% (GGBound).

Together, these contributions unify genomic information retention and utilization, with established components serving as implementation tools rather than the source of our methodological novelty.

Q2. Does proposal-conditioned genomic reasoning go beyond LLM-based ensembling, given the limited effectiveness of P-Qwen?

A: We thank the reviewer for this insightful comment. We agree that simply adding auxiliary proposals does not guarantee improvement, as shown by P-Qwen. Our task formulation instead emphasizes phenotype-relevant predictive information beyond auxiliary proposals (Eq. 1), motivating S2D-GRPO to learn to verify and refine proposals using genomic evidence, rather than merely aggregate them. As shown in Table 5, when exactly one proposal is correct, S2D-GRPO achieves 84.2% accuracy, compared with 44.6% for P-Qwen and 63.4% for standard GRPO. These results highlight the benefit of genome-grounded verification and refinement over simple proposal aggregation.

Q3. Is the P-Qwen comparison fair, and are S2D-GRPO gains independent of proposal-format exposure?

A: We thank the reviewer for this valuable suggestion. We agree that P-Qwen does not fully control for proposal-format exposure. We therefore introduce a proposal-conditioned SFT baseline (P-SFT) and further apply S2D-GRPO to the same checkpoint (P-SFT-S2D). As shown in the Table below, while P-SFT alone shows mixed improvements, P-SFT-S2D improves performance on 16 out of 17 phenotypes, achieving results comparable to GGBound. These results demonstrate that our gains cannot be attributed merely to proposal-format exposure, further validating the effectiveness of S2D-GRPO. We have include these additional results to the Appendix D.9 of our revised manuscript.

GroupPhenotypeMetricP-QwenP-SFTP-SFT-S2DGGBound
Phys. BoundaryTemperatureRMSE ↓\downarrow9.599.487.757.72
pHRMSE ↓\downarrow1.351.341.241.23
SalinityRMSE ↓\downarrow3.873.893.263.29
Phys. OptimumTemperatureRMSE ↓\downarrow6.376.045.035.00
pHRMSE ↓\downarrow1.131.010.830.83
SalinityRMSE ↓\downarrow3.033.522.562.57
MetabolismCarbon SourceF1 ↑\uparrow0.460.570.620.63
Electron AcceptorF1 ↑\uparrow0.630.630.880.87
Electron DonorF1 ↑\uparrow0.820.640.860.89
Nitrogen SourceF1 ↑\uparrow0.540.760.830.83
CategoricalOxygen ToleranceAcc.(%) ↑\uparrow60.160.566.867.8
Energy TypeAcc.(%) ↑\uparrow80.863.384.684.6
MorphologyCell ShapeAcc.(%) ↑\uparrow47.642.154.354.4
MotilityAcc.(%) ↑\uparrow50.574.577.077.5
FlagellaAcc.(%) ↑\uparrow85.170.088.586.2
Spore FormationAcc.(%) ↑\uparrow49.888.591.092.5
Gram StainAcc.(%) ↑\uparrow69.096.096.096.5

Q4. Identical Predictions Across LLMs

A: We thank the reviewer for this careful observation. We inspected the original generation logs and saved predictions, confirming prediction collapse in the sequence-only setting: DeepSeek-V4 and Kimi-K3 produced nearly constant predictions (0.56–0.91% unique prediction ratios). We verified that this behavior originated from the model outputs rather than parsing or fallback mechanisms. We have clarified these findings in Appendix D.1.

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

AuthorsOpenReview ↗

Gentle Follow-up on the Rebuttal

Dear Reviewer SGVV,

We sincerely appreciate your thoughtful review, valuable suggestions, and the time and effort you have devoted to evaluating our submission. We have provided additional experiments and clarifications in response to your valuable feedback.

We hope our responses and additional experimental results could address your concerns. We would be very grateful if you could share any further feedback or remaining concerns at your convenience. We would be happy to provide additional clarifications if needed.

Best regards,

Paper 15002 Authors

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

The paper targets predicting microbial life-boundary phenotypes from genomic sequences. It formulates the task as proposal-conditioned phenotype refinement, where genomic evidence is used to refine the proposals. It introduces the GGBound and Sufficiency-to-Decision GRPO methods. It also constructs a strain-centric benchmark. Experiments show that GGBound consistently outperforms genome-based baselines and LLMs.

Critical strengths
  1. Novel methodology: It introduces the GGBound and Sufficiency-to-Decision GRPO methods. It also constructs a strain-centric benchmark.

  2. Solid experiments: The paper introduces one benchmark and shows better performance compared with genomic baselines and LLMs.

Critical weaknesses
  1. Limit the genomic model to Evo 2: No evidence of how much benefit Evo 2 brings compared to other genomic foundation models.

  2. A little hard to understand the theory: It is a little hard to understand what is the meaning of the theory shown in the paper for people without much math background.

Questions for authors
  1. If we change Evo 2 to other models, like DNABERT-2, Nucleotide Transformer, or GenomeOcean [1], how will the whole performance change? [1] GenomeOcean: An Efficient Genome Foundation Model Trained on Large-Scale Metagenomic Assemblies

  2. Could you please provide some plain and easy-to-understand language for the theory shown in the paper?

4 comments

We sincerely thank you for your valuable comments. We hope our response below could address your concerns.

Q1. How much benefit Evo 2 brings compared to other genomic foundation models

A: We thank the reviewer for this valuable suggestion. We replace Evo2 with three alternative genomic encoders: DNABERT-2 (D2), Nucleotide Transformer (NT), and GenomeOcean (GO), while keeping the rest of GGBound framework unchanged. As shown below, Evo2 achieves the strongest overall performance, while alternative encoders generally yield lower scores. Nevertheless, our framework consistently outperforms the corresponding MLP baselines on all 17 phenotypes across all three alternative backbones, demonstrating its generalizability beyond Evo2. We have added these results to Section 5.4 of the revised manuscript.

GroupPhenotypeMetricD2-MLPD2-S2DNT-MLPNT-S2DGO-MLPGO-S2DE2-MLPGGBound
Phys. BoundaryTemperatureRMSE ↓9.219.058.708.629.318.757.927.72
pHRMSE ↓1.301.261.281.261.291.271.271.23
SalinityRMSE ↓3.963.933.563.504.143.883.343.29
Phys. OptimumTemperatureRMSE ↓7.026.906.386.067.346.515.455.00
pHRMSE ↓1.040.970.950.911.060.920.870.83
SalinityRMSE ↓3.353.202.872.803.142.592.662.57
MetabolismCarbon SourceF1 ↑0.450.590.480.620.570.630.620.63
Electron AcceptorF1 ↑0.480.620.520.650.330.770.520.87
Electron DonorF1 ↑0.540.810.430.770.530.820.540.89
Nitrogen SourceF1 ↑0.510.570.480.590.520.750.600.83
CategoricalOxygen ToleranceAcc.(%) ↑29.551.331.659.143.558.066.867.8
Energy TypeAcc.(%) ↑0.076.93.969.27.769.276.984.6
MorphologyCell ShapeAcc.(%) ↑4.331.912.844.721.348.045.654.4
MotilityAcc.(%) ↑4.068.06.570.559.065.576.077.5
FlagellaAcc.(%) ↑14.973.623.079.347.180.081.686.2
Spore FormationAcc.(%) ↑25.675.932.783.476.985.091.592.5
Gram StainAcc.(%) ↑21.584.529.589.583.087.096.096.5

Q2. A little hard to understand the theory.

A: We thank the reviewer for this suggestion. Our objective (Eq. 1) is to maximize I(Y;Y^∣A)I(Y;\hat Y\mid A), which measures how much information the final prediction captures about the true phenotype beyond auxiliary proposals. Higher values indicate more informative predictions, potentially leading to better prediction accuracy. Achieving this requires both retaining phenotype-relevant genomic information and effectively utilizing it during decoding (Eq. 2). We introduce genomic utility UK\mathcal U_K to quantify the contribution of the observed genome relative to alternatives. Theorem 1 establishes bounds on I(Y;Y^∣A)I(Y;\hat Y\mid A), showing that higher utility strengthens its lower bound when irrelevant genomic information is controlled. Optimizing genomic utility potentially leads to better prediction performance. This provides a theoretical motivation for improving predictions through genomic evidence utilization. We have revised Section 4.1 of our manuscript to clarify the theoretical intuition and its connection to our optimization framework.

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

AuthorsOpenReview ↗

Gentle Follow-up on the Rebuttal

Dear Reviewer rfRW,

We sincerely appreciate your positive and thoughtful review, valuable suggestions, and the time and effort you have devoted to evaluating our submission. We have provided additional experiments and clarifications in response to your valuable feedback.

We hope our responses and additional experimental results could address your concerns. We would be very grateful if you could share any further feedback or remaining concerns at your convenience. We would be happy to provide additional clarifications if needed.

Best regards,

Paper 15002 Authors

Reviewer_rfRWOpenReview ↗

Thanks for the rebuttal

Thanks for the detailed response. These solve my concerns, and I keep me positive score.

We sincerely appreciate your time and constructive feedback. We are deeply encouraged that our responses have addressed your concerns. Thank you again for your support!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

We sincerely thank all reviewers for their time, thoughtful feedback, and constructive evaluations. Reviewer rfRW recognized the novel methodology of GGBound and S2D-GRPO, the strain-centric benchmark, and the solid experimental results. Reviewer SGVV also highlighted token-level genomic utility as a critical strength. Reviewer NXUm commended our well-justified framework, thorough component ablations, and strong empirical performance. In particular, Reviewer NXUm described S2D-GRPO as an ingenious mechanism for encouraging genuine genomic evidence utilization and acknowledged its effectiveness in resolving conflicting proposals.

Based on the reviewers' valuable suggestions, we have further strengthened our manuscript and conducted additional evaluations, with revision marked blue:

  1. Genomic Encoder Generalization: We evaluated three alternative genomic encoders (DNABERT-2, Nucleotide Transformer, and GenomeOcean), demonstrating the generalizability of GGBound beyond Evo2 (Section 5.4 and Appendix D.8).
  2. Comparison Fairness: We introduced an additional proposal-conditioned SFT baseline to disentangle the improvements from S2D-GRPO and proposal-format exposure (Appendix D.9).
  3. Theoretical and Methodological Clarity: We clarify how our information-bottleneck formulation motivates genomic information retention and utilization, and how Theorem 1 connects genomic utility to predictive information. Together, these insights motivate the overall design of GGBound. (Section 4.1)
  4. Statistical Reliability: We conducted 20,000 strain-level bootstrap resamples to establish confidence intervals for our main results and ablation studies (Appendices D.1 and D.6), and investigated prediction collapse in sequence-only LLM baselines (Appendix D.1).
  5. Reasoning Supervision and Validation: We clarified teacher model selection, evidence-grounded rationale refinement, and the scope and limitations of biological claim validation (Appendix B.3).

We sincerely thank all reviewers again for their constructive feedback, which has helped us strengthen the empirical evidence, methodological clarity, and overall presentation of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.