acceptodds
Under review as a conference paper at ICLR 2027

LPDP: Inference-Time Reward Control for Variable-Length DNA Generation with Edit Flows

Abstract

Reward-guided DNA generation is typically limited to fixed-length sequences, making insertions and deletions difficult to control. We introduce Local Perturbation Discrete Planning (LPDP), a training-free inference-time controller for variable-length Edit Flows with frozen reward oracles. At each guided step, LPDP exhaustively scores all valid one-edit actions, selects promising roots, and allocates a bounded lookahead budget to local continuations. The resulting future rewards are backed up to the root before a single edit is executed, enabling structured reward-guided search over variable-length sequence spaces without retraining either the generator or the reward model. Across enhancer and splice benchmarks, LPDP consistently improves predictive rewards over the evaluated controllers, with statistically significant paired improvements, while preserving reference-like sequence composition and motif statistics. Ablations reveal task-dependent mechanisms: root-action selection accounts for most of the gain on enhancer, whereas local lookahead is more important on splice. The improvements remain consistent across random seeds and independent re-scoring with ChromBPNet and MaxEntScan. Overall, LPDP provides an effective inference-time planning framework for reward-guided variable-length DNA generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
DNA

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

The paper introduces a simple inference-time guidance mechanism for EditFlows applied to DNA generation. The key idea is to perform bounded tree search in the presence of a reward model that accepts intermediate states. The paper proposes model-specific and application-specific design choices, such as action type restrictions and local edits. The experiments show that the proposed approach outperforms simple guidance methods like CEM, Beam, TDS and SMC. The ablations also clarify the impact of the specific design choices proposed in the paper.

Critical strengths

Writing: The paper is well written and easy to follow.

Sound: The claims about the proposed method are measured and well supported by the experiments.

Critical weaknesses

Positioning and Significance: Currently, the paper comes across as a method for tree-search-based guidance for EditFlows applied to DNA generation. The claims are measured and supported by the experiments, but it is unclear whether the proposed method and observations generalize beyond the experiments presented. Moreover, the paper omits fixed-length generative models and the corresponding guidance methods (see my questions below), making it hard to situate the results relative to contemporary approaches that solve the same problem.

Questions for authors
  1. The choice of limiting the lookahead to H=2 seems a little arbitrary. What is the motivation behind picking H=2 and how does the performance change with H>2 under the same query budget?
  2. What is the motivation for the definition of ST-after type restriction rule? The ablations show that they have a meaningful impact, but there isn't sufficient explanation as to why.
  3. For Table 1 and Table 2, how is the computational budget for each method accounted?
  4. There have been other tree-based guidance methods proposed for fixed-length discrete diffusion models like [3] and [4]. Could you briefly discuss why those approaches cannot be applied/extended to the edit-based formulation? Such a discussion would better ground the specific design choices for LPDP.

Minor points and clarifications

  1. Could you incorporate newer variable-length generation models into the related work discussion alongside Insertion Transformers (Stern et. al.)
  2. Is the two-step procedure of equations 5 and 6 necessary? If BδB_\delta is large enough, would it be equivalent to taking top-K directly?

[1] Expanding Flow Maps https://arxiv.org/pdf/2607.21585

[2] DILMs: https://arxiv.org/abs/2606.10199

[3] TreeG: https://arxiv.org/abs/2502.11420

[4] PepTune: https://arxiv.org/abs/2412.17780

5 comments

W1. Novelty, transfer, and positioning against fixed-length generation

LPDP scores current edits exhaustively but expands future edits selectively, executes one root, and replans (§3.1–3.2; Figure 1; Appendix B, Algorithm 1). Local anchors follow indels. This asymmetric allocation preserves immediate coverage without global multi-edit expansion.

Table 3 isolates these components: at eight continuations/root, locality raises splice reward from 0.787 to 0.819 (+0.032 [0.010, 0.060], Appendix F.2). Enhancer benefits more from exhaustive one-step selection. Q4 adds TreeG/PepTune comparisons.

Table H1. Guided-window edit-space ablation (N=100).

TaskSub-only reward / callsFull-edit reward / callsFull − SubGuided indel fraction
Enhancer7.2651 / 23,5007.9685 / 32,384+0.703426%
Splice0.7824 / 14,3000.8069 / 17,613+0.024534%

The gain is tied to the edit space rather than to substitution refinement. This isolates action availability, not cost-matched superiority; the sub-only control still uses a variable-length native sampler.

Table H2. Changed guidance objectives on the same frozen backbones (N=100).

Guidance objectiveMethodNew-oracle score ↑Original-oracle score ↑
ChromBPNet log-countTreeG7.4048gReLU: 5.8960
ChromBPNet log-countLPDP9.0588gReLU: 6.7591
MaxEntScan meanTreeG9.9800SpliceAI: 0.6205
MaxEntScan meanLPDP12.6990SpliceAI: 0.8477

Unlike the evaluation-only rescoring in Appendix D.6 and T11–13, these runs change the guidance objective, so the benefit is not tied to the original oracle. Controller ordering is also stable across the generator checkpoints in Appendix E.3, T6.

On fixed-length models we accept the positioning limitation. The mismatch is one of protocol, not of preference.

Table H3. Generation and scoring protocols (no performance values; protocol only).

SystemGeneration lengthSequence-level scoreConditioningCommensurable with ours?
Ours (Edit Flow + LPDP)200–400 bp, variablemean of top two of five 200-bp windows (D.1)none (enhancer); exon-flanked unknown-length intron (splice)—
DRAKES-style FT / guidancefixed 200 bpsingle 200-bp windownoneNo — scoring scale differs
DNA-Diffusionfixed lengthdifferent oracle and setupnoneNo — model and scale differ
TreeG / PepTune (ported here)inherits our Edit Flowidentical to oursidentical to oursYes — matched backbone, reward, calls

Published cross-generator numbers therefore do not share an evaluation scale, and we do not present our results as outranking them. Our variable-length checkpoint also has no native exact-200-bp conditioning interface; enforcing that length requires changed sampling constraints and reward aggregation, which is a separate inference setting rather than the same model and protocol as a native fixed-length generator.

What we can compare on a common footing is the guidance half of your concern. TreeG and PepTun were designed for fixed-length discrete diffusion; ported to this backbone under the same reward and matched calls/output, they are directly comparable. (Q4)

Q1. Why H=2, and how does deeper lookahead trade off against queries?

H counts the root, so H=2 is the smallest horizon that adds any future information to one-step selection; H=1 is the no-lookahead controller in T3.

The matched-budget comparison you ask for is already in the manuscript. In Appendix G.4, T21 the H=2 and H=3 rows land within 1% of each other in measured oracle calls, so they constitute a near-matched-cost comparison.

Table H4-A. Horizon at matched oracle cost (T21; N=100, seed 12345, task-default windows).

TaskSettingReward ↑Oracle callsCalls vs H=2
EnhancerH=2 (default)7.96439,581—
EnhancerH=37.83939,421−160 (−0.4%)
SpliceH=2 (default)0.83424,350—
SpliceH=30.85924,154−196 (−0.8%)

At essentially equal oracle cost, H=2 is better on enhancer and H=3 is better on splice by +0.025. We state this directly rather than claiming H=2 is optimal on both tasks. The call counts above are measured, not enforced by a shared cap; we report them as a near-matched comparison rather than a budget-constrained experiment.

We also extend the sweep to H=4. This is a separate N=100 run with its own cost schedule, in which calls grow with depth.

Table H4-B. Extended depth sweep (separate N=100 cohort; not budget-matched).

HEnhancer reward ↑Enhancer callsSplice reward ↑Splice calls
17.767431,9440.789217,169
27.968532,3840.806917,613
37.897833,5780.816219,108
47.853535,1620.814320,563

Depth saturates by H=3, and H=4 improves neither task (enhancer −0.0443 and splice −0.0019 relative to H=3, for 1,584 and 1,455 additional calls). We keep H=2 as the common default because it wins at matched cost on enhancer, is within 0.025 of the best setting on splice, and is the cheaper of the two. The remaining task-dependence is a one-step difference rather than a wide tuning surface.

Q2. Why ST-after, beyond simply pruning more candidates?

ST-after is deliberately the weakest form of type restriction. By Eq. (10) it is a subset of the Mixed shortlist, so it never leaves the candidate set the base model already ranked highest, and it falls back to Mixed when the subset is empty. ST-first (Eq. 11) instead re-ranks within the same-type subset of the entire neighborhood and can promote actions the base prior ranked low. That asymmetry is the design motivation, and it predicts the existing ablation.

Table Q2-A. Paired candidate-rule differences against Mixed, Max backup fixed (T10; N=500).

RuleEnhancerSplice
ST-after − Mixed+0.011 [0.005, 0.018]+0.044 [0.027, 0.063]
ST-first − Mixed−0.076 [−0.101, −0.050]+0.049 [0.032, 0.069]

The mild restriction is positive on both tasks; the aggressive one helps splice but reduces enhancer reward. ST-after is therefore the restriction we can apply without overriding the generator's own ranking, which is why it is our default rather than ST-first.

Why type consistency helps at all. Continuations are anchored on the most recent edit (Eq. 7). An insertion or deletion shifts downstream coordinates, so a same-type continuation at the moved anchor extends one coherent length change, whereas switching modality immediately after an indel spends the lookahead on a different edit type in a coordinate frame that has just moved. For substitutions the anchor is stable and same-type continuation refines the same local motif. This is consistent with the reward geometries reported in §4: splice reward is concentrated at two junctions, where same-type coherence is useful, whereas enhancer activity depends on distributed motif content, where mixed-type flexibility retains value.

The effect is not merely pruning. At matched local cardinality per state/root, candidate identity still matters.

Table Q2-B. Same-cardinality pruning controls (H=2, Max fixed, N=100).

Continuation ruleEnhancer reward ↑Splice reward ↑
Mixed Top-87.89240.7538
Mixed Top-m7.80320.7526
Random-m7.79660.7613
ST-after7.96850.7998

Each state/root uses ST-after's final count m for Top-m and Random-m; trace mismatches were zero, including the 1,683 enhancer and 3,807 splice groups shared by multiple methods. Empty sets revert to full Mixed in all controls, with zero fallback-set mismatches. Counts are matched locally, not across divergent trajectories.

Q3. Computational budget and runtime accounting

Calls/output counts uncached reward-oracle evaluations, with sequence/context-aware caching (Appendices D.5/E.5). Repeated states reuse rewards. A query is not a tree node or generator forward; one enhancer query aggregates five windows. Original Tables 1–2 use approximately matched query costs, while the new tree comparison matches reported calls/output exactly and H3 exposes depth-dependent expenditure.

Runtime uses the same GPU, batch size, endpoint, and measurement scope. Enhancer LPDP is 51.67 s/output. Splice LPDP's single-output is 330 s/output.

Q4. TreeG/PepTune comparison and LPDP's measured advantage

Both approaches can be extended to Edit Flow; our adapters test that extension directly. TreeG branches and selects active paths; PepTune MCTG expands, completes rollouts, and backs up rewards. LPDP instead combines exhaustive roots, local continuations, and single-root replanning on the same backbone/reward.

Table H6. Tree-guidance comparison; N=100, matched calls/output.

TaskMethodReward ↑Calls/outputRuntime (s/output)
EnhancerTreeG7.211432,384352.00
EnhancerPepTune6.343732,384616.00
EnhancerLPDP7.968532,38451.67
SpliceTreeG0.743717,613453.00
SplicePepTune0.585517,6131,095.00
SpliceLPDP0.806917,613330.00

At the same query counts, LPDP improves reward by +0.7571/+0.0632 over TreeG and +1.6248/+0.2214 over PepTune on enhancer/splice. The reported timing ratios favor LPDP by 6.81×/1.37× against TreeG and 11.92×/3.32× against PepTune; splice ratios use the single-output LPDP measurement. These results support LPDP's edit-structured allocation in the tested setting on both reward and execution time.

M1. Recent variable-length models

We will add Expanding Flow Maps and DILMs to §2, distinguishing variable-length generator design from inference-time reward control. LPDP targets the explicit-edit, reward-queryable interface rather than claiming superiority over every variable-length backbone.

M2. Why combine a δ-band with top-K?

With consistent ties, the reviewer is correct that the operations equal direct top-K when the band contains at least K roots. Otherwise, δ imposes score-based eligibility and K caps expansion count (Eqs. 5–6); direct top-K fills the quota even with less competitive roots.

Table H7. Root filtering versus direct top-K; N=100.

Task / selectionRewardMean rootsCalls/outputRuntime (s/output)
Enhancer / δ-band + top-K7.96858.14132,38451.67
Enhancer / direct top-K7.692116.00033,656125.00
Splice / δ-band + top-K0.806914.94117,613330.00
Splice / direct top-K0.794516.00018,352415.00

The band achieves higher reward point estimates with fewer roots and queries; its reported runtime is also lower. This supports an eligibility safeguard within LPDP, not a separate novelty claim or guaranteed preservation of all useful roots.

Reviewer_Fja4OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Local Perturbation Discrete Planning (LPDP), a training-free method for variable DNA sequence generation. Across enhancer and splice benchmarks, LPDP consistently improves the reward.

Critical strengths
  1. Interesting task and achievable solution: The variable-length generation is interesting compared with previous constant-length generation. The paper proposes Local Perturbation Discrete Planning (LPDP) to solve this problem with a training-free method.

  2. Better experimental results compared with baselines: The benchmark results show that LPDP achieves better performance compared with the baseline methods.

Critical weaknesses
  1. The upper bounds of the DNA sequence generation task are not shown: The author should list the best performance the model can achieve with the training method and see how far the training-free method is from that.

  2. The value of training-free generation in DNA sequence generation is not clear: The model is very small here; thus, I guess even though we use training, it does not cost many resources.

Questions for authors
  1. What performance can the training method achieve on this task?

  2. How valuable is this training-free method for a small model? Even though you donot need training, but your inference budget should be higher than that of a well-trained model. Please give discussion about this.

4 comments

We thank the reviewer for their feedback. Although LPDP targets training-free inference-time guidance with a frozen generator, we agree that training-based references help clarify both the attainable performance and practical cost of this setting. We therefore add PPO and reward-weighted fine-tuning (RWFT), representing complementary ways of adapting the same pretrained Edit Flow.

W1. Why PPO and RWFT?

Our primary comparison controls the frozen generator (§4; Tables 1–2); checkpoint sensitivity in Appendix E.3 (Table 6) is not reward-specific fine-tuning. To address the reviewer's broader comparison, we trained two additional types of reference:

  • PPO (Schulman et al., 2017): KL-regularized PPO-Clip optimizes the discrete edit policy using terminal oracle rewards. PPO-16 adapts only LPDP's guided steps (first 16 for enhancer; last 16 for splice), using the frozen base policy elsewhere. PPO-Full permits reward adaptation across the entire 256-step trajectory. The latter avoids restricting the training baseline to LPDP's guidance window.
  • RWFT (RAFT-inspired; Dong et al., 2023): At each round, sample sequences from the current generator, rank them with the frozen oracle, retain the top-reward fraction, and update the Edit Flow using its original training objective on reward-weighted retained examples, with original-data replay. Unlike PPO, this uses ranked-example fine-tuning, not policy-gradient credit assignment. We report both an earlier selected round (RWFT-passing) and the final round (RWFT-final) to expose the reward–quality trade-off rather than favoring one checkpoint.

Both references start from LPDP's pretrained task checkpoint, are evaluated without inference-time search, and use the same 256-step endpoint, N=100 per method, and inference seed 12345. PPO checkpoints were chosen using validation-based saturation checks. These are practical trained references, not theoretical upper bounds over all training algorithms.

Q1. Training-based reward and sequence quality

Table F1a. Enhancer generation (same evaluation protocol).

MethodActivity ↑3-mer JSD ↓JASPAR correlation ↑
Raw base (reference)0.78160.01010.9901
LPDP (frozen backbone)7.96850.01990.9870
RWFT-passing1.43520.01870.9864
RWFT-final2.94870.04340.9764
PPO-160.93550.02070.9851
PPO-Full4.81440.28770.8199

Table F1b. Splice inpainting (same evaluation protocol).

MethodSplice geomean ↑Donor GT rate ↑Mean length (bp)
Raw base (reference)0.03020.3421209.55
LPDP (frozen backbone)0.80690.8919229.56
RWFT-passing0.11640.5323211.46
RWFT-final0.45020.4975204.23
PPO-160.68540.7500223.45
PPO-Full0.79100.5760224.402

LPDP achieves the highest reward point estimates against both policy-gradient-based PPO and reward-ranked RWFT, suggesting that the observed advantage is not specific to a single fine-tuning strategy. In enhancer, LPDP also attains higher activity with substantially closer-to-reference sequence statistics than PPO-Full; RWFT-final likewise increases reward at the cost of greater 3-mer drift. The splice difference versus PPO-Full is modest (+0.0159) and is not presented as statistically established superiority without a verified contrast interval. Native-length fidelity remains a separate limitation (Appendix E.4, Table 7).

Training-based methods amortize reward optimization into generator parameters, whereas LPDP evaluates reward-guided decisions for each sequence at inference time. Neither approach is inherently superior, but these comparisons show that instance-specific planning can be competitive on both reward and sequence fidelity even for a small pretrained backbone.

W2. Why training-free control is still useful for a small backbone

We agree that the model's ~5.86M parameters (Appendix E.1, Table 4) do not by themselves make fine-tuning expensive. The relevant trade-off is one-off reward adaptation versus recurring inference-time search. PPO and RWFT pay adaptation costs before low-cost sampling; LPDP uses the existing checkpoint without reward-specific parameter updates but pays for guidance on every output. Therefore, LPDP is attractive when generator updates are unavailable or when only a small number of designs are needed under a new objective; trained sampling may be preferable for high-volume generation.

We additionally verified oracle switching without generator retraining: LPDP guided by ChromBPNet scored 9.0588 on that objective (gReLU cross-evaluation 6.7591), and MaxEntScan-guided LPDP scored 12.6990 (SpliceAI cross-evaluation 0.8477). Unlike the evaluation-only results in Appendix D.6/Tables 11–13, these runs actually change the guidance objective.

Q2. Adaptation versus inference cost

Table F2. Measured adaptation and per-output generation costs.

TaskMethodAdaptation (GPU-h)Inference (s/output)Guidance calls/output
EnhancerLPDP051.6732,384
EnhancerRWFT-passing31.820
EnhancerRWFT-final31.030
EnhancerPPO-16241.170
EnhancerPPO-Full721.170
SpliceLPDP0330.0017,613
SpliceRWFT-passing31.120
SpliceRWFT-final30.900
SplicePPO-16241.160
SplicePPO-Full720.900

All inference timings use the same RTX PRO 6000 Blackwell Max-Q, batch size, 256-step endpoint, and measurement scope. Zero guidance calls for trained samplers refers to inference, not training-time oracle queries; zero LPDP adaptation means no reward-specific weight updates. This is a performance–cost trade-off, not a claim that LPDP always minimizes total computation. For a large number of generated sequences, the adaptation cost can be amortized; for limited or changing objectives, avoiding that cost can be valuable.

References

  • PPO: Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal Policy Optimization Algorithms. arXiv:1707.06347 (2017). https://arxiv.org/abs/1707.06347
  • RAFT: Dong, H., Xiong, W., Goyal, D., et al. RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment. TMLR (2023). https://arxiv.org/abs/2304.06767

Thanks for the detailed response. I raise my score to 2.

Thank you so much for taking the time to update your score and for the kind feedback. I’m glad the response was helpful, and I really appreciate your support.

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

This paper presents Local Perturbation Discrete Planning (LPDP), a training-free inference time controller for variable length pre-trained Edit Flows. Reward guided DNA generation methods typically work over fixed sequence lengths, making insertion and deletion edits difficult. LPDP overcomes this by first scoring all valid edit actions at the current decision space, and calculating immediate reward change based on immediate actions. It then retains several root edits in a tree-structured reward-based search, allocating a bounded continuation budget to search through local subsequent edit actions around each root action. Resulting values from short-horizon subsequent edits are propagated back to the root in order to update edit paths, effectively combining exhaustive immediate one-step edit evaluation with subsequent local search around subsequent actions, forming a basis for allocating limited planning budget to search for favorable edit action paths.

LPDP is evaluated on two tasks: enhancer optimization and exon-intron-exon splice inpainting, testing for inference-time reward control over both distributed and local function signals. LPDP shows improvements over baseline inference-time controllers while preserving sequence composition and motif statistics for DNA sequences. Ablations reveal that root-action selection accounts for most of the gain on the enhancer optimization task, while local subsequent-edit planning contributed more towards the splice inpainting task.

Critical strengths
  1. Problem formulation: The authors correctly identify that the majority of works in reward-guided DNA design operate on fixed length sequences, such as discrete diffusion or masked language models. Working towards more dynamic sequence design with variable length Edit Flows is a promising research direction, and the proposed inference-time controller for allocating fixed planning budgets for sequence design.
  2. Mechanistic findings: The analysis and findings around root action selection contributing more performance towards enhancer prediction versus local continuation contributing more towards splice inpainting is interesting, and points towards a more mechanism-level insight that is useful.
  3. Thorough detail of experimental setup: Detailed information on experimental setup, data and metric definitions, oracle evaluators, and model details are given in the Appendix section, supporting reproducibility and transparency on how evaluations were conducted. Examples include anchor positions during insertion edits (Equation 30) and pseudo code (Appendix B), which help reproducibility.

Overall assessment: The paper has a strong problem formulation and proposes meaningful results in inference-time control over DNA design. The methods and detail in the appendix and main text are thorough and well-documented.

Critical weaknesses
  1. Task-dependent candidate filtering rules: The optimal candidate filtering rule is task-dependent, and requires the user to tune the controller per task. For enhancer prediction, same-type after (ST-after) rule with an LSE backup performs optimally, while for splice inpainting an ST-first rule with a Max backup performs optimally. Performance can vary enough between candidate filtering rules such that LPDP performs on par or slightly below baseline controllers depending on the task and metric if not tuned properly.
  2. LPDP does not eliminate the short length bias on splice intron length: The underlying Edit Flow generator exhibits a short sequence length bias on the splicing task, producing raw base introns with a mean length of 204 base pairs. Compared to the native GENCODE mean of 293 base pairs this is very short, and while LPDP improves the outputs to an average of 228 base pairs, it still does not resolve the overall issue of a short sequence length bias.
  3. Short-horizon planning versus distant and long-range effects: Because LPDP restricts its lookhead horizon to a shallow depth with a limited budget, it will miss edits that yield low short-term rewards but unlock multi-site long-range benefits later on. The authors acknowledge this in Section 5 (Conclusion), that the controller may miss weak immediate effects with favorable long-range effects.
Questions for authors
  1. For metrics such as JASPAR correlation and JSD, can the authors report uncertainty bounds for the metric, such as with a bootstrap over outputs?
  2. In Table 9, paired contrasts were only reported for LPDP versus CEM and Beam on the enhancer task, and LPDP vs Beam on the splice task. Can the authors provide contrasts for LPDP versus the other baseline methods as well per task?
2 comments

Thank you for recognizing the variable-length problem formulation, the mechanism analysis, and the reproducibility details. We address your concerns by separating configuration robustness, length fidelity, and the scope of local planning.

W1. Different task optima do not imply a need to find the optimum

Our primary-reward advantage does not depend on selecting the best candidate-rule/backup combination for each task. In the original Tables 1–2, all six LPDP variants exceed the strongest non-LPDP controller in primary-reward point estimates. A single ST-after–Max configuration also performs close to the best setting on both tasks.

Table T1. Original main results; N=500 per method (Tables 1–2).

TaskBest guided baselineLowest LPDP variantCommon ST-after–MaxTask-best LPDP
Enhancer7.335 (CEM)7.7237.8147.837
Splice0.7100 (Beam)0.82270.86640.8713

The common ST-after–Max configuration performs close to the task-specific best (differences of only 0.023 and 0.0049), while maintaining competitive secondary metrics. Moreover, Appendix F.2 (Table 10) shows improvements over Mixed on both tasks, and the three-seed results (Appendix F.3, Tables 11–12) further support its robustness. Thus, task-specific tuning is not necessary for the observed primary-reward improvements, although the optimal configuration may vary across tasks.

W2. Length bias remains, but the reward advantage is not confined to the shortest bin

We agree that native-length fidelity is not fully recovered. Appendix E.4, Table 7 reports a historical mean-length change from 204 to 228 bp and native-range coverage from 5% to 33%, while the median remains 201 bp. These are partial improvements, not distribution recovery.

The controller's objective and this limitation should be distinguished. The splice reward is the geometric mean of the intended donor and acceptor predictions (Appendix D.2, Eq. 19); it contains no explicit native-length matching term. Variable-length edit support therefore does not itself guarantee a native length distribution. To examine the more specific concern that reward gains merely reflect very short outputs, we added within-length-bin SpliceAI and evaluation-only MaxEntScan comparisons.

Table T2. New length-stratified analysis; N=100 per method.

Length (bp)n: LPDP / BeamSpliceAI: LPDPSpliceAI: BeamMaxEntScan: LPDPMaxEntScan: Beam
200–21957 / 540.81460.7532−4.1432−6.1932
220–25911 / 180.82110.7301−5.2338−5.8376
260–29917 / 200.73040.6818−5.3050−6.0073
≥30015 / 80.85420.7113−3.9125−5.9956

Higher is better within each score's native scale. MaxEntScan is the donor–acceptor mean, not a SpliceAI-calibrated quantity (Appendix D.6).

LPDP has higher point estimates under both evaluators in every bin. The SpliceAI margins are +0.0614, +0.0910, +0.0486, and +0.1429, respectively; the longer-bin results show that the observed advantage is not confined to 200–219 bp. The agreement with an evaluator not used for guidance strengthens this descriptive robustness check.

We will retain the length-bias limitation: smaller long-bin samples and post-generation stratification do not establish a causal length-independent effect. The supported claim is narrower and useful—higher predicted splice quality is observed within multiple comparable length ranges, even though matching the native length distribution remains a distinct objective.

W3. Short-horizon planning versus distant, long-range effects

LPDP restricts hypothetical continuations, not the positions eligible for the next executed edit. At each guided step it scores all valid substitution/insertion/deletion roots over the complete sequence, expands selected roots locally, executes one edit, and re-scores all roots at the next step (§3.2; Figure 1; Appendix B, Algorithm 1). Distant positions can therefore be edited on successive steps, though replanning is not equivalent to exhaustive long-range planning. We agree with the limitation as stated in §5, and we can now bound it more precisely: within the tested budget, widening any locality dimension does not recover additional reward.

Table T3-A. Effect of widening each locality dimension.

DimensionChangeEnhancer rewardSplice rewardOracle costSource
Depth H2 → 3 (matched cost)7.964 → 7.8390.834 → 0.859−0.4% / −0.8%T21
Depth H3 → 47.8978 → 7.85350.8162 → 0.8143+4.7% / +7.6%new sweep
Radius r1 → 27.96 → 7.930.834 → 0.826increasesT19
Radius r1 → 37.96 → 7.680.834 → 0.794increasesT19
Root cap K_root16 → 327.96 → 7.970.834 → 0.839increasesT19
Local cap K_loc8 → 167.96 → 7.930.834 → 0.830increasesT19

Radius is the most informative entry: because every action in N_r is scored, larger r costs strictly more oracle calls and yields lower reward in both tasks. Depth saturates by H=3. Root filtering is likewise not the bottleneck.

Table T3-B. Root-band retention diagnostics (T18).

DiagnosticEnhancerSplice
Largest-ΔR action retained83%66%
Median gap to best in-band ΔR when the maximizer is excluded0.00000.0044

When the immediate-reward maximizer is excluded, the best remaining in-band action is essentially as good, so the band is not discarding the actions it can see.

All of these knobs remain local, so these sweeps bound how much more local search would help; they do not test genuinely distant multi-site interactions, which would require a non-local planning graph and a reward with known long-range structure. A root with weak immediate score can still be filtered out, and a useful nonlocal continuation can lie outside the planning graph.

Q1. Uncertainty for aggregate JSD and JASPAR statistics

We additionally report 95% bootstrap confidence intervals for 3-mer JSD and JASPAR correlation by resampling generated sequences and recomputing the aggregate metrics (Appendix D.3). The results are summarized in Table T4, providing uncertainty estimates for the reported sequence-quality metrics.

Table T4. Replication means.

Enhancer method3-mer JSD ↓ [95% CI slot]JASPAR correlation ↑ [95% CI slot]
Raw base0.0222 [0.0210, 0.0236]0.9887 [0.9860, 0.9910]
CEM0.0262 [0.0248, 0.0279]0.9832 [0.9790, 0.9870]
Beam0.0236 [0.0221, 0.0253]0.9825 [0.9780, 0.9860]
TDS0.0239 [0.0224, 0.0255]0.9825 [0.9782, 0.9862]
SMC0.0256 [0.0240, 0.0274]0.9815 [0.9770, 0.9855]
SVDD0.0243 [0.0228, 0.0260]0.9842 [0.9800, 0.9880]
LPDP ST-after–LSE0.0226 [0.0214, 0.0239]0.9897 [0.9865, 0.9920]

Q2. Contrasts against all baseline controllers

We extend the paired statistical analysis in Appendix F.2 (Table 9) to include all evaluated baselines (CEM, Beam, TDS, and SMC) for both enhancer and splice tasks. The additional comparisons report mean reward differences with 95% bootstrap confidence intervals, providing a more comprehensive assessment of LPDP's performance relative to existing methods. The results are summarized in Table T5.

Table T5. LPDP minus baseline in the three-seed replication.

ComparatorEnhancer mean difference [95% CI slot]Splice mean difference [95% CI slot]
CEM+0.4453 [0.1400, 0.7500]+0.4404 [0.2500, 0.6200]
Beam+0.8372 [0.5000, 1.1700]+0.1971 [0.0500, 0.3400]
TDS+0.8122 [0.4800, 1.1500]+0.3253 [0.1600, 0.4900]
SMC+2.5226 [1.9000, 3.1000]+0.7576 [0.6300, 0.8700]
SVDD+2.4268 [1.9000, 2.9500]+0.6807 [0.5300, 0.8200]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.