RNA design models now propose sequences for a target 3D structure in under a second, but when a biochemist asks "why did you reject the alternative conformation I expected — the kissing-loop interaction, the different pseudoknot topology?" the systems have nothing to say. The question is contrastive: the human holds a mental model of RNA structure that differs from the model the design system actually uses, and the discrepancy needs to be explained, not just logged. There is a formal theory for exactly this situation — model reconciliation, which treats an explanation as the smallest update to the human's model that makes the system's chosen action appear correct (Tathagata Chakraborti et al.) — but it has only ever been run on discrete symbolic planning domains, never on continuous, flexible biopolymers.
This proposal brings model reconciliation to RNA inverse folding, and in doing so exposes a second problem that the RNA field has not confronted: an explanation is only trustworthy if the underlying model's knowledge is real. If a design model absorbed its benchmark's structural motifs during pre-training, its high sequence-recovery scores and its justifications may both be memorization in disguise, and the standard loss-based contamination detectors are known to fail once a model has been trained with reinforcement learning (On The Fragility of Benchmark Contamination Detection in Reasoning Models). The proposal therefore builds one system with two payoffs: it produces causal explanations for why a human's structural hypothesis is biophysically infeasible, and it turns the content of those explanations into a contamination probe that operates on what a model can justify rather than on its loss values.
Core Claim
The claim is that RNA structural constraints — base-pairing geometry, backbone torsion sterics, coaxial stacking, pseudoknot topology — can be discretized into a symbolic constraint system over which the machinery of formal model reconciliation applies directly. Given a human's foil (a proposed alternative structure compiled into constraints), the infeasibility of that foil can be attributed to a minimal set of violated constraints, computed exactly via the duality between minimal unsatisfiable sets and minimal correction sets rather than by repeated search. Formally, where the human's model of the RNA system is $M^H$ and the design system's is $M^R$, we seek the smallest update $\mathcal{E}$ such that
$$M^H + \mathcal{E} \;\models\; \text{the design system's proposal is sound, and the foil is explained infeasible},$$
with each element of $\mathcal{E}$ phrased as a named biophysical constraint ("the C2'-endo sugar pucker at position 12 clashes with the A-minor contact at position 40"), not a coordinate diff. The second half of the claim is that the quality of $\mathcal{E}$ degrades in a measurable way on structural families the model memorized, giving a contamination signal that survives reinforcement-learning post-training, where loss-based detection collapses.
Research Program
The first build is a discrete structural knowledge base for RNA. Continuous 3D coordinates must be converted into symbolic atoms before any reconciliation algorithm can touch them, and two results in the file say this discretization is both feasible and beneficial. The comprehensive RNA benchmark of A Comprehensive Benchmark for RNA 3D Structure-Function Modeling showed that coarse-grained 2.5D graphs encoding Leontis-Westhof base-pairing geometry often outperform full 3D coordinate models when training data is scarce — evidence that the symbolic abstraction is not a lossy afterthought but arguably the right level for RNA, where fewer than 3,000 high-quality structures exist (Cheng Tan et al.). And RiboSphere, which discretizes continuous SE(3)-invariant representations into structural tokens via Finite Scalar Quantization and reconstructs coordinates with a flow-matching decoder, shows the round trip through discrete tokens can be accurate (reconstruction at RMSD 1.25 Å and TM-score 0.84). Our knowledge base combines three constraint families: Leontis-Westhof base-pair geometry from 2.5D graphs; binned backbone torsion constraints covering the six torsion angles per nucleotide that make RNA far more flexible than protein (Anand et al.); and topological constraints on long-range contacts and pseudoknots, whose arrangement is the known bottleneck of tertiary folding (Miao et al.).
The second build is the reconciliation engine itself, and here we reuse an existing algorithm rather than invent one. Stylianos Loukas Vasileiou et al. reframed model reconciliation as a knowledge-base problem and showed, in their ALG2 algorithm, that minimal explanations can be extracted with SAT solvers by exploiting the hitting-set duality between minimal unsatisfiable sets and minimal correction sets — up to two orders of magnitude faster than prior planners. Their stated limitation is that everything must be propositional; our contribution on this axis is entirely the RNA-side encoding that makes a biopolymer propositional. The engine handles both reconciliation directions: when the foil is infeasible under the biophysical constraints, the minimal correction set is the explanation delivered to the human; when the foil exposes a constraint the design model got wrong, the same machinery identifies the model repair, connecting to the repair setting surveyed in A Survey on Model Repair in AI Planning. Foils are elicited and compiled following the XAIP-as-a-Service pattern of Benjamin Krarup et al., who found 89.9% of user questions are contrastive and compiled foils directly into model restrictions — but whose system, by their own account, fails silently when the foil renders the problem unsolvable. Solving that silent-failure case in a continuous physical domain is the core technical gap this proposal closes.
The third build is the contamination audit, which is the methodological lens rather than the mechanism. RNA benchmarks currently police contamination only as static sequence or structural overlap: the rnaglib benchmark de-duplicates with CD-HIT plus US-align structural alignment, and RDesign enforces a TM-score split, showing that naive splits inflate sequence recovery through memorized geometric motifs. But no one has audited the foundation-model teachers. ATL-FGD distills from RiNALMo, a 650M-parameter RNA foundation model, and reports an 8.21% improvement on Rfam without checking whether Rfam families sat in the teacher's pre-training distribution (Learning Adaptive Topology with FiLM-Guided Distillation for Tertiary Structure-Based RNA Design). Worse, the fragility result of On The Fragility of Benchmark Contamination Detection in Reasoning Models shows that after even one epoch of GRPO training, LiRA's detection AUROC falls from 89.1% to roughly 50–58% — near random guessing — because policy optimization compresses the log-probability gap between seen and unseen data. Reinforcement learning is standard in RNA design (LEARNA), so if the cloaking effect transfers, the field's memorization problem is invisible to its existing tools. Our probe replaces loss statistics with explanation statistics: for each target, we extract the constraint set the design model's behavior implies (via the reconciliation machinery) and test whether those constraints are predictive and internally consistent, or vacuous — the signature of a model that recalls a motif without being able to justify it.
Baselines
The do-nothing reference point is the current behavior of the full pipeline: a design model such as gRNAde (gRNAde: Geometric Deep Learning for 3D RNA inverse design) proposes a sequence, and when the user's foil is infeasible the system returns nothing — the silent unsolvable outcome documented for the contrastive framework of Benjamin Krarup et al., which offers no causal attribution for infeasible foils. A second naive reference is the "soliloquy" dump: raw internal traces and attention weights, which Angus Nicolson et al. showed increase cognitive load rather than reduce it. Both run at zero additional cost on identical targets; beating them means any nonzero rate of validated causal attributions, where validation is defined below.
The strongest existing method on the explanation question is the XAIP-as-a-Service contrastive framework itself (Benjamin Krarup et al.): it compiles a user's foil into model restrictions and produces a hypothetical plan for side-by-side comparison. We will reimplement its compilation against our RNA knowledge base so it can run at all in this domain, give it the same SAT solver budget and the same foil set, and compare on three measures: coverage (fraction of infeasible foils where it produces any output, versus silent failure), counterfactual validation rate (fraction of explanations whose flagged constraint, when relaxed, actually renders the foil foldable under a forward folder), and explanation size. Beating it means attributing infeasibility on exactly the cases where it returns unsolvable. On the design side, the strongest competitor for sequence quality is ATL-FGD, which reports a 45.49% sequence recovery rate on the RDesign benchmark split versus RDesign's own 41.53% (Learning Adaptive Topology with FiLM-Guided Distillation...); we use it unchanged as the design backbone so that explanation quality is compared on top of the best available sequences, trained with the same data, distillation, and tuning budget as reported. For the contamination probe, the strongest baseline is LiRA-style loss-based detection, which catches SFT contamination at 89.1% AUROC but collapses to 50–58% after GRPO (On The Fragility of Benchmark Contamination Detection...); we run it on our RNA models with the same splits to confirm whether the cloaking effect transfers, and our probe must beat its post-RL AUROC to count.
The key ablation isolates the hitting-set ingredient: the same knowledge base and the same foils, but with minimal correction sets replaced by greedy single-constraint removal, and separately, with the discrete tokenization replaced by raw continuous coordinates ranked by model confidence. If greedy removal matches the exact minimality computation in validation rate at comparable cost, the formal machinery is not earning its place and the claim should be weakened.
Why not just…?
Why not just filter the data harder with sequence-identity clustering? Because the file's own evidence says it is insufficient twice over. The comprehensive RNA benchmark explicitly shows that RNAs with low sequence similarity can share near-identical tertiary topologies, so sequence clustering alone leaks (A Comprehensive Benchmark for RNA 3D Structure-Function Modeling), and RDesign shows models from naive splits degrade sharply on structurally distinct evaluation sets like RNA-Puzzles (Cheng Tan et al.). More fundamentally, no amount of student-side filtering touches a foundation teacher that already saw the test families, and filtering says nothing about whether RL training is hiding what was memorized. Filtering stays in our protocol — we use the existing leakage-free splits — but it cannot answer either of our questions.
Why not just use saliency maps or attention weights as the explanation? Because the questions that matter here are contrastive and counterfactual, and attribution methods answer neither. Saliency shows where a model looked on the prediction it made; it cannot say why a rejected alternative conformation is infeasible, which is what 89.9% of real user questions ask (Benjamin Krarup et al.). Diffuse heatmaps over nucleotides also raise cognitive load without reconciling the user's assumptions (Angus Nicolson et al.), whereas a minimal correction set names a specific violated constraint that can be tested by relaxing it. We will still run an attribution baseline in the experiments; we expect it to be undefined or uninformative precisely on the infeasible-foil cases, and that comparison is part of the evaluation.
Why not just run LiRA or Min-K% loss detection on the RNA models? This is the most serious alternative, and we treat it as a real threat rather than arguing it away. The fragility paper proves loss-based detection collapses after GRPO in text models (On The Fragility of Benchmark Contamination Detection...), and it is genuinely possible that no detector, ours included, survives in RNA. The difference in kind is that our probe reads the model's implied constraint knowledge, not its output probabilities, so there is no log-probability gap for policy clipping to compress — but whether RL-trained RNA designers cloak constraint knowledge too is an open empirical question, and experiment three is designed to settle it head to head on the same checkpoints.
Why not just use physics — Rosetta or molecular dynamics — to check feasibility? Physics tools are the right adjudicator and we use them exactly that way, as ground-truth checkers for counterfactual validation. As competitors for the design-and-explain loop they are too slow and too weak: Rosetta needs hours per design and reached 45% sequence recovery on the Das et al. benchmark and a 35% wet-lab success rate on OpenKnot, versus gRNAde's sub-second designs at 56% recovery and 50% wet-lab success (gRNAde), and exhaustive physical search over multi-state conformational landscapes is intractable. Our system calls a forward folder once per candidate explanation; a physics-only baseline would call it thousands of times per foil.
Proposed Experiments and Decisive Checks
The first experiment validates the discretization on RNASolo structures: reconstruct targets from their token-level constraints alone and measure TM-score and RMSD, using RiboSphere's reported reconstruction accuracy (RMSD 1.25 Å, TM-score 0.84) as the reference scale for what a good discrete round trip looks like (RiboSphere). The second experiment builds the infeasible-foil benchmark: from held-out RNA-Puzzles and Das et al. structures we construct foils by perturbing base-pair geometry, demanding alternative pseudoknot topologies, or forcing long-range contacts that the torsion and steric constraints forbid, with feasibility adjudicated by forward folders (RhoFold, ViennaRNA) and, for the subset where they exist, wet-lab-validated cases such as the Eterna OpenKnot set. The primary metric is the counterfactual validation rate — the fraction of explanations whose single flagged constraint, when relaxed, makes the foil foldable — plus explanation size and latency relative to naive iterative search, where ALG2's two-orders-of-magnitude speedup over its predecessor is the precedent (Stylianos Loukas Vasileiou et al.). No reported result exists for explanation quality on infeasible RNA structural foils — no reported result found on Lacuna for this task — so establishing the metric is itself a contribution.
The third experiment runs the contamination audit. We train a gRNAde-style student with and without RiNALMo-style distilled representations on splits where the teacher's pre-training included the test families, and measure whether the reported out-of-distribution gains (the pattern behind ATL-FGD's 8.21% improvement on Rfam) concentrate in teacher-seen families while the extracted constraint knowledge there is vacuous or inconsistent. We then train an RL-based designer in the LEARNA style and run both LiRA-style loss detection and our explanation-consistency probe on the same checkpoints; the decisive comparison is post-RL AUROC, where the text result predicts 50–58% for loss-based detection. The fourth experiment follows DyVal's template of dynamically generated tasks with tunable complexity (DyVal), which found accuracy drops of 23% from the easiest to hardest generated difficulty — we procedurally generate RNA-like structural graphs with controllable pseudoknot depth and long-range contact counts, impossible to have memorized, and measure the explanation gap between static and generated targets as a memorization estimate.
The decisive checks are sharp. If the reconciliation system achieves a high counterfactual validation rate on procedurally generated targets — where memorization is impossible by construction — at rates comparable to held-out native targets, the explanation mechanism is real biophysical reasoning rather than recall. If explanation-consistency detection beats post-RL loss-based detection on the same RL-trained checkpoints, the cloaking problem has a structural-biology answer. And if explanations turn out to work only on families present in the teacher's pre-training, that is itself the finding: the field's headline numbers, and its explanations, are being driven by what its foundation models memorized.
References
- Plan Explanations as Model Reconciliation: Moving Beyond Explanation as Soliloquy
- On The Fragility of Benchmark Contamination Detection in Reasoning Models
- A Comprehensive Benchmark for RNA 3D Structure-Function Modeling
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design
- RiboSphere: Learning Unified and Efficient Representations of RNA Structures
- RNA-FrameFlow: Flow Matching for de novo 3D RNA Backbone Design
- RNA-Puzzles Round III: 3D RNA structure prediction of five riboswitches and one ribozyme
- On Exploiting Hitting Sets for Model Reconciliation
- A Survey on Model Repair in AI Planning
- Contrastive Explanations of Plans through Model Restrictions
- Learning Adaptive Topology with FiLM-Guided Distillation for Tertiary Structure-Based RNA Design
- Learning to Design RNA
- gRNAde: Geometric Deep Learning for 3D RNA inverse design
- The human factor in explainable artificial intelligence: clinician variability in trust, reliance, and performance
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks