acceptodds
Withdrawn from ICLR 2027

A Reasoning-Informed Verifier-Guided Editing Agent for RNA Inverse Folding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

RNA function follows from RNA structure, so designing sequences that reliably fold into a prescribed shape is central to RNA therapeutics and synthetic biology. Inverse folding is difficult because the sequence-to-structure map is highly degenerate and rugged, candidate quality is observable only after folding, and optimizing against a single folding oracle invites overfitting. Existing designers emit a sequence in one open-loop pass, provide no rationale for individual nucleotide choices, and accept candidates through hand-tuned thresholds. We present RIVET, which recasts RNA inverse folding as verifier-guided, model-based reinforcement learning over interpretable edits. Instead of emitting a sequence directly, RIVET repairs one through a propose, imagine, select, verify, and refold loop built from four modules. SKETCH is a reasoning policy that proposes typed, biophysically grounded edits with explicit rationales and is trained by group-relative policy optimization. ECHO is a learned world model that ranks edits without folding. SCOUT forwards only the most promising edits to costly verification. GAUGE is an energy-based verifier that turns folding feasibility into a calibrated dense reward and conformal accept-reject gate. Two guarantees make the loop safe: potential-based shaping provably cannot bias the design objective, and the gate bounds false acceptance at a chosen rate. The same loop supports tertiary-structure design, secondary self-consistency, and binding-conditioned design for RNA-ligand and RNA-RBP targets. Under a multi-tool protocol whose reporting predictors are disjoint from the loop-internal folder, RIVET attains state-of-the-art native recovery of 56.3%, the best tertiary and secondary self-consistency, and generalizes robustly to held-out RNA-Puzzles and CASP15 targets. Our code is available at https://anonymous.4open.science/r/RIVET-4517/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA Inverse Folding, Reasoning, Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ERSNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

RIVET is a framework for RNA inverse folding which uses iterative seq editing instead of generating a complete seq in a single pass. The framework has 4 components: 1) SKETCH, an LLM-based policy which proposes edits with reasoning traces, 2) ECHO, a learned world model to predicts changes in structure quality, 3) SCOUT, selects promising edits for verification, 4) GAUGE, an energy-based verifier which provides reward signals & conformal acceptance criterion. Experiments show increased performance over the baselines on the RNA 3D IF task, and include the 2D structure agreement and generalization on RNA families, and predicted property proxies.

Critical strengths
  1. The formulation of RNA inverse folding as iterative editing. The framework is flexible to accommodate additional seq and functional constraints. The combination of the world model is novelty & fancy with structural selective verification.
  2. The evaluation is complete including seq recovery, 3D sc-score, 2D agreement, and different RNA families, which is more mature compared with other counterparts in the RNA inverse folding area.
Critical weaknesses

(Major) RIVET intros a new backbone conditioned initializer + SKETCH, ECHO, SCOUT, GAUGE. The framework looks promising at the first glance, but those fancy components seems to be enforced together form an extremely novel system with limited empirical improvements.

The table 1 only compares the whole system with other methods. I have this concern because, in Table 4 ablation results, the full depth-3 editing improves mean scTM by only 0.0058 (seq recovery lowered by 0.0042 after editing) on the 60-targets. The reported no-edit reference has mean scTM 0.418, and ECHO-only guidance improves mean scTM by 0.002, while uniform-random slate selection improves by 0.0021 (with even lower σ\sigma). Based on the reported mean results, ECHO-only guidance does not outperform uniform-random selection from the same legal slate. It's very weird that the whole system show competitive performance improvement. Does the ECHO collaborate well with the other modules well?

(Major) Also, say, if there are 60 targets and the average improvement is 0.0058, and I think the optimization on each targets is trivial. For the structure prediction problem, say for AlphaFold, if you simply change the random seeds, the structure prediction score will be increased somehow, and I think the increase will not be as trivial as 0.0058. Also, for the protein design problem, we often include a relatively larger benchmark so that the performance improvement will be more reliable. Of course I understand that the authors may argue that the RNA structure is scarce and no sufficient data.

(Major) A central claim is that SKETCH produces interpretable and biophysically grounded reasoning for residue edits. However, the authors later state that SKETCH is prompted with the target identifier, current sequence, step index, remaining edit budget, and legal alphabet. I wonder how the LLM directly assesses the target's structural geometry, folding defects, base-pair relationships, 3D contacts, previous verification feedback, etc. E.g.,, deciding whether an edit should resolve a steric clash or satisfy an A-minor interaction normally requires information about the structural environment. A target identifier and sequence alone do not explicitly supply this information.

(Major) On the world model formulation and its necessity for multi-step reasoning-guided editing. I am not fully convinced that ECHO functions as a world model in the proposed framework, or that its necessity has been sufficiently demonstrated. As described, ECHO mainly predicts a scalar quality change for a candidate edit, rather than the resulting RNA conformation or a complete state transition. This makes it closer to a learned edit-quality scoring function than a conventional dynamics model. While such an abstraction could still support planning, I'm confused on how ECHO enables meaningful multi-step reasoning, since the search procedure doesn't perform folding at intermediate nodes.

(Minor) Proposition 1 defines the task reward as state-quality function and used a shaped reward. The paper claims that the optimal policy under this reward is necessarily the same as under the original task reward. However, the potential depends on the reasoning trace r, but r is not included in the state definition in Section 3.1. The Markov-state assumptions behind the stated theorem probably need clarification. More importantly, GAUGE is not used only through potential-based reward shaping. Its raw energy also affects initialization, candidate ranking and acceptance. These operations can change which designs are selected, regardless of whether the shaping term preserves an optimal policy.

(Minor) Several in-consistency issues, see the questions part.

Questions for authors
  1. Why the table 1 shows the gRNAde trRosettaRNA scTM as 0.235, while the table 9 reports the tR score as 0.220, 0.226 and 0.206 by 3 length groups. Did I miss any splitting details?
  2. How did you implement the GAUGE module? You described in the manuscript as backbone/seq/text encoders + cross-attention, but your released codes only include the GAUGE as a scalar-feature MLP EnergyGauge, at rivet_system/gauge_energy.py. In your codes , the crossattention seems to be in a separate initializer module, which doesn't establish in the GAUGE modules. I wonder, what does the GAUGE's actual implementation and contribution? Please let me know if I miss any details.
  3. What are the full-test results for the backbone-conditioned initializer without any editing? How much do GRPO training, reasoning traces, ECHO, and GAUGE each improve performance under the same candidate and verification budget? I understand the full ablation experiments may take lots of computations, but since the existing ablation results are not convincing, I hope the authors can clarify or offer some insightful discussions.
  4. How does Proposition 1 handle finite-horizon termination, the task-quality difference reward, and the trace-dependent potential?
  5. How are the structural quality features of unverified intermediate states obtained? How does ECHO account for interactions between consecutive edits?
Reviewer_vwDCOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents RIVET, a closed-loop RNA inverse-folding system. An initializer generates a sequence for a target RNA backbone. A language-model policy proposes typed edits and textual rationales. A learned world model ranks edits. A selector sends chosen candidates to folding tools. An energy verifier supplies a dense shaping signal and an accept-reject gate.

The paper reports tertiary inverse-folding results on a structure-similarity split and additional results on RNA-Puzzles, CASP15, secondary-structure consistency, and predicted binding. RIVET reaches 56.3% native recovery and the highest mean self-consistency score among the compared backbone-conditioned methods. The paper also presents a potential-based shaping argument and a split-conformal false-acceptance bound under exchangeability.

Critical strengths
  1. The paper evaluates structure consistency with several reporting predictors that are separate from the loop-internal folder. This is a meaningful attempt to reduce single-oracle overfitting. The ordering is consistent across the three reported predictors in the main table.

  2. The benchmark construction tries to control sequence and structure similarity across splits. The union of sequence-identity and structure-similarity components is a stronger split design than a simple random split.

Critical weaknesses
  1. The incremental contribution of the closed-loop editing system is not yet clear. On the reported ablation subset, editing gives a relatively small improvement over the no-edit reference, and some component comparisons use different budgets. A budget-matched comparison would help separate the effect of the initializer from the effect of the editing loop.

  2. The conformal guarantee depends on assumptions that are not evaluated in the reported experiments. The guarantee assumes exchangeability, while the candidates are selected adaptively and the reported runs do not use online calibration refresh. The paper should state this scope more clearly and provide basic empirical calibration results.

Questions for authors
  1. Could the authors provide a budget-matched comparison between the initializer, a simple edit-search baseline, and the full RIVET loop?

  2. Could the authors clarify how the conformal calibration set is constructed and report the empirical false-acceptance rate in the evaluated settings?

  3. Are the main system comparisons matched in the number of generated candidates and folding-tool calls? If not, it would be helpful to report these budgets more explicitly.

  4. Is there evidence that the textual rationale contributes beyond the typed edit and sequence context? A brief ablation or clarification would help define the role of reasoning in the method.

Reviewer_nB1BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes RIVET, which models RNA inverse folding as a process combining sequence editing and verification. The method uses a language model to propose edits with biologically physically interpretable explanations, filters candidates through a learned quality prediction model, and then assists selection with an energy verifier and folding tools. The paper also proposes potential-based reward shaping and conformal acceptance gating, respectively used to preserve the optimization objective and control the false acceptance rate.

Critical strengths
  1. Extending one-shot sequence generation to inspectable local edits and using surrogate models to reduce expensive folding calls provides a meaningful direction for RNA design under limited computational budgets.
  2. The paper reports sequence recovery, tertiary structure, and secondary structure metrics simultaneously, and includes held-out targets such as RNA-Puzzles and CASP15. The method's advantages in Table 1 appear under multiple predictors, providing a positive signal for further research.
Critical weaknesses
  1. The main text claims that the internal folding tools and the reported tools do not overlap, but the validation tools listed in Appendix A.10.2 and A.12 include the same Protenix, Boltz-2, and trRosettaRNA family tools, so the independence of the evaluation is questionable. In addition, Table 3 gives the number of test targets as 4,301, whereas A.10 writes 369. Weighting the length-group counts in Table 3 does not match the reported 0.418. These issues directly affect the credibility of the main performance conclusions.
  2. In Table 4, three-step editing brings only a 0.0058 improvement in average scTM, and there is a lack of same-budget controls without GRPO, without reasoning text, and confidence intervals, making it difficult to determine whether the advantage comes from the initialization model or the proposed editing mechanism. On the theoretical side, the terminal potential term under a finite horizon is not handled.
Questions for authors
  1. Please explain the tool overlap, the difference in test set size between 4,301 and 369, and the reason why the grouped results in Table 3 are inconsistent with the overall mean.
  2. Under the same initialization, number of candidates, and folding budget, how do the variants without GRPO, without reasoning text, and using only the initialization model perform respectively?
  3. How does Proposition 1 handle the terminal term γTΦ(sT)\gamma^T\Phi(s_T) in the finite-horizon shaped return?
  4. What are the sample size nn, error rate α\alpha, and infeasible criterion used in conformal calibration? After adaptive search, does the actual false acceptance rate and its confidence interval still satisfy the stated bound?

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

Related papers

Discussion (0)

Sign in to comment.