WorksComputer Science
Emergent Steering Beyond Endpoint Alignment in Chemical Reaction Models
Abstract
Machine learning models for chemical reaction prediction are usually evaluated by endpoint accuracy, but endpoint correctness alone does not reveal how a model represents chemical transformation or why it fails. We introduce UniRxnAxis, a unified forward-retrosynthesis framework that models both tasks as direction-conditioned endpoint redistribution in a shared latent space. By aligning source states, transformed states, and encoded target endpoints, UniRxnAxis enables directional auditing of learned latent updates. This shared latent geometry allows the learned displacement (from the encoder source state to the decoder-facing transformed state ) to be decomposed into an endpoint-parallel component and an endpoint-orthogonal residual, which we call emergent steering. Our evaluation shows that emergent steering is functionally important: in retrosynthesis, it recovers most of the model's utility, remains effective under random, shuffled, and prototype controls, and directly affects which bond edits are promoted or suppressed. The residual-over-parallel ordering persists across alignment strengths, after decoder-interface calibration, against angle-matched random axes, and in a second graph-edit architecture, although its magnitude depends on the reference axis and optimization. The decomposition also supports a proxy taxonomy of model-side failures, showing how reaction models can be audited beyond endpoint accuracy.
Introduction
Machine learning models have been widely applied to chemical reaction modeling by translating between molecular representations, such as SMILES strings or molecular graphs [7, 5] (Wan et al. 2022; Shi et al. 2020; Zhong et al. 2023). In this paradigm, models map reactants to products for forward reaction prediction, or products to reactants for retrosynthesis. In both cases, the central goal is to model the underlying chemical transformation between the input and output molecules, including which bonds are broken, which bonds are formed, and how atoms are rearranged during the reaction. On standard benchmarks, forward reaction prediction has achieved high accuracy (measured by comparing model outputs with the recorded ground truth) [1, 17]. In contrast, retrosynthesis remains inherently ambiguous [11, 3] (Tu and Coley 2022): multiple valid reaction routes can lead to the same product, yet datasets typically record only a single target, making exact-match evaluation underrepresent true performance.
Does this apparent progress mean that machine learning models are ready for routine use by chemists? Not quite. Despite strong benchmark results, these models are still largely perceived as black boxes. Chemists are often reluctant to rely on predictions that lack transparency or cannot be interrogated in a scientifically meaningful way. Importantly, this limitation goes beyond standard notions of interpretability, which often rely on post hoc analyses such as attention weights, gradient-based attribution, or input perturbations [20, 13]. While useful, these methods do not fully capture the underlying transformation process.
For practical adoption in chemistry, models must provide more than endpoint correctness. Concretely, they should expose audit handles that allow chemists to inspect promoted or suppressed bond edits, assess whether computation concentrates near plausible reaction centers, and diagnose where transformations fail. In this paper, we develop models that support such auditing, enabling chemists to interrogate and validate predicted transformations rather than passively accept final endpoints.
Since many models use an encoder–decoder framework, we study chemical transformations in latent space. Given an encoded molecule , corresponding to either reactants or a product, the model maps it to a transformed representation , which is then decoded into the output molecule (product in forward prediction or the reactants in retrosynthesis). Let . We can ask whether can be decomposed into chemically meaningful components whose roles can be tested through decoder behavior, bond-edit decisions, and reaction-center localization.
We propose to decompose into two components: a target-aligned component that captures the intended chemical transformation, and a complementary orthogonal component that captures emergent steering:
where explains the coarse commitment motion, and represents additional latent-space shifts that may steer the model toward chemically plausible sample-specific outcomes, conceptually shown in Figure 1. To define the decomposition, we use the encoded ground-truth endpoint as a reference. For a source state , let be the supervised endpoint direction. We define as the projection of onto , and as the remaining component orthogonal to .
Figure 1. Directional decomposition and functional asymmetry. (Left) The model update is decomposed into endpoint-parallel content and an orthogonal residual relative to . (Right) In retrosynthesis, endpoint-parallel decoding collapses () to 10% top-1 accuracy, while the orthogonal residual () preserves nearly full utility (63.0% vs. 64.4%).
To make these latent quantities comparable and the decomposition measurable, we introduce UniRxnAxis: Unified Reaction Modeling with Directional Latent Axes, which is a joint forward–retrosynthesis framework with bidirectional endpoint alignment. UniRxnAxis uses a shared encoder to place reactant and product endpoints in a common latent coordinate system, a direction-conditioned transition module with task adapters to distinguish forward prediction from retrosynthesis, and a source-conditioned residual decoder to reconstruct the target endpoint from the source state and learned latent update. This design serves not only endpoint prediction, but also auditability: it makes the source state , transformed state , and encoded target endpoint directly comparable. As a result, defines a target-reference axis for analyzing the realized update .
Our evaluation shows that recovers most retrosynthesis utility, survives control substitutions, and influences which bond edits are promoted or suppressed. Because the encoded target is itself weakly decoder-executable, part of this asymmetry reflects the choice of reference axis; the gap shrinks but does not close when the axis is made more executable (Sec. 4.3). As uses the ground-truth endpoint, emergent steering is an audit-time diagnostic, not an inference-time control.
To summarize, our contributions are:
We introduce UniRxnAxis, a unified forward–retrosynthesis framework that formulates both tasks as direction-conditioned endpoint redistribution in a shared latent coordinate system.
We propose a directional decomposition of the learned latent update into an endpoint-parallel component and an endpoint-orthogonal residual (emergent steering), providing an audit handle for model-side reaction steering.
Extensive evaluation shows that emergent steering is functionally important: preserves most retrosynthesis performance, remains robust to control substitutions, and modulates bond-edit decisions. We characterize how the effect depends on the reference axis, the alignment weight, and the decoder interface, and reproduce its qualitative ordering in a second architecture (Graph2Edits). The same decomposition also enables a proxy taxonomy of model-side failures.
Related Work
Reaction Prediction and Retrosynthesis. Reaction prediction and retrosynthesis are commonly formulated as encoder–decoder problems. Sequence-to-sequence [30, 31, 14, 27, 24, 23, 18] (Tetko et al. 2020; Wang et al. 2020; Zhang et al. 2021; Wan et al. 2022), graph-to-graph [16, 7, 10, 28, 1, 26, 5] (Shi et al. 2020; Somnath et al. 2021; Zhong et al. 2023; Zhao et al. 2025; Yao et al. 2024), template-based [32, 6, 8, 4] (Xie et al. 2023), template-free [23, 18, 14] (Tetko et al. 2020; Tu and Coley 2022; Wan et al. 2022), and hybrid systems [Tu and Coley, 2022] have all made this framing effective: given one side of a reaction, the model is trained to recover the recorded other side. Prior methods have sought to improve latent alignment between input and output molecules, primarily as a means to enhance endpoint prediction accuracy [2, 19, 21] (Shen and Zhang 2026). In contrast, our work treats latent alignment as a tool for analyzing the transformation itself. Prior bidirectional models use forward and retrosynthetic signals mainly to improve reactant–product mapping at the molecular level [24, 22] (], Xu et al. 2025). In contrast, we use bidirectional modeling not only for prediction, but also to characterize latent steering.
Interpretability of reaction models. Typical post hoc probes of trained reaction-translation models have been used to explain the predictions. For example, Integrated Gradients is employed to attribute the predictions of a GCN-based retrosynthesis [15]. Similarly, T5-Chem [24] uses SHAP [25] values to explain the importance of functional groups in reaction tasks. Also, attribution analysis is employed in interpreting the trained transformer’s dependency on input tokens of a reaction. [20] These methods are useful for explaining which input features or molecular regions influence a prediction, answering which input features contributed to this output. However, our goal is different: UniRxnAxis audits the model’s internal transformation process and identify which part of that process is decoder-executable. Moreover, our work is distinct from standard mechanistic interpretability. In chemistry, “mechanism” usually refers to arrow-pushing steps, reactive structures, or isolable intermediates, and we do not claim to recover these directly. Instead, we study the model mechanism: a functionally consequential organization of latent computation that is aligned with reaction-relevant structure.
Methodology
Problem formulation: reaction as endpoint redistribution. We formulate reaction modeling as role-conditioned endpoint redistribution. Molecular reactions transform one molecular graph into another through bond formation, bond breaking, and atom rearrangement. Thus, forward prediction maps the reactant endpoint to the product endpoint, while retrosynthesis maps the product endpoint to the reactant endpoint. Let an endpoint molecular graph be , where denotes atom slots, bond orders, formal charges, and aromaticity labels. The model therefore learns a direction-conditioned map
Rather than reconstructing the target graph entirely from scratch, UniRxnAxis parameterizes this map as a source-conditioned redistribution of graph structure and atom states. For bond prediction, UniRxnAxis uses a residual structural form:
while atom-level target properties are predicted from the decoder representation. This mirrors the chemical fact that many reactions preserve a large scaffold and concentrate the difficult prediction on sparse bond and state changes.
UniRxnAxis Framework
UniRxnAxis consists of three coupled components, as shown in Figure 2. First, a shared molecular encoder maps each discrete reaction endpoint into a continuous latent state. Second, a shared direction-conditioned transformation moves the encoded source toward the encoded target endpoint. In our implementation, this transformation is factorized into a module followed by a lightweight direction-conditioned adapter . Third, a source-conditioned residual decoder combines the original source encoding with the predicted latent displacement to reconstruct the target graph.
Figure 2. Overview of the UniRxnAxis architecture. It jointly models forward prediction and retrosynthesis with shared endpoint encoders, a shared direction-conditioned attention-style update module , a direction-conditioned adapter , and a source-conditioned residual decoder. The encoded target endpoint supplies the latent alignment reference used during training and audit.
In short, the pipeline is , and is used only as a training and audit reference, never as an inference-time input. The same architecture is used for both forward prediction and retrosynthesis. Rather than training separate models, UniRxnAxis conditions the , , and decoder on a learned direction embedding , where . Thus, forward and retrosynthetic examples share the same encoder and transformation machinery, while the reaction direction is specified through learned conditioning.
Component 1: shared endpoint encoder. Given a source endpoint graph , the shared molecular encoder produces a latent embedding
The same encoder also encodes the target endpoint during training and audit:
The source and target endpoints are therefore represented by the same encoder and occupy the same model-native latent coordinate system.
Component 2: direction-conditioned latent transformation. The direction-conditioned transformation is implemented as
where produces an intermediate updated state and maps it to the final direction-adapted state , which is trained to align with the encoded target endpoint .
Component 3: source-conditioned residual decoding. The decoder reconstructs the target graph as:
where . Here, denotes feature concatenation, so fuses the original source state with the predicted latent displacement , conditioned on the direction embedding . In implementation, the bond decoder predicts residual edits relative to the source graph rather than absolute target bonds from scratch. Let denote the source bond-order matrix extracted from . Then
Additional decoder heads predict the remaining target-side atom states, such as formal charges, aromaticity labels, and, in retrosynthesis, atom types for dummy slots.
Training with Dual Endpoint Constraints
UniRxnAxis is trained with two coupled constraints: structured endpoint prediction and latent endpoint alignment. First, for each direction , the structured endpoint prediction minimizes:
Here, , , and supervise bond orders, formal charges, and aromaticity labels, respectively. The is used only in retrosynthesis, where product-to-reactant decoding must fill dummy atom slots on the reconstructed reactant side. Each term is implemented as a masked cross-entropy loss over its corresponding discrete target labels.
Second, UniRxnAxis aligns the transformed source state with the encoded target endpoint. During training, the target graph is encoded as , and is encouraged to approach it:
where selects valid target/active slots. The full objective is
where controls the strength of the latent endpoint-alignment constraint.
It is worth noting that the latent alignment constraint specifies where the target endpoint lies in representation space, but does not prescribe how the direction-conditioned stack transforms into the decoded target. Whether the learned update follows the direct endpoint direction, and whether deviations from that direction are functionally important, therefore remains an empirical question addressed by the directional audit introduced next.
Directional Decomposition
The dual constraints make directional auditing possible: the prediction loss trains the model to reconstruct the target endpoint, while the alignment loss defines a target direction in latent space against which the learned update can be decomposed.
For each reaction and direction, the dual constraints define two latent displacements. The supervision-implied endpoint displacement is
while the model’s realized update is
Let . We decompose into an endpoint-parallel component and an endpoint-orthogonal residual:
The endpoint-parallel counterfactual state is then
We interpret as endpoint-aligned attraction: the part of the learned update that points toward the encoded target endpoint. In contrast, we call emergent steering: the endpoint-orthogonal part of the update that is produced by the model, is not directly specified by the alignment loss, but may still influence decoder decisions. The decomposition is relative to the reference axis , which need not be decoder-executable and requires the ground-truth endpoint; is thus an audit-time quantity (see Sec. 4.3). This interpretation is falsifiable. If is only noise, random replacements or shuffles should preserve decoding. If it carries a real steering signal, then removing, replacing, or scaling it should change decoder decisions, and its mass should align with reaction-relevant structure. Section 4 therefore tests whether emergent steering is functionally necessary, chemically organized, and distinct from random or shuffled alternatives.
Experiments
Metric conventions. In this section, all audits use USPTO-50K Schneider et al. [ [29] data following standard splits. Code is available at https://github.com/yilish/UniRxnAxis. Reaction classification means predicting the USPTO-50K reaction-class label (10 classes) from a frozen latent representation with a lightweight linear probe. A bond edit is any atom-pair bond whose order changes between the source and target molecules; formed bonds are the subset absent in the source and present in the target, while broken bonds are present in the source and absent in the target. Edit, formed, broken, and center F1 compare decoded predictions against these bond-edit tensors, with center atoms defined as atoms incident to any bond edit. Bond-logit probes use the model’s continuous bond logits before rounding into bond orders.
We organize the evaluation around six questions.
Q1: Is the audited model a valid substrate? Before interpreting internal latent directions, the model must first be an effective endpoint predictor for both forward prediction and retrosynthesis.
Q2: Does endpoint alignment explain the executable update? After evaluating decoding on (in Q1), we decode the endpoint-parallel state , and the residual-only state . If performs poorly while the residual-only state recovers most of the full performance, endpoint attraction is not a sufficient decoder-executable explanation of the learned update.
Q3: Is the residual structured and sample-specific rather than an artifact? To test whether the residual’s utility is an artifact of neural networks, we replace with norm-matched random directions, token/sample shuffles, same-class shuffles, and class prototypes.
Q4: Does the residual act through specific decoder bond decisions? We inspect bond logits before final discretization, including boundary crossing, target-error reduction, and local sensitivity. This tests whether acts on specific bond edits rather than merely moving the latent state into a globally decoder-preferred region.
Q5: Is residual steering organized around reaction-relevant structure? We test whether update components and residual-heavy atoms are related to reaction type, reaction-center enrichment, and atom-level suppress/rescue effects.
Q6: Can the decomposition support model-side failure diagnosis? Finally, we use the decomposition to build a proxy taxonomy that separates endpoint symptoms from latent-stage diagnostic hypotheses. As a narrow proof of concept, we also test whether residual dosage exposes local correction headroom for a subset of near-miss failures.
Q1: Is the audited model a valid substrate?
Table 1 reports the benchmarking result under the standard USPTO-50K retrosynthesis, and select representative baselines across template-based, semi-template-based, template-free, unified reaction-model, and large-pretraining paradigms. The joint UniRxnAxis checkpoint is competitive on USPTO-50K, reaching 64.4% retrosynthesis. On forward prediction tasks, our model reaches Top-1 accuracy and 88.3% under joint training and 88.7% under forward only training settings.
| Paradigm | Model | Top-1 |
| Template-based | GLN [8] | 52.5 |
| LocalRetro [4] | 53.4 | |
| Semi-template-based | RetroPrime [Wang et al., 2021] | 51.4 |
| Graph2Edits [Zhong et al., 2023] | 55.1 | |
| Template-free | Graph2SMILES [Tu and Coley, 2022] | 52.9 |
| NAG2G [Yao et al., 2024] | 55.1 | |
| EditRetro [12] | 60.8 | |
| Unified | T5Chem [24] | 46.5 |
| RXNGraphormer [Xu et al., 2025] | 51.0 | |
| Pretraining | RSGPT [9] | 63.4 |
| Ours | UniRxnAxis (retro only) | 61.7 |
| UniRxnAxis (joint) | 64.4 |
Table 1. USPTO-50K retrosynthesis top-1 accuracy with reaction class unknown.
Directional audit also requires both reactant-side and product-side endpoints to exist as meaningful latent states. Joint training of both forward and retro direction enables better semantic representation across all settings. Reaction-class classification proves this pattern across all tasks (Table 2, top).
| Endpoint role | Forward-only | Joint | Retro-only |
| Product as / forward target | 0.4658 | 0.7731 | 0.7299 |
| Product as / retro source | 0.5674 | 0.7257 | 0.6846 |
| Reactant as / forward source | 0.7103 | 0.8243 | 0.6842 |
| Reactant as / retro target | 0.6950 | 0.8172 | 0.6532 |
| Setting | Object | AUC | R@1 |
| Split direct | 0.5131 | 0.0000 | |
| Split-aligned | 0.9404 | 0.5574 | |
| Split-aligned | 0.9437 | 0.4068 | |
| Split-aligned | 0.9472 | 0.3393 | |
| Joint UniRxnAxis | 0.9796 | 0.4907 |
Table 2. Audit substrate checks: joint training improves endpoint role readout (top) and native forward–retro update retrieval (bottom).
Bidirectional training also enables a more aligned transformation geometry. We test the forward–retro retrieval, which treats one direction’s latent object as a retrieval query and asks whether the paired opposite-direction object is nearest among test examples. AUC and R@1 therefore measure whether forward/retro transformation objects are directly comparable in the same coordinate system. Without alignment, split forward/retro checkpoints leave update objects near chance. Post-hoc split-space alignment can recover substantial similarity, but it requires an extra learned map; joint UniRxnAxis exposes comparability directly in native coordinates (Table 2, bottom). Thus bidirectional training appears to organize the decomposed transformation component in a shared audit space rather than only making static endpoint states alignable.
Q2: Does endpoint alignment explain the executable update?
Minimizing (eq. 10) will move the model update toward the direct endpoint direction .
Table 3 reports a stronger relevance of and in retrosynthesis than in forward prediction. However, the large orthogonal fraction (mean ) shows that much of the update lies outside the endpoint direction. The target encoding thus acts mainly as a reference axis around which we measure the learned motion.
| Dir. | Mean | Mean |
| Forward | 0.3282 | 0.9322 |
| Retro | 0.6273 | 0.7405 |
Table 3. Per-sample Geometry metrics.
Geometry alone does not tell us whether the endpoint-parallel motion is useful to the decoder. We therefore decode the endpoint-parallel state . This intervention keeps only the part of the learned update explained by the target-reference direction and removes the endpoint-orthogonal residual . If endpoint alignment were sufficient for the transformation, should retain most of the full decoding performance.
In retrosynthesis, however, this is not the case. Table 4 shows that the endpoint-parallel state reaches only 0.0997 top-1 accuracy, far below the full transformed state at 0.6441. Endpoint-directed motion alone is thus not a sufficient decoder-executable retro program; the decoder needs update content beyond .
| State | Top-1 | Edit | Formed | Center |
| 0.0405 | 0.3832 | 0.0684 | 0.4674 | |
| 0.0997 | 0.4998 | 0.1927 | 0.6046 | |
| 0.6303 | 0.7999 | 0.8761 | 0.9371 | |
| 0.6441 | 0.8868 | 0.8868 | 0.9418 | |
| 0.1110 | 0.5082 | 0.2071 | 0.6138 |
Table 4. Retro latent-state decoding. Residual-only () nearly matches full , while and remain weak.
The audit geometry is bidirectional, but the two directions are not equally diagnostic. Forward prediction is empirically easier, with a more compact geometry (Appendix § 7.§ 3). The high forward baseline is also consistent with the copy-scaffold inductive bias of the residual decoder being nearly sufficient for forward prediction, making forward a negative-control direction for the audit rather than the main analysis target. In contrast, retrosynthesis is a harder one-to-many inverse problem with disconnection ambiguity, and therefore provides a more sensitive setting for auditing whether the carries decoder-relevant signal. We therefore focus the analyses mainly on the retro direction. Moreover, is dominant in retrosynthesis, but is not merely decorative: when the true residual is fixed, the full state still outperforms residual-only, while anti-parallel or shuffled parallel vectors degrade performance (Appendix Table 27).
Q2 (cont.): Is the asymmetry an artifact of the reference axis or the model?
The encoded target itself decodes at only 0.1110 (Table 4), and the decoder is trained only on realized displacements . The weak could therefore reflect a reference axis that the decoder cannot execute, rather than a property of the learned update. We test this alternative in four ways; details are in Appendix § 7.§ 10.9.
Alignment strength. We retrain with under the same budget (Table 6). Geometry changes monotonically: cosine alignment rises and the residual fraction falls as grows. At the target axis becomes substantially executable (, Parallel ), and the gap shrinks to +8.27 points, yet Residual remains higher. At the residual is only 15.7% of the update norm, yet it still exceeds Parallel by 17.18 points, at the cost of lower accuracy.
| cos | Par. | Res. | Res.−Par. | ||||
| 0.001 | .738 | .649 | 18.41 | 55.74 | 60.56 | 14.88 | +37.33 |
| 0.01 | .649 | .734 | 49.69 | 57.96 | 59.62 | 49.39 | +8.27 |
| 0.1 | .741 | .627 | 9.97 | 63.03 | 64.41 | 11.10 | +53.06 |
| 1 | .414 | .869 | 23.21 | 39.92 | 52.99 | 24.97 | +16.71 |
| 10 | .157 | .934 | 11.10 | 28.28 | 49.95 | 11.10 | +17.18 |
Table 6. Retro sensitivity to (test top-1, %). . 0.1 is the audited checkpoint.
Decoder-interface calibration. With the model frozen, a label-free, zero-intercept linear map , fit on the training split to match realized displacements, raises axis decoding from 11.10% to 21.25%. Re-decomposing around still gives 52.73% Residual versus 21.25% Parallel. Fine-tuning the decoder on raises decoding to 64.45%, but also yields 10.39% with the displacement zeroed. This reveals an shortcut, so we read it as a supervised-recoverability upper bound rather than pre-existing axis utility.
Matched-random axes. For each test reaction we sample eight null axes with the same angle to as , which matches component norms and only changes orientation. The target-defined axis gives a 53.07-point Residual−Parallel separation versus 37.13 for matched axes (difference 15.94, 95% CI [14.83, 17.10]). Generic projection geometry thus contributes but does not explain the full effect. Scale mismatch is not dominant either (; decoding peaks at 19.53%).
Other architectures and datasets. We add an endpoint-alignment loss to the autoregressive graph-edit model Graph2Edits (Zhong et al., 2023) (55.44% top-1) and audit the state change across the ground-truth first edit. Its frozen edit head predicts the second edit at 16.58% from the parallel-only state versus 48.79% from the residual-only state (: 75.80%). On USPTO-MIT (Jin et al., [16]), retrosynthesis again remains more endpoint-aligned than forward prediction (, ). This transfer is geometric only; we do not test decoder executability there.
Interpretation. The gap’s magnitude depends on the reference axis, the alignment weight, and the decoder interface; its sign is stable in every setting we tested. Two factors contribute. The decoder is exposed only to realized updates, which explains part of the gap. In addition, only softly attracts toward the endpoint, while demands exact, sample-specific bond decisions that can occupy degrees of freedom endpoint attraction leaves free. Our claim is therefore not that orthogonal directions are intrinsically more useful, but that, relative to the supervised endpoint axis, the native decoder-executable computation concentrates in the residual more than for matched arbitrary axes.
Q3: Is structured and sample-specific rather than an artifact?
is not a magnitude-only artifact. A functional residual could still be an artifact if its utility came from magnitude alone. Norm-matched random controls do not recover the true residual’s performance or edit-boundary crossing (Table 5, top), which rejects this hypothesis.
| Control | Variant | Top-1 |
| Reference | Full | 0.6441 |
| 0.6303 | ||
| Norm | Global | 0.1007 |
| Per-token | 0.1022 | |
| Prototype | Class subspace | 0.2606 |
| Class NN | 0.1460 | |
| Class mean | 0.1558 | |
| Shuffle | Token | 0.0855 |
| Sample | 0.1162 | |
| Same-class | 0.1188 |
Table 5. artifact controls. The true works; norm, prototype, and shuffle controls fail.
is not template or shuffle shortcuts. A second hypothesis is that the residual may be a class-level template direction rather than a sample-specific steering signal. Class means, nearest-neighbor class residuals, and class subspace projections fall far below the true residual and the full model (Table 5, bottom). Nor is any nearby empirical residual good enough: token shuffling, sample shuffling, and same-class shuffling all collapse accuracy. Thus, the residual is therefore not a norm, prototype, or shuffle-stable artifact. Instead, it is sample-specific.
Q4: Does Delta z-perp act through specific decoder bond decisions?
The previous controls show that the residual is structured, but they do not yet show where it acts in the decoder. A remaining hypothesis is that merely moves the latent state into a globally decoder-preferred region, rather than pushing bond-edit decisions. We test this hypothesis from three decoder-facing views: boundary crossing, sensitivity, bond logit refinements.
selectively pushes bonds across decoder boundaries. We compare bond logits before and after adding the residual step from to . As shown in Table 7A, selectively pushes correct edit and formed edges across decoding thresholds, rarely pushes wrong edges across thresholds, and suppresses competing wrong edges. We quantify this using target-edge crossing, wrong-edge crossing, and competing-edge suppression. The high target-edge crossing and near-zero wrong-edge crossing indicate that emergent steering acts selectively on bond-edit decisions rather than globally increasing decoder confidence.
| A: boundary crossing | ||
| Target edit crosses | 0.5217 | |
| Target formed crosses | 0.6054 | |
| Wrong non-target crosses | 0.0004 | |
| Wrong competitors down | 0.8525 | |
| B: local derivative | ||
| on target formed | +1.0401 | |
| on target formed | +0.1839 | |
| Random on target formed | -0.0017 | |
| on competitors | -0.4104 | |
| C: bond-logit refinement | Shift | TER |
| Edge group | ||
| Target edit | 0.1926 0.3740 | 0.1611 0.5143 |
| Target formed | 0.3288 0.5129 | 0.1425 0.6721 |
| Wrong competitors | 0.5532 -0.5427 | -0.5542 0.5423 |
Table 7. Decoder-facing bond probes. crosses target thresholds, has the strongest tested local derivative, and selectively refines bond logits.
Decoder sensitivity is strongest along . The decoder’s local score sensitivity is concentrated along on the same edges where the residual improves finite decoding behavior. Finite interventions are complemented by a local derivative test. Among tested directions, has the strongest local directional derivative on target formed bonds and the strongest negative derivative on selected competing non-target edges, much larger than or a random direction (Table 7, Panel B).
selectively refines bond logits. We next test whether the boundary effects reflect targeted score correction or a broad score shift. Using pre-discretization bond logits, we measure bond-logit shift and target-error reduction (TER) for target edit/formed bonds and high-scoring false-positive competitors. The step gives coarse support by raising target-edge scores, but also increases false-positive competitors. In contrast, raises target edit/formed-edge scores more strongly while suppressing competitors, yielding the largest TER on target edges and positive TER on false positives (Table 7C). Thus, refines bond logits toward the target pattern rather than simply amplifying decoder confidence.
A scalar residual-dosage sweep also peaks near the learned residual magnitude rather than growing monotonically (Appendix Table 10).
Q5: Is residual steering organized around reaction-relevant structure?
If the residual is part of reaction computation, it should reflect reaction-level organization, not only endpoint accuracy. Stage-wise probes show that reaction-class readability increases from early encoder layers to , then decreases after the update. In retrosynthesis, however, the decomposed components and are more class-readable than the raw update (Appendix Figure 4). This suggests a program-to-refinement pattern: the encoder builds a class-readable reaction program, while the update converts it into an instance-specific bond-edit state for decoding. Appendix § 7.§ 10.3 includes a qualitative joint forward/retro t-SNE visualization of the same update components, used only as supporting semantic visualization rather than primary evidence.
As supporting chemistry-facing evidence, atom-level diagnostics show that update components are enriched around reaction centers. Under the raw-anchor convention, carries the largest retro reaction-center mass (0.4918), while remains enriched over the label fraction (0.3255 vs. 0.1596). The residual is thus not the most localized component: already points to the reaction center, while selects among nearby candidate edits (Table 7). Residual-heavy atoms are also functionally tied to decoding: suppressing top residual atoms strongly degrades formed-bond F1, and rescuing them recovers part of the loss (Appendix § 7.§ 10.4). These supporting evidences suggest that residual structure is chemistry-facing.
Q6: Can the decomposition support model-side failure diagnosis?
UniRxnAxis provides two-level failure labels. Separating from makes failure diagnosis a two-level question: what is wrong in the decoded endpoint, and where along does the transformation become unreliable? We therefore build a hierarchical, proxy-based taxonomy with endpoint symptoms on the output side and latent-stage diagnostic hypotheses on the model side. We present the taxonomy in Appendix § 7.§ 10.7 in detail.
Both levels use five matched categories: program (parallel), refinement (residual), dosage, nonresponsive/off-manifold, and decoder-realization. Among 1782 wrong retro test samples, the merged endpoint taxonomy assigns a non-mixed label to all samples, and the latent taxonomy assigns a non-mixed label to 96.6% (Appendix Figure 3A–B).
Off-diagonal audit provides insight into models’ internal behavior. The endpoint–latent crosstab first acts as a sanity check: the conceptual diagonal carries substantial mass (mean available row fraction 0.650), so the taxonomy recovers obvious stage correspondences when endpoint symptoms and latent traces agree. The most informative cases are off-diagonal: the endpoint symptom says one thing, but the latent trajectory suggests another. The largest such pattern is endpoint refinement/edge-selection errors mapping to latent program/parallel failures (503/1217, row fraction 0.413), rather than to latent residual refinement failures (433/1217, row fraction 0.356). Some apparent final competing-edge mistakes may thus originate in an atypical or undercommitted program state before residual refinement. Appendix Figure 3D shows one such case: the final output has a local-looking edge/refinement symptom, while the decoded state and proxy trace point upstream to program formation.
On residual-dosage headroom. The failure taxonomy suggests a broader implication: the residual is not only a diagnostic signal, but also a place where the model may already contain recoverable solutions. We choose one simple one-dimensional proof of concept: residual dosage. If some wrong samples fail mainly because the residual amount is slightly under- or over-applied, can a different residual dosage recover them? We first measure the headroom with an oracle over the fine grid for . This oracle identifies 247 failures where that could be corrected by changing only , raising official accuracy from 0.6441 to 0.6934. When test a simple metric-gated selector: keep by default, apply dosage changes only to samples whose decoded proxy signals look dosage-sensitive, and choose the alternative from the fine-grid proxy score (Appendix § 7.§ 2). The best conservative rule reaches 0.6529, rescuing 46/247 oracle-rescuable failures (18.6%) while breaking 2/3225 alpha-1 correct samples (0.06%) (Table 8).
| Run | Fixed | Oracle | Oracle-rec. | Selector | Rescue | Break |
| Retuned fine grid | 0.6441 | 0.6934 | 247 | 0.6529 | 46 (18.6%) | 2 (0.06%) |
Table 8. Fine-grid residual-dosage rescue on the retro test split.
The claim is not that scalar solves retrosynthesis, but that the audit exposes a narrow subset of near-miss failures whose decoded outcome changes when the residual is dosed differently. Because itself requires the ground-truth endpoint, this is not yet a deployable corrector.
Conclusion
Endpoint correctness is necessary for reaction modeling, but it is not enough to make a model scientifically inspectable. UniRxnAxis aligns both sides of the reaction in a shared latent geometry, which enables decomposing the learned update into endpoint-parallel attraction and endpoint-orthogonal emergent steering. Emergent steering preserves most retrosynthesis utility, survives control substitutions, and modulates bond-edit decisions; its advantage shrinks but persists when the reference axis is made more executable. Reaction models should thus be evaluated not only by their endpoints but also by how their latent transformations organize chemical change.
Limitations. The decomposition uses the ground-truth endpoint, so emergent steering is an audit-time diagnostic; the gap’s size depends on the reference axis, , and the decoder interface, and only its sign is stable in our tests. Executable evidence covers retrosynthesis on USPTO-50K with two architectures, with geometric transfer only on USPTO-MIT, and we report deterministic top-1 only. Reaction-center and failure labels are unvalidated automatic proxies.
Acknowledgments and Disclosure of Funding
This work was supported by the National Science Foundation under the NSF Center for Computer Assisted Synthesis (C-CAS), grant number CHE-2202693.
Competing interests. The authors declare no competing interests.
Appendix
Audit State and Metric Definitions
Table 9 summarizes the latent states used in the counterfactual audit. The decomposition states depend on , the encoded ground-truth endpoint latent, and are therefore analysis-time states rather than deployable inference-time states. The implemented endpoint chain is . Thus is the raw encoded source, while is the transformed source state aligned to . The encoder is shared and direction-agnostic; the update module, adapter, and decoder interface all receive .
| State | Definition | Uses GT endpoint? |
| Raw encoded source endpoint, | No | |
| Shared attention-style updated state before task adaptation | No | |
| Natural transformed source state, aligned to the encoded target | No | |
| Encoded ground-truth endpoint latent | Yes | |
| Endpoint direction used for projection | Yes | |
| Model update displacement | No | |
| Projection of onto | Yes | |
| Yes | ||
| , program-stage counterfactual | Yes | |
| Residual-only | Yes | |
| Yes |
Table 9. Audit state definitions.
Latent tensors are represented internally as token states . We use two aggregation domains. For active-target graph-level readouts, let when token is active in the update window and valid on the target endpoint, matching the endpoint-alignment mask . The graph readout is the masked mean
For encoder-layer stage probes that do not use a target endpoint, the same masked mean is applied with the source-side raw valid-token mask. Decoder counterfactuals, atom localization, edge-score probes, and dosage interventions do not use this pooled vector: they keep the token state, apply the decomposition under the active-target mask, and pass the resulting token tensor through the frozen decoder.
Residual-Dosage Selector Details
Residual-dosage sanity check. In addition to the main decoder-facing bond probes, we test whether the residual direction behaves like a calibrated correction when only its scalar magnitude is changed. Starting from , we decode for a small alpha grid. Performance rises sharply from the endpoint-parallel state, peaks near the learned residual scale, and mildly declines when over-dosed. We treat this as supporting evidence for calibrated residual correction, not as the primary mechanistic claim.
| Top-1 | Edit F1 | Formed F1 | Center F1 | |
| 0 | 0.0997 | 0.4998 | 0.1927 | 0.6046 |
| 0.25 | 0.4590 | 0.8055 | 0.7448 | 0.8819 |
| 0.5 | 0.6369 | 0.8881 | 0.8868 | 0.9420 |
| 1 | 0.6441 | 0.8868 | 0.8868 | 0.9418 |
| 2 | 0.6345 | 0.8852 | 0.8849 | 0.9410 |
Table 10. Retro dosage response for . The residual behaves like a calibrated correction rather than an unbounded force. Top-1 accuracy bootstrap CIs are reported in Appendix A.6.
The adaptive-alpha analysis is evaluated only after the audit-time decomposition has constructed and the residual direction . For each test sample, we decode the fine grid with . The oracle selector reports the best grid point under label-available exact-match evaluation and is therefore only an upper bound.
The GT-free proxy rule in Table 8 uses no ground-truth exact match, edit F1, center F1, GT edge overlap, oracle alpha, or true failure label. It is GT-free only after the audit decomposition has already constructed and . The rule keeps unless the decoded alpha sweep passes proxy gates for a local dosage change; among passing candidates it chooses the alpha with the largest proxy improvement over .
The thresholds are selected by a small grid sweep on a fixed 512-sample sensitivity subset. For each candidate policy, we evaluate two label-available quantities after selection: rescued, the number of samples wrong at but correct after selection, and broken, the number of samples correct at but wrong after selection. Candidate policies are restricted to a conservative frontier with at most two broken samples on the sensitivity subset and nonzero changed-sample fraction. We rank the remaining rules by
then by official accuracy and rescued count. Thus the selected thresholds are tuned for cautious rescue, not for maximizing edit F1, center F1, or oracle agreement. The final reported rule is one of the sensitivity-selected GT-free proxy rules evaluated on the full retro test split.
| Quantity | Definition |
| Oracle alpha | For each sample, choose the alpha that maximizes official exact-match correctness; ties are broken by edit F1, center F1, distance to , and then alpha value. This is label-available and used only as an upper bound. |
| Oracle-rescuable | A sample that is wrong at but correct for at least one alpha on the grid under the oracle rule. There are 247 such retro test samples. |
| Rescued / broken | Rescued means wrong at and correct after the selector; broken means correct at and wrong after the selector. |
| Reconstruction score | Negative mean rounding error between continuous bond logits and their rounded decoded bond orders over active atom pairs. Higher means the decoded graph is closer to a stable decoder boundary. |
| Predicted-edge margin | The weakest predicted changed-edge strength minus the strongest non-changed competitor strength. Higher means the decoded edit set is more separated from competing edits. |
| Consensus count | Number of alpha values that decode to the same edit signature. Higher means the candidate output is stable across nearby dosage values. |
| Internal alpha score | , where is reconstruction score, is predicted-edge margin, is consensus count, is invalid-output penalty, is source-equal penalty, and is zero-edit penalty. The selector uses score relative to . |
| Proxy gates | A candidate alpha must have score , be no less valid than , have consensus count at least 1, and have predicted-edge margin at least . Downward dosage is allowed only when the overdrive-risk score is in the top 60%; upward dosage is allowed only when the under-refinement-risk score is in the top 60% and edit-count explosion is at most 2. |
| Risk scores | Overdrive risk is the sum of residual-norm -score, residual/perpendicular-ratio -score, and 0.5 times combined-instability -score. Under-refinement risk is the residual/perpendicular-ratio -score plus 0.5 times residual-norm -score plus 0.25 times residual-top-1-mass -score. |
| Reported selector | The reported rule has score threshold , minimum consensus 1, risk quantile 0.4, maximum program-atypicality -score 10, maximum upward edit-count explosion 2, and margin threshold . On the full retro test split, it changes 3.42% of samples, rescues 46, breaks 2, and reaches official accuracy 0.6529. |
Table 11. Definitions used by the fine-grid dosage-rescue audit.
| Audit family | Latent domain | Notes |
| Endpoint role readout | Graph-level pooled | Cached pooled or with linear ridge probes |
| Update-object alignment and endpoint geometry | Graph-level pooled | Uses , , and |
| Stage/component class readout | Graph-level pooled | Encoder stages use raw valid-token pooling; decomposition components use active-target-mask pooling |
| Prototype, nearest-neighbor, subspace, and program-centroid analyses | Graph-level pooled | Class means/centroids are fit in pooled or space |
| Functional state decoding, residual-only, and dosage | Token-level tensor | Constructs , residual-only, or in before decoding |
| Boundary crossing, TER, and Bond/ | Token/edge-level | Decoder scores are edge-level; directions are token tensors under the active-target mask |
| Norm controls and token/sample shuffles | Token-level tensor | Random or shuffled residual tensors are decoded; summaries report sample means |
| Atom center mass, gradients, suppress/rescue | Atom/token-level | Uses per-token norms/gradients over active atoms, not graph-pooled vectors |
| Failure taxonomy and adaptive dosage | Mixed derived proxies | Reuses token-level decoded states plus atom/edge/sample proxy metrics |
Table 12. Latent aggregation domains used by Section 4 audits.
Table 13 defines the main metrics used in Section 4. Unless otherwise stated, confidence intervals use sample-level bootstrap resampling; edge-level quantities are first aggregated to sample means when sample identifiers are available.
| Metric | Definition / aggregation | Caveat |
| Official accuracy | Mean deterministic top-1 exact-match correctness | Not beam search |
| Edit F1 | Sample-level F1 over edited atom pairs / bond edits | Tensor proxy |
| Formed F1 | Sample-level F1 over formed bonds | Empty-set convention follows local implementation |
| Broken F1 | Sample-level F1 over broken bonds | Changed-order edits tracked separately |
| Center F1 | Sample-level F1 over atoms incident to bond edits | Reaction-center proxy |
| Center mass ratio | Residual token-norm mass on center atoms / active mass | Model-side localization |
| Boundary crossing | Fraction of selected edges crossing rounded/clamped decode thresholds | Decoder-score proxy |
| Competing-edge suppression | Suppression of high-score non-GT candidate edges | Candidate definition is automatic |
| Target-error reduction | Reduction in continuous | Not calibrated probability |
| Bond/ derivative | Local gradient directional derivative of bond score | Model sensitivity, not physical causality |
| Program atypicality | Distance of to train-set family centroid | Correlational uncertainty proxy |
| Failure modes | Automatic endpoint/latent taxonomy labels on wrong samples | Not human chemistry labels |
Table 13. Audit metric definitions and caveats.
| Direction | Norm ratio |
| Forward | 0.9307 |
| Retro | 0.6859 |
Table 14. Supplementary endpoint-geometry aggregate omitted from the main table. The norm ratio is , a norm-weighted counterpart to the per-sample mean residual fraction reported in Table 3.
Effective-Dimension Check for Directional Asymmetry
As supporting geometry for the directional-asymmetry discussion in § 4, we compare the transformed endpoint state in forward and retro on the full USPTO-50K test split. Global metrics use the covariance spectrum of centered graph-level features; mean local intrinsic dimension uses the Levina–Bickel kNN estimator with .
| State | Effective rank | Participation ratio | Top-1/trace | Mean local ID |
| Forward | 8.21 | 3.07 | 0.5585 | 12.60 |
| Retro | 11.25 | 5.66 | 0.2996 | 13.22 |
Table 15. Effective-dimension evidence for the transformed endpoint state. Forward is more compact than retro by effective rank, participation ratio, and mean local intrinsic dimension.
These dimension metrics are used only as supporting evidence for task asymmetry: lower effective dimension does not by itself imply easier chemistry or weaker semantic content.
Artifact-Control Diagnostic Proxies
| Control | Variant | Edit F1 | Cross |
| Reference | Full | 0.8868 | 0.5217 |
| Norm-only random | Global norm | 0.5020 | 0.0163 |
| Norm-only random | Per-token norm | 0.5071 | 0.0288 |
| Prototype | Class subspace | aggregate-only | |
| Prototype | Class nearest neighbor | aggregate-only | |
| Prototype | Class mean | aggregate-only |
Table 16. Available diagnostic proxy details for the residual artifact controls in Table 5. “Cross” denotes decoder-boundary crossing as defined in § 4.
Detailed Section 4 Audit Protocols
Tables 17–19 spell out the aggregation domain used by each Section 4 audit. The important distinction is that graph-level readouts are masked means over latent tokens, while decoded counterfactuals, bond-boundary probes, dosage sweeps, and atom-localization audits keep the full token tensor until after decoder evaluation.
| Audit | Implementation family | Latent domain | Protocol / reported quantity | |
| Endpoint context | performance | Main endpoint evaluation | Decoded graph output | Reports deterministic top-1 endpoint exact match for the trained joint checkpoint; no latent pooling enters the score. |
| Opposite-endpoint role readout | role probe | Graph-level pooled | Fits linear/ridge role probes on cached pooled or vectors to test whether endpoint role information is readable. | |
| Update-object alignment | delta- analysis | Graph-level pooled | Computes , , , then reports cosine/projection/orthogonal residual geometry. | |
| Endpoint geometry | delta- analysis | Graph-level pooled | Uses the same pooled endpoint vectors to measure distance-to-target, parallel fraction, residual norm share, and related aggregate geometry. | |
| Staged decoding | selective ablation | Token-level tensor | Builds , , residual-only, and as tensors under the active-target mask, decodes each with the frozen residual decoder, and reports official/edit/formed/broken/center metrics. | |
| Target-encoder oracle | oracle decoding | Token-level tensor | Replaces the transformed endpoint state by before the same frozen decoder to test whether target encoder latents are directly decoder-executable. | |
| Functional split | selective ablation | Token-level tensor | Compares full, endpoint-parallel, and endpoint-orthogonal counterfactual states after decoding; no pooled vector is fed to the decoder. | |
| Parallel preconditioner | parallel dosage | Token-level tensor | Keeps the true residual fixed and varies the endpoint-parallel component before decoding to test whether the residual needs a coarse endpoint-conditioned state. |
Table 17. Detailed protocols for substrate, geometry, and staged decoding audits.
| Audit | Implementation family | Latent domain | Protocol / reported quantity |
| Norm-matched random controls | selective ablation | Token-level tensor | Replaces by random directions scaled to the same masked norm, decodes the resulting token state, and reports endpoint/edit proxies. |
| Token/sample/class shuffles | shuffle controls | Token-level tensor for decoding | Permutes or substitutes residual token tensors before decoding; graph-level class prototypes may be fit for selection, but the decoder intervention itself is token-level. |
| Class prototype controls | prototype controls | Mixed | Fits class mean/nearest-neighbor/subspace controls in pooled residual space, then evaluates decoded prototype-derived token controls where available. |
| Boundary crossing | bond-boundary audit | Token/edge-level | Evaluates continuous decoder bond scores on selected edges and counts threshold crossings caused by endpoint-parallel versus residual movement. |
| Score delta and TER | bond-boundary audit | Token/edge-level | Computes score changes and target-error reduction on decoder edge scores after decoding counterfactual token states. |
| Directional derivative | decoder-gradient audit | Token/edge-level | Uses masked token-space directions and decoder gradients to measure local alignment for target and competing edges. |
| Residual dosage | dosage response | Token-level tensor | Sweeps before decoding and summarizes official/edit/center response curves. |
| Audit | Implementation family | Latent domain | Protocol / reported quantity |
| Encoder-stage class readout | layerwise readout | Graph-level pooled | Pools each encoder layer over raw valid source tokens and fits cross-validated linear reaction-class probes. |
| Component class readout | component readout | Graph-level pooled | Pools , , , , and under the active-target mask before linear probing. |
| Program centroid / atypicality | program centroid | Graph-level pooled | Fits class/family centroids in pooled or residual space and relates distance to residual demand or failure. |
| Atom center mass | node residual audit | Atom/token-level | Computes per-token residual norm mass on active atoms and compares mass on reaction-center atoms against total active mass. |
| Atom gradients and suppress/rescue | atom/gradient audit | Atom/token-level | Uses per-token gradients, residual norms, and atom suppress/rescue interventions; no graph-level latent vector is used for the atom proxy itself. |
| Merged failure taxonomy | failure taxonomy | Mixed derived proxies | Combines decoded state metrics, dosage response, atom center mass, edge competition, and program atypicality into automatic endpoint/latent failure labels. |
| Weak alpha rescue | adaptive dosage | Token-level plus derived selector | Applies residual dosage choices to token-level decoded alpha sweeps; GT-free selectors are GT-free only after the audit decomposition has been constructed. |
Table 18. Detailed protocols for residual controls and decoder-grounding audits.
Bootstrap Confidence Intervals for Key Audit Results
Tables 20–22 report bootstrap intervals for the main sample-level claims in Section 4. We resample test samples with replacement; edge-level quantities are first aggregated to sample means whenever sample identifiers are available. Entries marked aggregate-only in the main text are not assigned invented intervals.
| Quantity | Metric | Mean [95% CI] |
| Retro | Official | 0.0405 [0.0350,0.0463] |
| Retro | Official | 0.0997 [0.0913,0.1075] |
| Retro | Official | 0.6441 [0.6305,0.6571] |
| Retro residual-only | Official | 0.6303 [0.6171,0.6437] |
| Forward | Official | 0.8694 [0.8602,0.8784] |
| Forward | Official | 0.8816 [0.8724,0.8902] |
| Forward | Official | 0.8832 [0.8744,0.8926] |
Table 20. Bootstrap CIs for staged decoding and functional split.
| Family | Quantity | Mean [95% CI] |
| Residual control | True residual official | 0.6441 [0.6301,0.6563] |
| Residual control | Random global-norm r0 official | 0.1029 [0.0949,0.1112] |
| Residual control | Random per-token-norm r0 official | 0.1051 [0.0969,0.1138] |
| Residual control | Token-shuffle r0 official | 0.0839 [0.0765,0.0907] |
| Residual control | Sample-shuffle r0 official | 0.1176 [0.1084,0.1270] |
| Residual control | Same-class shuffle r0 official | 0.1190 [0.1104,0.1280] |
| Decoder boundary | Aggregate edit residual crossing | 0.5217 [0.5081,0.5353] |
| Decoder boundary | Aggregate formed residual crossing | 0.6054 [0.5980,0.6082] |
| Decoder boundary | Aggregate wrong-edge crossing | 0.0004 [0.0004,0.0005] |
| Decoder boundary | Aggregate competing-edge suppression | 0.8525 [0.8302,0.8732] |
| Aggregate on GT formed edges | +1.0401 [+0.9134,+1.1767] | |
| Aggregate on GT formed edges | +0.1839 [-0.1566,+0.2137] | |
| Aggregate on GT formed edges | -0.0017 [-0.0085,+0.0050] | |
| Aggregate suppressing competing edges | -0.4104 [-0.5126,-0.3219] | |
| Atom localization | Retro center mass | 0.3255 [0.3210,0.3300] |
| Atom localization | Retro center mass | 0.4918 [0.4875,0.4961] |
| Atom function | Top-3 suppress formed F1 drop | 0.6112 [0.5972,0.6245] |
| Atom function | Top-3 rescue formed F1 gain | 0.2960 [0.2834,0.3081] |
Table 21. Bootstrap CIs for residual controls, decoder grounding, and atom-level organization.
| Family | Quantity | Mean [95% CI] |
| Dosage | official | 0.0997 [0.0917,0.1079] |
| Dosage | official | 0.4590 [0.4452,0.4737] |
| Dosage | official | 0.6369 [0.6231,0.6495] |
| Dosage | official | 0.6441 [0.6303,0.6561] |
| Dosage | official | 0.6345 [0.6209,0.6481] |
| Failure diagonal | Program Program/Parallel | 0.858 [0.810,0.903] |
| Failure diagonal | Dosage Dosage | 0.886 [0.771,0.971] |
| Failure diagonal | Decoder Decoder realization | 0.969 [0.906,1.000] |
| Failure off-diagonal | Refinement Program/Parallel | 0.413 [0.387,0.440] |
| Failure off-diagonal | Nonresponsive Residual failure | 0.504 [0.445,0.555] |
| Alpha selection | Fixed official | 0.6441 [0.6303,0.6581] |
| Alpha selection | Oracle alpha official | 0.6934 [0.6810,0.7062] |
| Alpha selection | Conservative selector official | 0.6529 [0.6395,0.6659] |
Table 22. Bootstrap CIs for residual dosage, failure taxonomy, and weak alpha selection.
Failure-Diagnosis Visualization

Figure 3. Two-level failure diagnosis on wrong retro samples (). Endpoint symptoms and latent-stage diagnoses agree on the diagonal, while off-diagonal cases reveal hidden program/parallel failures behind local-looking refinement errors.
Qualitative Visualization of Update-Component Semantics
Figure 5 visualizes forward and retro update components in a jointly fit t-SNE space and colors points by reaction class. The plot is consistent with the linear readout evidence in Figure 4: in full-dimensional probes, is more readable than for both reaction class (macro-F1 0.762 vs. 0.288) and bond-edit bucket (macro-F1 0.687 vs. 0.451). We use this figure only as a qualitative semantic signal, since t-SNE layouts depend on visualization choices and do not establish mechanism by themselves.
Figure 4. Reaction-class linear readouts across stages and update components. (a) Macro-F1 mostly builds through encoder layers up to E8/S and drops after the UniRxnAxis update to S. (b) Using the same 5-fold ridge-probe protocol on the active-target-mask pooled decomposition states, retro component separation increases reaction-class readability relative to the raw update, whereas forward remains slightly higher than either component.

Figure 5. Qualitative joint t-SNE visualization of forward and retro update components colored by reaction class. Coordinates are fit jointly over both directions within each component panel; circles denote forward examples and crosses denote retro examples. The visualization is appendix-only and is not used as primary evidence.
Atom-Level and Sample-Level Supporting Diagnostics
Chemistry alignment should also appear in atom-level coordinates, especially around reaction centers. Under the raw-anchor convention, retro carries the largest reaction-center mass, while remains strongly center-enriched and almost always includes a reaction-center atom among its top residual atoms (Table 23, top). Thus the atom-level result is not “residual mass is uniquely most localized”; it is that both endpoint alignment and residual refinement concentrate around chemically edited atoms, with the residual retaining localized structure after removing the endpoint-reference component.
| Reaction-center localization | ||
| Direction | Metric | Value |
| Retro | center mass | 0.4918 |
| Retro | center mass | 0.3255 |
| Retro | center mass | 0.4075 |
| Retro | Label fraction | 0.1596 |
| Forward | center mass | 0.2340 |
| Forward | center mass | 0.1526 |
| Forward | center mass | 0.1928 |
| Functional atom-level evidence | ||
| Probe | Metric | Value |
| Gradient | Center mass | 0.5897 |
| Gradient | Spearman grad norm vs residual norm | 0.4549 |
| Gradient | Top-3 gradient/residual atom overlap | 0.6491 |
| Suppress | Edit F1 drop | 0.3340 |
| Suppress | Formed F1 drop | 0.6112 |
| Suppress | Center F1 drop | 0.2641 |
| Rescue | Edit F1 gain | 0.1436 |
| Rescue | Formed F1 gain | 0.2960 |
| Rescue | Center F1 gain | 0.1581 |
Table 23. Atom-level organization of emergent steering.
Localization alone could be a visual prior, so we test whether the localized atoms matter to the decoder. Decoder gradients concentrate on reaction-center atoms and overlap with high-residual atoms. Suppressing top residual atoms causes large edit, formed-bond, and center F1 drops; rescuing those atoms restores part of the performance (Table 23, bottom). Thus residual localization is functionally connected to bond-edit computation.
Table 23: Atom-level organization of emergent steering. Top: both endpoint-parallel and residual components concentrate mass on reaction-center atoms, with the endpoint-parallel component carrying the largest center mass under the raw-anchor convention. Bottom: localized residual atoms are functionally connected to decoder bond-edit computation. Bootstrap CIs are reported in Appendix A.6.
The parallel/residual division is also visible at the sample level. If provides a coarse program, atypical coarse programs should require more residual refinement. We define program atypicality as the distance from a sample’s to the training centroid for its reaction class. Larger distance predicts a larger residual-to-parallel norm ratio and slightly worse edit and center quality, but it is not a strong predictor of final correctness or residual-stage gain (Table 24). This weakens the old “program atypicality explains residual demand” story: the coarse program geometry is a useful sample-level covariate, not a sufficient account of residual content.
| Correlation | Spearman |
| Distance vs | 0.2161 |
| Distance vs | 0.3754 |
| Distance vs edit F1 | -0.0387 |
| Distance vs center F1 | -0.0492 |
| Distance vs edit gain | -0.0738 |
| Distance vs center gain | -0.0399 |
Table 24. Program atypicality is a weak residual-demand covariate.
Table 24: Program atypicality is a weak residual-demand covariate. Values are Spearman correlations using reaction-class centroids of .
Audited UniRxnAxis Checkpoint Configuration
Table 25 lists the implementation details of the checkpoint used for all directional audits in § 4. These details are not needed to understand the method, but they define the audited model instance.
Endpoint-Parallel Preconditioning Controls
Table 26 reports the full dosage sweep used to test the conditional contribution of when the true residual steering direction is fixed. The state is ; thus is the residual-only intervention, is the full learned state, and negative reverses the endpoint-parallel component.
| Top-1 | Edit F1 | Formed F1 | Center F1 | |
| -1.0 | 0.3843 | 0.4346 | 0.2219 | 0.6276 |
| -0.5 | 0.5482 | 0.7101 | 0.5816 | 0.8242 |
| 0.0 | 0.6303 | 0.8799 | 0.8761 | 0.9371 |
| 0.25 | 0.6415 | 0.8867 | 0.8879 | 0.9416 |
| 0.5 | 0.6423 | 0.8870 | 0.8877 | 0.9413 |
| 1.0 | 0.6441 | 0.8868 | 0.8868 | 0.9418 |
| 1.5 | 0.6425 | 0.8885 | 0.8879 | 0.9429 |
| 2.0 | 0.6431 | 0.8880 | 0.8872 | 0.9428 |
Table 26. Endpoint-parallel dosage under fixed true residual steering.
Table 27 compares the true endpoint-parallel component against controls that preserve coarse statistics or empirical vectors while breaking the sample-specific endpoint-aligned direction. The true parallel component improves over residual-only, while random and shuffled controls remain below it.
| Variant | Top-1 | Edit F1 | Formed F1 | Center F1 |
| True parallel | 0.6441 | 0.8868 | 0.8868 | 0.9418 |
| Residual-only | 0.6303 | 0.8799 | 0.8761 | 0.9371 |
| Anti-parallel | 0.3843 | 0.4346 | 0.2219 | 0.6276 |
| Random global-norm parallel | 0.6275 | 0.8782 | 0.8733 | 0.9358 |
| Random per-token-norm parallel | 0.6209 | 0.8694 | 0.8598 | 0.9303 |
| Sample-shuffled parallel | 0.6089 | 0.8670 | 0.8593 | 0.9302 |
| Same-class shuffled parallel | 0.5990 | 0.8660 | 0.8579 | 0.9293 |
| Token-shuffled parallel | 0.5690 | 0.8528 | 0.8436 | 0.9254 |
Table 27. Endpoint-parallel controls under fixed true residual steering.
Failure Taxonomy Definitions
The failure taxonomy is an automatic proxy audit over retro test examples that are wrong under the natural decoded output. It is not a human chemical-error adjudication. Endpoint labels describe the decoded output symptom, while latent labels describe which part of the trace appears unreliable. Multiple fine proxy triggers may fire for one sample; the reported merged label groups related triggers into the coarse categories below, with mixed/ambiguous used when no single coarse explanation is dominant.
| Axis | Coarse label | Meaning / typical proxy evidence |
| Endpoint symptom | Refinement / edge-selection | The final decoded graph has local-looking edit errors, such as a wrong competing edge, center mislocalization, or an incomplete final edit. |
| Endpoint symptom | Program | The decoded output appears wrong at a broad reaction-program level, beyond a single final edge choice. |
| Endpoint symptom | Nonresponsive / collapse | The decoded output barely responds to the residual path, collapses to a scaffold-like output, or remains near an unedited state. |
| Endpoint symptom | Decoder realization | The latent/bond proxies look plausible, but the discrete decoder still misses exact endpoint realization. |
| Endpoint symptom | Dosage | Changing the residual magnitude appears to improve the proxy path, suggesting under- or over-application of the residual. |
| Latent-stage diagnosis | Program / parallel | The program-stage trace is atypical, undercommitted, mislocalized, or poorly anchored before residual refinement. |
| Latent-stage diagnosis | Residual refinement | is partly plausible, but the residual direction is insufficient, mislocalized, or points toward a competing edit. |
| Latent-stage diagnosis | Decoder realization | Latent localization and edit proxies are plausible, but exact graph realization fails after decoding. |
| Latent-stage diagnosis | Dosage | The residual direction appears useful but the learned scalar amount is too weak, too strong, or otherwise off its best local range. |
| Latent-stage diagnosis | Nonresponsive / off-manifold | Sweeping residual dosage does not produce a useful decoded response, suggesting the local latent path is ineffective. |
| Latent-stage diagnosis | Mixed / ambiguous | Multiple coarse latent explanations fire without a dominant assignment. |
Table 28. Coarse failure-taxonomy labels used by the endpoint–latent crosstab.
Concrete thresholding convention. The fine labels below use fixed, quantile-based thresholds computed over the available USPTO-50K retro test rows before filtering to wrong examples ( unless noted). We write low() for values in the bottom quartile, very_low() for the bottom decile, high() for the top quartile, and very_high() for the top decile. When a low cutoff equals the observed maximum, as happens for saturated F1 proxies, the rule uses strict so that perfect F1 values are not labeled low only because the distribution is saturated. Rules fire only on wrong natural outputs; correct examples are tracked only as diagnostic warnings. If a proxy is unavailable for a sample, the corresponding evidence item simply does not fire.
Table 29: Numeric cutoffs used by the automatic failure-taxonomy rules. Low and high correspond to and ; very-low and very-high correspond to and .
| Proxy | |||||
| edit F1 | 5007 | 0.0000 | 0.1292 | 0.6667 | 1.0000 |
| center F1 | 5007 | 0.0000 | 0.3077 | 0.8000 | 1.0000 |
| edit F1 | 5007 | 0.4000 | 1.0000 | 1.0000 | 1.0000 |
| center F1 | 5007 | 0.8000 | 1.0000 | 1.0000 | 1.0000 |
| edit-F1 gain | 5007 | 0.0000 | 0.1818 | 0.6000 | 1.0000 |
| center-F1 gain | 5007 | 0.0000 | 0.2000 | 0.5000 | 1.0000 |
| Residual edit/center gain | 5007 | 0.0000 | 0.2667 | 0.5303 | 1.0000 |
| Program atypicality | 5007 | 0.3055 | 0.3586 | 0.5584 | 0.7023 |
| 5007 | 0.0113 | 0.0148 | 0.0302 | 0.0534 | |
| 5007 | 0.0228 | 0.0246 | 0.0321 | 0.0391 | |
| residual norm | 5007 | 0.2571 | 0.2840 | 0.4059 | 0.5779 |
| Center mass ratio | 5007 | 0.1730 | 0.2038 | 0.4000 | 0.5896 |
| Off-center mass ratio | 5007 | 0.4104 | 0.6000 | 0.7962 | 0.8270 |
| Residual entropy | 5007 | 0.9342 | 0.9501 | 0.9741 | 0.9794 |
| Mislocalized residual score | 5007 | 0.3838 | 0.5699 | 0.7752 | 0.8081 |
| Combined instability score | 5007 | 1.7549 | 2.1491 | 2.5664 | 2.6317 |
| Residual boundary-crossing rate | 5002 | 0.0000 | 0.0000 | 0.5000 | 0.9333 |
| Wrong-edge crossing rate | 686 | 0.0000 | 1.0000 | 1.0000 | 1.0000 |
| Competing-edge suppression | 686 | -0.1608 | 0.2962 | 0.7595 | 0.9319 |
| Residual target-error reduction | 5002 | 0.0231 | 0.1177 | 0.5325 | 0.8698 |
| edit-score margin | 5002 | -1.0034 | -0.0128 | -0.0067 | -0.0061 |
| Top non-GT edit magnitude | 5007 | 0.0026 | 0.0027 | 0.0033 | 0.6106 |
| Best dosage | 5007 | -0.2500 | 0.2500 | 0.5000 | 0.5000 |
| Same-class residual-control edit F1 | 5007 | 0.0000 | 0.2222 | 0.7778 | 0.8889 |
| True minus same-class edit F1 | 5007 | 0.0000 | 0.1111 | 0.5556 | 0.9394 |
| GT-edit residual derivative | 1023 | -0.1357 | 0.0009 | 0.5611 | 1.0839 |
| GT-edit residual gradient cosine | 1023 | -0.0820 | -0.0128 | 0.1151 | 0.2095 |
Table 29.
Table 30: Endpoint-side fine triggers before coarse merging. All rows are conditioned on the natural decoded output being wrong.
| Fine endpoint trigger | Thresholded proxy condition |
| Program-level failure | Low edit F1 or center F1, together with high program atypicality or high family-centroid distance. |
| Center mislocalization | Low center mass ratio, high off-center mass ratio, or a top-3/top-5 residual-center miss. |
| Competing-edge failure | At least two of: low competing-edge suppression, high wrong-edge crossing, high top non-GT edit magnitude, or low edit-score margin. |
| Under-refinement | proxy quality is low, and either best , improves over , or the dosage sweep is monotonic-improving. |
| Overshoot | Dosage response is overshoot, and either best or is worse than a lower dose. |
| Nonresponsive / collapse | Dosage response is nonresponsive, and the stage path remains low, the alpha path is flat, or the alpha-sweep proxy range is unavailable. |
| Refinement-level failure | edit and center F1 are not low, but edit/center F1 or the residual gain is low. |
| Template/prototype insufficiency | The true sample residual control is high, the same-class/bucket control is low, and the true-minus-control gap is high. |
| Program-atypical hard case | Program atypicality is very high, residual norm is high, and edit or center F1 is low. |
| Residual-instability risk | At least two of: high residual entropy, high off-center mass, low edit-score margin, high mislocalized residual score, or high combined instability score. |
| Untyped proxy-insufficient | No endpoint rule fires with enough available evidence; used only as a low-confidence fallback before coarse merging. |
Table 30.
The endpoint coarse merge maps program-level, program-atypical, and template/prototype triggers to Program; center, competing-edge, and refinement-level triggers to Refinement / edge-selection; under- and over-dosage triggers to Dosage; and nonresponsive triggers to Nonresponsive / collapse. Decoder realization is reserved for exact-output misses where the latent and decoder-score proxies are otherwise plausible.
Table 31: Latent-stage fine triggers before coarse merging. All rows are conditioned on the natural decoded output being wrong.
| Fine latent trigger | Thresholded proxy condition |
| Anchor mislocalization | edit F1 or center F1 is very low. This rule is marked limited-support because readout evidence is sparse. |
| Program atypicality | Program atypicality or family-centroid distance is high, and either edit/center F1 is low or residual norm is high. |
| Program undercommitment | At least two gains are low across edit, formed-bond, and center proxies, with low or low quality. |
| Wrong program commitment | edit/center F1 is low, a competing/wrong-edge signal fires, the program proxy or is high, and residual gain remains low. |
| Residual mislocalization | At least two of low center mass, high off-center mass, top- center miss, or high residual entropy, paired with low center quality or low center mass. |
| Residual misdirection | At least two of low/negative GT-edit derivative, high wrong-edge crossing, low competing-edge suppression, or low target-error reduction, with low or negative residual gain. |
| Residual insufficiency | edit/center F1 is not very low, residual gain is low, and dosage or crossing evidence suggests that is too weak. |
| Residual overdrive | Dosage response is overshoot, and best , is worse than a lower dose, or residual gain is negative. |
| Latent nonresponsive | Dosage response is nonresponsive, and the alpha path is flat, the stages stay low, or residual gain is low. |
| Decoder realization | Residual localization and decoder-score proxies look plausible, edit/center proxies are not low, but the exact decoded product is still wrong. |
| Residual template-like failure | Same-class/bucket residual control is both close to the true residual and relatively strong. |
| Latent high-risk untyped | Multiple high-risk proxies fire, but no specific latent rule has enough support; also used as the fallback when no rule fires. |
Table 31. Latent-stage fine triggers before coarse merging. All rows are conditioned on the natural decoded output being wrong.
The endpoint primary label is chosen by evidence count with a fixed tie priority, while all secondary fine labels are retained for auditing. The latent primary priority is nonresponsive, program, residual/refinement, dosage, decoder-realization, and then mixed/ambiguous; explicit best-alpha or response-type evidence can promote dosage over residual refinement.
Broader Impact and Asset Licenses
Societal impact. This work aims to improve the inspectability of reaction models, which may support safer and more reliable use of machine-learning tools in chemical prediction and retrosynthesis. The main risk is indirect dual use: improved reaction-model analysis could be integrated into synthesis-planning workflows that might be used for unsafe or harmful chemical design. Our method is an audit-time diagnostic framework rather than an autonomous synthesis planner, and it should be used with expert chemical oversight.
Existing assets and licenses. We use the public USPTO-50K Schneider et al. [2016] and USPTO-MIT Jin et al. [2017] reaction benchmarks following prior work and cite the original dataset/source papers. Baseline numbers are taken from the cited papers under their reported USPTO-50K reaction-class-unknown settings. The transfer analyses in Appendix § 7.§ 10.9 use the public implementations of Graph2Edits Zhong et al. [2023], EditRetro Han et al. [2024], and Graph2SMILES Tu and Coley [2022] released by their authors, under their respective licenses. Molecular preprocessing and evaluation use RDKit, distributed under the BSD license. We do not use or release scraped proprietary datasets.
Reference-Axis, Alignment, and Transfer Analyses
This section gives the full results summarized in Section 4.3. Unless stated otherwise, results are retrosynthesis top-1 exact-match accuracies (%) on the USPTO-50K test split.
Alignment-weight sweep. Each setting in Table 6 is a separate checkpoint trained with the same nominal training budget as the audited checkpoint. All other hyperparameters follow Table 25.
| Component | Configuration |
| Latent/token dimensions | , atom slots |
| Backbone encoder/decoder | 8 Transformer layers, 8 heads, dropout 0.1 |
| Shared update | , shared parameters, 2 layers, hidden dim 512, 8 heads |
| Direction conditioning | learned embedding , retro=0 and forward=1 |
| Task adapter | direction-conditioned 2-layer attention adapter applied to in the supervised transformation |
| Decoder input | fuse mode, before residual graph decoding |
| Bond decoder | direction-conditioned residual bond-change decoder, depth 3, FF multiplier 2 |
| Endpoint alignment weight | |
| Training mixture | forward/retro ratio 0.5/0.5 |
| Training hardware | 4× NVIDIA A10, 10 h (40 GPU-hours) |
Table 25. UniRxnAxis checkpoint used in the directional audit. The model is trained jointly on forward and retrosynthetic directions; reaction-class conditioning and cycle loss are disabled in this checkpoint.
Decoder fine-tuning. Keeping the encoder and transformation frozen, we fine-tune the decoder on paired inputs . Because such pairs can train a new readout even when the original decoder cannot execute , and can also introduce an shortcut, this intervention changes the representation-to-output interface rather than the representation. Table 32 scales the displacement fed to the fine-tuned decoder.
| Displacement dosage | 0 | 0.25 | 0.5 | 1 |
| Top-1 (%) | 10.39 | 48.11 | 62.41 | 64.45 |
Table 32. Fine-tuned decoder, decoding . Accuracy at exposes the shortcut, so the value is a supervised-recoverability upper bound. The original decoder gives 4.05% at and 11.10% at .
Frozen-model linear calibration. We fit a zero-intercept ridge map on the training split so that matches the realized displacement , using latent matching only (no output labels); the model remains frozen. Mapped-axis decoding rises from 11.10% to 21.25%. Re-decomposing around gives 52.73% residual-only versus 21.25% parallel-only, a 31.48-point gap, compared with 53.07 points around .
Matched-random null axes. For each test reaction we draw eight random unit axes with . This matches the parallel and residual component norms, and the distances of the intervened states to and ; only the orientation changes. With the model frozen, the target-defined axis yields a 53.07-point Residual–Parallel separation versus 37.13 points for matched-random axes (difference 15.94, 95% bootstrap CI [14.83, 17.10]). Along the residual dosage path , , the target-defined residual’s accuracy AUC is 6.61 points higher than that of the matched-random residuals.
Relative magnitude and target-axis sweep. On the test split, (s.e.; median 0.9553), so and lie at comparable distances from . Table 33 decodes along the target axis directly.
| 0 | 0.25 | 0.5 | 0.75 | 1 | 1.25 | 1.5 | 2 | 2.5 | 2.75 | 3 | 4 | 6 | |
| Top-1 | 4.05 | 4.59 | 6.55 | 8.59 | 11.10 | 13.40 | 15.20 | 17.74 | 18.99 | 19.53 | 19.29 | 17.52 | 11.72 |
Table 33. Decoding with the original frozen decoder. The response is non-monotonic and peaks at , far below (64.41%).
Graph2Edits transfer. We train Graph2Edits [Zhong et al., 2023] with its native edit loss plus an analogous endpoint-state alignment loss (), adding no new parameters or decoder heads. The resulting model reaches 55.44%/83.74% official end-to-end retrosynthesis top-1/top-5. Because Graph2Edits acts through successive edits rather than one global latent transition, we use an edit-level audit. For 4,999 test reactions we supply the ground-truth first edit, which prevents earlier errors from contaminating the analysis. We take the recurrent states immediately before () and after () that edit, define by a fresh encoding of the target reactant graph, and decompose around . With all parameters frozen, the existing edit head predicts the second edit from each intervened state (Table 34).
| State | Parallel | Residual | |||
| Top-1 (%) | 0.68 | 16.58 | 22.74 | 48.79 | 75.80 |
Table 34. Edit-level audit of endpoint-aligned Graph2Edits: second-edit top-1 (%). Residual exceeds Parallel by 32.21 points.
Post-hoc split-space alignment for other architectures. Table 35 repeats the split-aligned forward–retro retrieval of Table 2 on the 5,007-reaction test set for two external architectures. EditRetro recovers strong cross-direction correspondence, whereas Graph2SMILES is much weaker, so post-hoc alignability depends on the architecture and representation. This experiment tests alignability only; it does not test whether is functionally dominant in these models.
Table 35. Post-hoc split-space forward–retro retrieval on -type states.
USPTO-MIT geometry. We apply the audited checkpoint to USPTO-MIT [16] and recompute the per-sample geometry of Table 3. Absolute geometry shifts modestly, with lower cosine alignment and larger orthogonal fractions, but the forward–retro contrast is stable across datasets and splits (Table 36). This supports cross-dataset transfer of the relative geometry, not of decoder executability.
| Dataset | Mean F/R | Mean F/R | ||
| USPTO-50K test | .3282 / .6273 | +.2991 | .9322 / .7405 | −.1917 |
| USPTO-MIT train | .2561 / .5503 | +.2942 | .9529 / .7824 | −.1705 |
| USPTO-MIT val | .2560 / .5522 | +.2962 | .9531 / .7813 | −.1718 |
| USPTO-MIT test | .2566 / .5516 | +.2950 | .9528 / .7812 | −.1716 |
Table 36. Per-sample geometry on USPTO-50K and USPTO-MIT. F/R = forward/retro; = Retro − Forward.
NeurIPS Paper Checklist
Claims
Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?
Answer: [Yes]
Justification: Yes. The abstract and introduction accurately reflect the paper’s main contribution: a directional latent audit framework for endpoint-supervised reaction models, with evidence that an endpoint-orthogonal residual is functionally important for retrosynthesis. We explicitly scope the claims as model-side audit rather than recovery of human chemical mechanisms or a deployable inference-time steering method, and state that the magnitude of the residual–parallel gap depends on the reference axis, the alignment weight, and the decoder interface (Section 4.3).
Limitations
Question: Does the paper discuss the limitations of the work performed by the authors?
Answer: [Yes]
Justification: Yes. § 5 contains an explicit Limitations paragraph: the decomposition requires the ground-truth endpoint and is therefore an audit-time diagnostic; the gap magnitude depends on the reference axis, , and the decoder interface; executable evidence covers retrosynthesis on USPTO-50K with two architectures (USPTO-MIT supports geometric transfer only); only deterministic top-1 is reported; and the reaction-center and failure labels are automatic proxies. Section 4.3 and Appendix § 7.§ 10.9 report the corresponding sensitivity analyses.
Theory assumptions and proofs
Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
Answer: [N/A]
Justification: This paper does not include theoretical result.
Experimental result reproducibility
Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
Answer: [Yes]
Justification: The paper provides the information needed to reproduce the main experimental results that support the claims. The main text specifies the UniRxnAxis architecture, including the shared endpoint encoder, direction-conditioned transformation with task adapters, source-conditioned residual decoder, structured prediction loss, and latent endpoint-alignment loss. It also defines the directional decomposition used by the audit, including the source state , transformed state , encoded target endpoint , endpoint direction , endpoint-parallel component , and endpoint-orthogonal residual . The appendix provides the training and implementation hyperparameters, audit-state definitions, aggregation domains, metric definitions, and intervention protocols needed to reproduce the staged decoding, residual replacement controls, decoder-boundary probes, dosage sweeps, atom-level diagnostics, and failure taxonomy. Appendix § 7.§ 10.9 specifies the reference-axis, alignment-weight, calibration, matched-random-axis, and cross-architecture/cross-dataset analyses. Code and processed data release details are addressed separately in the open-access checklist item.
Open access to data and code
Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
Answer: [Yes] Justification: Code is publicly available at https://github.com/yilish/UniRxnAxis, and the main text gives this link in § 4. The repository contains the UniRxnAxis implementation, the training and evaluation entry points, and the directional-audit scripts used for the staged decoding, residual replacement controls, decoder-boundary probes, dosage sweeps, atom-level diagnostics, and failure taxonomy, together with the environment specification and instructions for preparing the public USPTO-50K and USPTO-MIT benchmarks from their original sources. The paper also describes the data setting, splits, and evaluation protocol.
Experimental setting/details
Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?
Answer: [Yes]
Justification: The paper specifies the experimental setting needed to understand the reported results. The main text describes the USPTO-50K evaluation setting, the reaction-class-unknown retrosynthesis benchmark, the model architecture, the structured endpoint prediction loss, the latent endpoint-alignment loss, and the directional decomposition used in the audit. The appendix provides the full experimental details, including data splits and preprocessing, model hyperparameters, optimizer and training configuration, checkpoint selection, audit-state definitions, metric definitions, aggregation domains, bootstrap protocol, and the intervention procedures used for staged decoding, residual replacement controls, decoder-boundary probes, dosage sweeps, atom-level diagnostics, and failure taxonomy.
Experiment statistical significance
Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
Answer: [Yes]
Justification: The paper reports statistical uncertainty for the main experimental claims using bootstrap confidence intervals. The appendix explains that confidence intervals are computed by resampling test samples with replacement; edge-level quantities are first aggregated to sample-level means when sample identifiers are available, and aggregate-only quantities are explicitly marked rather than assigned invented intervals. Bootstrap intervals are reported for the staged decoding results, residual replacement controls, decoder-boundary and directional-derivative probes, dosage sweeps, atom-level diagnostics, failure-taxonomy statistics, the residual-dosage selector, and the matched-random-axis control in Appendix § 7.§ 10.9. These intervals cover the experiments supporting the main claims about endpoint-parallel insufficiency, residual utility, residual robustness, decoder-level bond-edit effects, and failure diagnosis.
Experiments compute resources
Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
Answer: [Yes]
Justification: Table 25 reports the training hardware for the audited checkpoint: NVIDIA A10 GPUs for roughly 10 hours, about 40 GPU-hours. Each checkpoint in Table 6 is trained under the same nominal budget, as stated in Appendix § 7.§ 10.9. The remaining audit analyses are post-hoc evaluations of trained checkpoints and cost substantially less than training.
Code of ethics
Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
Answer: [Yes]
Justification: We follow code of ethics of NeurIPS.
Broader impacts
Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
Answer: [Yes]
Justification: The paper discusses societal impact in Appendix A.13. Positive impacts include improving the inspectability and reliability of machine-learning models for chemical reaction prediction and retrosynthesis, which may help chemists diagnose model failures and use predictive tools more cautiously. Potential negative impacts are indirect: better reaction-model auditing and prediction tools could contribute to computer-aided synthesis workflows that might be misused for harmful chemical design or unsafe synthesis planning if deployed without domain oversight. The work is primarily a model-side audit framework rather than a deployable synthesis-planning system, and we emphasize that its outputs should be treated as diagnostic evidence rather than autonomous chemical recommendations.
Safeguards
Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?
Answer: [N/A]
Justification:The paper does not release a high-risk pretrained generative model, scraped dataset, image generator, language model, or autonomous synthesis-planning system. The experiments use the public USPTO-50K benchmark, and the proposed model is used as an audit-time diagnostic framework for analyzing reaction-model latent updates rather than as a deployed tool for autonomous chemical design. Therefore, the work does not require additional controlled-release safeguards of the type described in this checklist item.
Licenses for existing assets
Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
Answer: [Yes]
Justification:We use the public USPTO-50K reaction benchmark following prior retrosynthesis work and cite the original dataset/source papers. Baseline numbers are taken from the cited papers under their reported USPTO-50K reaction-class-unknown settings. We also use USPTO-MIT [16] and, for the transfer analyses in Appendix A.14, the public implementations of Graph2Edits [Zhong et al., 2023], EditRetro [12], and Graph2SMILES [Tu and Coley, 2022] released by their authors, which we cite and use under their respective licenses. Molecular preprocessing and evaluation use RDKit, which is distributed under the BSD license. We do not use or release scraped proprietary datasets.
New assets
Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
Answer: [Yes]
Justification: The paper does not introduce a new dataset or benchmark; the experiments use existing public reaction benchmarks. The accompanying code release at https://github.com/yilish/UniRxnAxis is documented in the repository, which states its license and gives usage instructions for training, evaluation, and the directional audit.
Crowdsourcing and research with human subjects
Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
Answer: [N/A]
Justification: The paper does not involve crowdsourcing, user studies, surveys, behavioral experiments, or research with human subjects. All experiments are conducted on existing chemical reaction benchmark data and model-generated outputs.
15. Institutional review board (IRB) approvals or equivalent for research with human subjects
Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
Answer: [N/A]
Justification: The paper does not involve crowdsourcing, user studies, surveys, behavioral experiments, or research with human subjects. All experiments are conducted on existing chemical reaction benchmark data and model-generated outputs, so IRB approval or equivalent human-subject review is not applicable.
16. Declaration of LLM usage
Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.
Answer: [N/A]
Justification: The core method and experiments do not use LLMs as an important, original, or non-standard component. The paper proposes and evaluates a latent reaction-model audit framework based on UniRxnAxis, directional decomposition, decoder interventions, and benchmark evaluation. Any use of LLM tools, if any, was limited to writing, editing, or formatting assistance and did not affect the core methodology, scientific claims, experimental design, or results.
References
- [1]Hangrui Bi, Hengyi Wang, Chence Shi, Connor Coley, Jian Tang, and Hongyu Guo. Non-autoregressive electron redistribution modeling for reaction prediction. In International Conference on Machine Learning, pages 904–913. PMLR, 2021.arxiv.org/abs/2106.07801
- [2]John Bradshaw, Brooks Paige, Matt J Kusner, Marvin Segler, and José Miguel Hernández-Lobato. A model to search for synthesizable molecules. Advances in Neural Information Processing Systems, 32, 2019.arxiv.org/abs/1906.05221
- [3]John Bradshaw, Anji Zhang, Babak Mahjour, David E Graff, Marvin HS Segler, and Connor W Coley. Challenging reaction prediction models to generalize to novel chemistry. arXiv preprint arXiv:2501.06669, 2025.arxiv.org/abs/2501.06669
- [4]Shuan Chen and Yousung Jung. Deep retrosynthetic reaction prediction using local reactivity and global attention. JACS Au, 1(10):1612–1620, 2021.
- [5]Ziqi Chen, Oluwatosin R Ayinde, James R Fuchs, Huan Sun, and Xia Ning. G 2 retro as a two-step graph generative models for retrosynthesis prediction. Communications Chemistry, 6(1):102, 2023.
- [6]Connor W. Coley, Luke Rogers, William H. Green, and Klavs F. Jensen. Computer-assisted retrosynthesis based on molecular similarity. ACS Central Science, 3(12):1237–1245, 2017.DOI
- [7]Connor W Coley, Wengong Jin, Luke Rogers, Timothy F Jamison, Tommi S Jaakkola, William H Green, Regina Barzilay, and Klavs F Jensen. A graph-convolutional neural network model for the prediction of chemical reactivity. Chemical science, 10(2):370–377, 2019.
- [8]Hanjun Dai, Chengtao Li, Connor W. Coley, Bo Dai, and Le Song. Retrosynthesis prediction with conditional graph logic network. In Advances in Neural Information Processing Systems, 2019.
- [9]Yafeng Deng, Xinda Zhao, Hanyu Sun, Yu Chen, Xiaorui Wang, Xi Xue, Liangning Li, Jianfei Song, Chang-Yu Hsieh, Tingjun Hou, et al. Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints. Nature communications, 16(1):7012, 2025.
- [10]Kien Do, Truyen Tran, and Svetha Venkatesh. Graph transformation policy network for chemical reaction prediction. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 750–760, 2019.
- [11]Michael E Fortunato, Connor W Coley, Brian C Barnes, and Klavs F Jensen. Data augmentation and pretraining for template-based retrosynthetic prediction in computer-aided synthesis planning. Journal of chemical information and modeling, 60(7):3398–3407, 2020.
- [12]Yuqiang Han, Xiaoyang Xu, Chang-Yu Hsieh, Keyan Ding, Hongxia Xu, Renjun Xu, Tingjun Hou, Qiang Zhang, and Huajun Chen. Retrosynthesis prediction with an iterative string editing model. Nature Communications, 15(1):6404, 2024.DOI
- [13]Friedrich Hastedt, Rowan M Bailey, Klaus Hellgardt, Sophia N Yaliraki, Ehecatl Antonio del Rio Chanona, and Dongda Zhang. Investigating the reliability and interpretability of machine learning frameworks for chemical retrosynthesis. Digital Discovery, 3(6):1194–1212, 2024.
- [14]Ross Irwin, Spyridon Dimitriadis, Jiazhen He, and Esben Jannik Bjerrum. Chemformer: a pre-trained transformer for computational chemistry. Machine Learning: Science and Technology, 3(1):015022, 2022.
- [15]Shoichi Ishida, Kei Terayama, Ryosuke Kojima, Kiyosei Takasu, and Yasushi Okuno. Prediction and interpretable visualization of retrosynthetic reactions using graph convolutional networks. Journal of chemical information and modeling, 59(12):5026–5033, 2019.
- [16]Wengong Jin, Connor Coley, Regina Barzilay, and Tommi Jaakkola. Predicting organic reaction outcomes with weisfeiler-lehman network. Advances in neural information processing systems, 30, 2017.arxiv.org/abs/1709.04555
- [17]Joonyoung F Joung, Mun Hong Fong, Nicholas Casetti, Jordan P Ilies, Ne S Dassanayake, and Connor W Coley. Electron flow matching for generative reaction mechanism prediction. Nature, 645(8079):115–123, 2025.DOI
- [18]Pavel Karpov, Guillaume Godin, and Igor V. Tetko. A transformer model for retrosynthesis. ChemRxiv, 2019.DOI
- [19]Eunji Kim, Dongseon Lee, Youngchun Kwon, Min Sik Park, and Youn-Suk Choi. Valid, plausible, and diverse retrosynthesis using tied two-way transformers with latent variables. Journal of Chemical Information and Modeling, 61(1):123–133, 2021.DOI
- [20]Dávid Péter Kovács, William McCorkindale, and Alpha A Lee. Quantitative interpretation explains machine learning models for chemical reaction prediction and uncovers bias. Nature communications, 12(1):1695, 2021.DOI
- [21]Xinjie Li and Abhinav Verma. Retromoe: A mixture-of-experts latent translation framework for single-step retrosynthesis. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 7509–7517, 2025.DOI
- [22]Haitao Lin, Junjie Wang, Zhifeng Gao, Xiaohong Ji, Rong Zhu, Linfeng Zhang, Guolin Ke, et al. Synbridge: Bridging reaction states via discrete flow for bidirectional reaction prediction. arXiv preprint arXiv:2507.08475, 2025.arxiv.org/abs/2507.08475
- [23]Bo Liu, Bharath Ramsundar, Parichay Kawthekar, Jie Shi, Joseph Gomes, Quoc Le Nguyen, Stephan Ho, Jack Sloane, Paul Wender, and Vijay Pande. Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS Central Science, 3(10):1103–1113, 2017.arxiv.org/abs/1706.01643
- [24]Jieyu Lu and Yingkai Zhang. Unified deep learning model for multitask reaction predictions with explanation. Journal of chemical information and modeling, 62(6):1376–1387, 2022.DOI
- [25]Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, 2017.arxiv.org/abs/1705.07874
- [26]Ziqiao Meng, Peilin Zhao, Yang Yu, and Irwin King. Doubly stochastic graph-based non-autoregressive reaction prediction. arXiv preprint arXiv:2306.06119, 2023.arxiv.org/abs/2306.06119
- [27]Giorgio Pesciullesi, Philippe Schwaller, Teodoro Laino, and Jean-Louis Reymond. Transfer learning enables the molecular transformer to predict regio-and stereoselective reactions on carbohydrates. Nature communications, 11(1):4874, 2020.
- [28]Mikołaj Sacha, Mikołaj Błaż, Piotr Byrski, Paweł Dąbrowski-Tumanski, Mikołaj Chrominski, Rafał Loska, Paweł Włodarczyk-Pruszyński, and Stanisław Jastrzębski. Molecule edit graph attention network: modeling chemical reactions as sequences of graph edits. Journal of Chemical Information and Modeling, 61(7):3273–3284, 2021.DOI
- [29]Nadine Schneider, Nikolaus Stiefl, and Gregory A Landrum. What’s what: The (nearly) definitive guide to reaction role assignment. Journal of chemical information and modeling, 56(12):2336–2346, 2016.DOI
- [30]Philippe Schwaller, Theophile Gaudin, David Lanyi, Costas Bekas, and Teodoro Laino. “found in translation”: predicting outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chemical science, 9(28):6091–6098, 2018.DOI
- [31]Philippe Schwaller, Teodoro Laino, Théophile Gaudin, Peter Bolgar, Christopher A Hunter, Costas Bekas, and Alpha A Lee. Molecular transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS central science, 5(9):1572–1583, 2019.DOI
- [32]Marwin HS Segler and Mark P Waller. Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry–A European Journal, 23(25):5966–5971, 2017.DOI


