acceptodds
Under review as a conference paper at ICLR 2027

ChemVD: Constructing Open-world Reaction Variable Spaces via Hierarchical Contrastive Alignment

Abstract

Reaction condition recommendation should construct compact, high-value variable subspaces from expanding candidate sets for more efficient experimental search. However, existing methods trained to recover historical conditions or rank predefined candidates struggle to generalize to unseen reagents under sparse comparative supervision, or to effectively model joint reaction–category–reagent dependencies. To address these limitations, we propose Chemical Variable-space Design (ChemVD), a structure-aware hierarchical cross-modal contrastive learning framework that reframes reaction condition recommendation as open-candidate retrieval and ranking for screening-space construction. ChemVD maps reaction, category, and reagent representations into a shared embedding space through hierarchical contrastive learning that explicitly captures their joint dependencies, while retaining generalization through multi-positive learning and chemically informed constraints. Experiments show that ChemVD generalizes effectively to reagents unseen during training. On HTE-OOD, with a 20% screening budget, ChemVD achieves a mean achievable maximum yield of 68.78% using an 8B-parameter backbone, outperforming a 1.6T-parameter LLM baseline by 20.44 percentage points. By enabling chemically informed, hierarchical ranking over openworld candidates, this work provides a new modeling paradigm for proposing compact variable spaces to support AI-driven scientific discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chemical variable-space design, Open-candidate space, Hierarchical contrastive alignment, Multi-positive learning, AI for chemistry

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ULRrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper reframes reaction-condition recommendation as budget-constrained "variable-space design": given a reaction and an open library of candidate reagents organized into functional categories (catalyst, ligand, base, solvent), select per-category subsets whose Cartesian product is a compact screening subspace that retains high-yield combinations under an experimental budget. The proposed method, ChemVD, encodes reactions (DRFP), categories (frozen Llama-3.1-8B) and reagents (Morgan fingerprints) into a shared embedding space, retrieves categories first, then ranks reagents with a category-conditioned query jointly encoding reaction and category. Training uses multi-positive contrastive learning, a per-positive coverage loss, yield-aware positive weighting, and several chemically informed regularizers (low-yield suppression, explicit-negative margin loss, category reweighting, frequency-based hard negatives). Validation: (i) Reagent-Split retrieval on a ChemRANG-derived dataset, stratified into seen/unseen reagents; (ii) HTE-OOD subspace design on 15 Suzuki high-throughput subsets expanded with 24 infeasible salt "catalysts" (from 384 to 9,600 conditions), evaluated by mean achievable maximum yield (MAMY) under per-category budgets; (iii) a controlled XOR / Independent distribution experiment and same-encoder ablations isolating the hierarchical architecture. Headline: at a 20% budget, ChemVD (8B backbone) reaches MAMY 68.78%, beating the zero-shot DeepSeek-V4-Pro (1.6T) API baseline by 20.44 pp, while the same-encoder Pairwise ablation reaches only ~15% mean MAMY.

Critical strengths
  1. A genuinely useful reframing, validated by the right control experiment. Recasting condition recommendation as budget-constrained open-candidate retrieval for compact subspace construction seems to be novel (I could not find this formulation in the literature) and directly matches how experimentalists actually design screens. Crucially, the paper isolates the architectural contribution with a same-encoder Pairwise ablation (identical encoders, data, multi-positive objective; only the hierarchical query removed): mean MAMY 15.27% vs 78.14%, with consistent wAPC@K gains on both seen and unseen reagent retrieval, plus a clean XOR / Independent distribution experiment. This is the strongest evidence in the paper and it supports the central claim that joint reaction–category conditioning, not candidate-space openness alone (BCE also has open candidates and fails), drives the gains.
  2. Above-average experimental coverage. Baselines span closed-set classifiers (ReaCon, Label-mix), autoregressive models (Parrot, TextReact), zero-shot LLMs from 30B to 1.6T (local and API, with verbatim prompts), a nonparametric train-frequency baseline, and objective-level baselines; ablations cover every loss component, encoder swaps (Table 13), and data volume. Leakage-aware cluster splitting, shared candidate pools for the seen / unseen strata, three seeds with standard deviations, and full hyperparameter tables show real care.
Critical weaknesses
  1. The headline LLM comparison is misleading as presented. The abstract's "outperforming a 1.6T-parameter LLM baseline by 20.44 percentage points" selects the weakest LLM (DeepSeek-V4-Pro) at the single budget where it is weakest. Table 12 shows Llama-3.1-70B (71.26) and Llama-3.3-70B (72.16) both beat ChemVD (68.78) at the 20% budget, and Llama-3.3-70B also leads at 30%; ChemVD's wins at 40-80% budgets occur where the task is near-trivial (even Pairwise reaches 85.99 at 80%). The body text is honest about this, so the full data is available — but the abstract overstates the result. Relatedly, Figure 3(b) plots "model parameters" including ChemVD's frozen Llama-3.1-8B encoder, so "an 8B model beats 70B/1.6T LLMs" really means "a small trainable head on a frozen 8B encoder, trained on ~98k domain examples, beats zero-shot general-purpose LLMs" — a legitimate and interesting result that the figure's framing obscures. The OOD evidence is weaker than claimed. The HTE-OOD expansion adds 24 inorganic salts (acetates, chlorides, bromides, fluorides, sulfates, BF4-/PF6-) as pseudo-yield-0 catalysts — candidates no chemist would consider for Suzuki coupling. The demonstrated "generalization to unseen reagents" is therefore mostly the ability to exclude trivially infeasible salts, not ranking of plausible unseen reagents (e.g., other Pd complexes or new ligands). The abstract's claim that ChemVD "generalizes effectively to reagents unseen during training" rests mainly on Reagent-Split Unseen (n=1,100), where the margin over Pairwise narrows substantially. Additionally, MAMY only requires that one high-yield combination survive in the 20-condition subspace — a lenient metric with no complementary measure of overall subspace quality.
Questions for authors
  1. Table 12: Pairwise MAMY is 0.00±0.00 at 20 / 30 / 40% budget, then jumps to 1.60 at 50% and 85.99 at 80%. What causes this cliff? Does Pairwise consistently retain the same solvent / base (so only catalyst ranking matters at high budget)? Please explain or show per-category retention.
  2. The abstract claims ChemVD "generalizes effectively to reagents unseen during training." Can you evaluate subspace design on an OOD axis with plausible unseen catalysts (e.g., Pd complexes or ligands absent from training but mechanistically valid), rather than infeasible salts? Would the 20.44 pp gap over LLMs hold there?
  3. Is code and the processed ChemRANG split available (or will be, upon acceptance)? I could not find any availability statement in the paper.
  4. A.5.4: MAMY is highest at 10% training data, drops sharply at 20%, then fluctuates. What explains the non-monotonicity — is it seed variance, or does more data shift rankings toward historically frequent (but lower-yield) conditions?
  5. The category ground truth is constant per reaction family (every Suzuki reaction has {Catalyst, Ligand, Base, Solvent}). The "58.1% improvement" in All-truth hit@5 therefore measures ranking these four above distractors, not learning which roles a reaction requires. Do you agree this evaluation is near-trivial, and can you report category results on a setting where the required role set varies (e.g., reactions needing additives but no ligand)?
  6. Can you add some kind of a complementary subspace-quality metric? For instance, fraction of >90%-yield conditions retained, or mean yield over the selected subspace. That would make MAMY's leniency (one surviving high-yield combination suffices) explicit.
  7. All subspace-design evidence is Suzuki. Can the framework transfer to other reaction families (the training set includes Michael and indole) or to non-discrete variables such as temperature / time?
9 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (1/8): Description of Parameter Efficiency Advantages and LLM Comparisons

We thank the reviewer for the positive assessment of our work and for recognizing its potential value. We also appreciate the constructive suggestions. In response, we have refined our claims regarding LLM comparisons and conducted additional experiments and analyses on generalization to chemically plausible unseen reagents, category prediction under different reagent-role requirements, and subspace design quality. We provide detailed responses to each concern below.

W1: The headline LLM comparison is misleading as presented. The abstract's "outperforming a 1.6T-parameter LLM baseline by 20.44 percentage points" selects the weakest LLM (DeepSeek-V4-Pro) at the single budget where it is weakest. Table 12 shows Llama-3.1-70B (71.26) and Llama-3.3-70B (72.16) both beat ChemVD (68.78) at the 20% budget, and Llama-3.3-70B also leads at 30%; ChemVD's wins at 40-80% budgets occur where the task is near-trivial (even Pairwise reaches 85.99 at 80%). The body text is honest about this, so the full data is available — but the abstract overstates the result.

A W1: We thank the reviewer for this suggestion. We have revised the abstract to highlight that ChemVD achieves competitive subspace design performance through domain-specific representation learning with a substantially smaller frozen backbone, rather than claiming superiority over all zero-shot LLM baselines, as follows:

“On HTE-OOD, under a 20% screening budget, ChemVD achieves a mean achievable maximum yield of 68.78% through domain-specific representation learning with a frozen 8B-parameter backbone, closely competitive with the strongest LLM baseline (trailing by 3.38 percentage points) while outperforming the 1.6T-parameter LLM baseline by 20.44 percentage points.”

...Relatedly, Figure 3(b) plots "model parameters" including ChemVD's frozen Llama-3.1-8B encoder, so "an 8B model beats 70B/1.6T LLMs" really means "a small trainable head on a frozen 8B encoder, trained on ~98k domain examples, beats zero-shot general-purpose LLMs" — a legitimate and interesting result that the figure's framing obscures.

To avoid potential confusion regarding model size and training settings, we have revised the caption of Figure 3(b) to explicitly distinguish ChemVD’s domain-specific representation learning from the zero-shot evaluation of general-purpose LLM baselines, as follows:

“Figure 3(b) Average MAMY versus backbone model size. ChemVD employs a frozen 8B-parameter backbone with a domain-trained representation learning module, whereas the general-purpose LLM baselines are evaluated in a zero-shot setting.”

AuthorsOpenReview ↗

Response to Reviewer 3 (2/8): Reagent Retention and the Abrupt Increase in Pairwise MAMY

Q1: Table 12: Pairwise MAMY is 0.00±0.00 at 20 / 30 / 40% budget, then jumps to 1.60 at 50% and 85.99 at 80%. What causes this cliff? Does Pairwise consistently retain the same solvent / base (so only catalyst ranking matters at high budget)? Please explain or show per-category retention.

A1: We thank the reviewer for this insightful question. To investigate the abrupt increase in MAMY, we inspected the category-wise reagent rankings and tracked how the retained candidate sets evolve across screening budgets. Since presenting full retention breakdowns for every HTE-OOD subset would be impractical, and their MAMY trajectories closely mirror the overall trend (Table R3.1), we reported Sub3-neg as a representative case.

Our analysis reveals that the abrupt increase is primarily associated with the delayed inclusion of a critical catalyst (Table R3.2). Specifically, the candidate pools contain 25 catalysts, 12 ligands, 8 bases, and 4 solvents. Notably, Pairwise ranks Pd(OAc)₂, the sole effective catalyst in Sub3-neg, 17th among 25 catalyst candidates. Consequently, Pd(OAc)₂ is excluded at screening budgets of 60% or lower and is only retained when the budget reaches 80%. By contrast, key ligands such as PPh₃ and P(o-tol)₃ are already retained at much smaller budgets. Furthermore, due to the relatively small candidate pools for bases and solvents, increasing the screening budget introduces only a limited number of additional candidates in these categories. Together, these findings explain the abrupt MAMY increase and highlight how category-wise ranking errors can bottleneck reaction discovery under limited screening budgets.

Table R3.1: Pairwise MAMY (mean ± standard deviation) across HTE-OOD subsets at different screening budgets.

BudgetMAMY
20%–40%0.00 ± 0.00
50%1.60 ± 6.20
60%4.03 ± 8.90
80%85.99 ± 15.32

Table R3.2:Pairwise reagent rankings by category and retention cutoffs in Sub3-neg.

RankCatalyst (N=25)Ligand (N=12)Base (N=8)Solvent (N=4)
1Sr(OAc)₂PPh₃Et₃NTHF (20%)
2Mg(OAc)₂AmPhos (20%)LiOtBu (20%)DMF (40/50%)
3NaOAcPCy₃NaHCO₃ (40%)MeCN (80%)
4KOAcdppfK₃PO₄ (50%)MeOH
5Ca(OAc)₂ (20%)P(o-tol)₃ (40%)None--
6LiOAcXantphos (50%)CsF (80%)--
7Ba(OAc)₂dtbpfNaOH--
8Al(OAc)₃cataCXium AKOH--
9NaBF₄None----
10NaPF₆ (40%)P(t-Bu)₃ (80%)----
11KBF₄SPhos----
12Na₂SO₄XPhos----
13NaF (50%)------
14KPF₆------
15KF------
16K₂SO₄------
17Pd(OAc)₂------
18LiBr------
19NaBr------
20KBr (80%)------
21LiCl------
22NaCl------
23KCl------
24MgCl₂------
25CaCl₂------
AuthorsOpenReview ↗

Response to Reviewer 3 (3/8): Generalization to Chemically Plausible Unseen Ligands

W1 & Q2: The abstract claims ChemVD "generalizes effectively to reagents unseen during training." Can you evaluate subspace design on an OOD axis with plausible unseen catalysts (e.g., Pd complexes or ligands absent from training but mechanistically valid), rather than infeasible salts? Would the 20.44 pp gap over LLMs hold there?

A2: We thank the reviewer for this important suggestion. Additional evaluation was conducted using unpublished experimental data from a 1,4-diene asymmetric 1,3-difunctionalization reaction newly developed by our collaborators. As this work is unpublished and involves confidential information, we cannot currently disclose the specific reactions, complete reaction-condition records, or reagent structures. We therefore report only the training setup and selected key evaluation results below.

In this reaction, a total of 12 chiral ligands were experimentally tested, among which TY-Phos series ligands, especially TY2, TY3, TY4, and TY5, performed well in the wet-lab experiments, with TY5 as the best-performing ligand. Other TY-phos ligands (TY1, TY6-9) exhibited lower performance, and the remaining Ming-phos (M1), Xu-phos (Xu1), and Xiang-phos (X1) ligands could barely promote the reaction. However, all 12 ligands are plausible choices for chemists prior to experiments, which provides us with a challenging scenario to evaluate ChemVD.

For model training, we collected 177 experimental records from this new reaction system, all involving TY5 as the ligand, and jointly trained ChemVD with the original Suzuki dataset using the same model configuration as in the manuscript. For evaluation, we considered seven reactions, including three represented in the training data and four previously unseen substrate pairs. The inference candidate pool contained 79 reagents, of which 68 were present in the training data, and 11 were newly introduced ligands entirely unseen during training. Specifically, among the 12 candidate ligands, only TY5 appeared in the training data, whereas the remaining 11 ligands, including the experimentally effective TY2, TY3, and TY4, were unseen. This setting allows us to assess whether ChemVD could choose privileged ligands (TY2-TY5) over other ligands. ChemVD and the LLM baselines were evaluated using identical candidate pools and evaluation conditions.

We evaluated recommendation performance using Exact@1 and Family@K. Exact@1 measures whether the top-ranked recommendation matches the experimentally identified reference combination, while Family@K measures the proportion of the top-K recommendations belonging to the experimentally validated ligand family (TY2-TY5). As shown in Table R3.3, ChemVD achieves an Exact@1 of 0.77 and Family@1, Family@3, and Family@4 scores of 1.00, 0.81, and 0.72, respectively, consistently outperforming all evaluated LLM baselines.

These results provide additional evidence that ChemVD can distinguish chemically plausible reagents with different experimental performance, including promising ligands entirely unseen during training, rather than merely excluding obviously infeasible candidates.

Table R3.3: Reagent recommendation performance for 1,4-diene asymmetric 1,3-difunctionalization.

ModelExact@1Family@1Family@3Family@4
Qwen3.5-35B0.080.370.470.48
Qwen3-30B0.000.420.480.49
Llama3.1-70B0.080.480.520.52
Llama3.3-70B0.030.520.480.47
DeepSeek-v4-pro0.210.560.670.65
ChemVD0.771.000.810.72
Reviewer_DaPDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper recasts reaction-condition recommendation as budget-constrained hierarchical retrieval over an open candidate library: categories are retrieved for a reaction, a reagent subset is selected within each category, and the product of these subsets forms the screening space. ChemVD encodes reactions with DRFP, category descriptions with a frozen Llama-3.1-8B and reagents with Morgan fingerprints in a shared space, scores reagents with a joint reaction–category query, and trains with multi-positive and per-positive coverage contrastive losses, yield weights and chemistry-motivated constraints. It is evaluated on seen/unseen strata of a cluster-based ChemRANG split and on HTE-OOD, against closed-set, autoregressive and LLM baselines. The formulation is practical, but the main OOD benchmark, the model selection and the baselines do not yet isolate the claimed gains.

Critical strengths

Building a compact, budgeted subspace from an expandable library matches how HTE campaigns are designed, and scoring candidates in a shared embedding space lets newly added reagents be scored without new output classes.

The pairing of a set-level multi-positive loss with a per-positive coverage loss is well motivated and both terms matter empirically. The split is built from reaction clusters, seen and unseen queries share candidate pools, the full LLM prompts are reported, and the XOR analysis is informative within its stated assumptions.

Critical weaknesses
  1. The main OOD evidence does not isolate ranking quality and is not cleanly held out.

(i) Each HTE-OOD subset has a single measured catalyst plus 24 added simple salts whose combinations all receive pseudo-yield zero, so the large majority of combinations are synthetic. With a per-category 20% budget, five of 25 catalysts are kept, and the maximum yield in the selected space is zero whenever the real catalyst is missed. The metric is therefore dominated by a near-binary catalyst-retention decision rather than by ranking over ligands, bases and solvents, which is consistent with three-seed standard deviations of roughly 15–20 points. ChemVD is also trained with explicit and hard negatives that the LLM baselines never see.

(ii) The final encoder combination was chosen because it gave the best overall balance across both benchmarks, which are the test benchmarks. The reported test numbers therefore partly reflect model selection on the test data. For computer science papers, we think this is not good.

  1. The baselines do not yet isolate the open-candidate and hierarchical claims.

(i) As described, the BCE baseline scores a candidate with a sigmoid of a linear function of the concatenated query and candidate features. The query term is then constant across candidates within a category, so the ranking does not depend on the reaction, which explains its near-zero wMRR and means the comparison changes scoring capacity as well as the objective. ReaCon and Label-mix are also below the training-frequency baseline, which suggests under-tuning.

(ii)Pairwise averages independent reaction–reagent and category–reagent scores, so it cannot represent a reaction–category interaction by construction, and the XOR result and category-level gains follow from the baseline design. The Unseen stratum is small, defined at the example level so it can contain seen reagents, and not characterized by structural similarity to training reagents, so generalization cannot be separated from nearest-neighbour fingerprint similarity.

Minor

The abstract reports the gain over one LLM baseline, while stronger LLM baselines are above ChemVD at the same budget. In abstract you should be more strict when claiming.

Questions for authors

See weakness

11 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (1/5): HTE-OOD Evaluation and Negative Examples

Thank you for your constructive feedback. In light of your comments, we have conducted new experiments and analyses to clarify and strengthen our findings. Below, we provide a point-by-point response to each of your concerns, and hope these additional results and clarifications adequately address your reservations regarding the soundness of our work.

Q1.1: Each HTE-OOD subset has a single measured catalyst plus 24 added simple salts whose combinations all receive pseudo-yield zero ... The metric is therefore dominated by a near-binary catalyst-retention decision rather than by ranking over ligands, bases and solvents ...

A1.1: To address this concern, we conducted a complementary evaluation using only the original experimentally measured reaction conditions, excluding all artificially added combinations and their associated pseudo-yields. Importantly, after removing the synthetic negatives, the search space contains only one experimentally tested catalyst, eliminating the catalyst-retention bias. Performance differences therefore depend on the selection of ligands, bases, and solvents.

As shown in Table R2.2 and Table R2.3, the overall results across the original HTE-OOD subsets demonstrate that ChemVD achieves competitive performance across different selection budgets, with particularly pronounced advantages in challenging subsets. Specifically, we highlighted the Sub3 subset because it represents a highly hard scenario due to the sparse distribution of high-yielding conditions (only 5.21% of conditions achieve yields above 60%, as shown in Table R2.1).

As illustrated in Table R2.3, ChemVD consistently achieves the highest achievable maximum yield across all evaluated selection budgets (20%–80%) compared to large language models (LLMs). This confirms that ChemVD's strong performance is not merely an artifact of binary catalyst filtering, but stems from robust hierarchical variable-space design across all reaction components.

...ChemVD is also trained with explicit and hard negatives that the LLM baselines never see.

Although the LLM baselines were not fine-tuned on these negative examples, our additional evaluations on reaction question answering and feasibility prediction show that these LLMs can accurately identify the chemical mechanisms underlying the examined negative cases (Table R2.4). This capability suggests that the LLMs possess relevant chemical knowledge. Moreover, ChemVD consistently outperforms all LLM baselines in the Sub3 evaluation after removing all artificially introduced negative combinations (Table R2.3). Together, these findings demonstrate that ChemVD’s advantage cannot be attributed solely to its exposure to explicit negative examples during training.

Table R2.1: Yield-interval distribution of original measured combinations in Sub3

yield_interval_percent[0,20)[20,40)[40,60)[60,80)[80,100]
count2209846200
percent57.29%25.52%11.98%5.21%0.00%

Table R2.2: Mean achievable maximum yield of each method on the original measured overall space across different budgets

Budget (Overall)ChemVDPairwiseQwen3-30BQwen3.5-35BLLaMa-3.1-70BLLaMa-3.3-70BDeepSeek-v4-flashDeepSeek-v4-pro
20%68.7860.1565.4658.2971.2672.1645.3048.34
30%73.2662.5067.7261.5472.7174.0557.5159.48
40%77.7472.4776.2966.3175.5177.0871.7475.52
50%79.6274.1079.2069.7480.5781.2276.4479.95
60%81.5174.7380.7970.2083.1482.3979.9581.01
80%87.9384.5284.4677.5885.0985.1985.4784.58

Table R2.3: Achievable maximum yield of each method on the original measured Sub3 space across different budgets

Budget (Sub3)ChemVDPairwiseQwen3-30BQwen3.5-35BLlama-3.1-70BLlama-3.3-70BDeepSeek-v4-flashDeepSeek-v4-pro
20%67.5343.7137.865.3844.4757.6828.918.91
30%67.5743.8037.865.3844.4757.6847.9962.45
40%67.6157.7557.617.0944.4757.6853.1562.45
50%67.6167.5364.2210.2964.2257.6858.2762.45
60%67.6167.5364.2210.2967.4967.4962.4162.45
80%74.8469.9467.6145.0767.6167.6169.167.61

Table R2.4: LLM responses on Suzuki–Miyaura mechanism and feasibility assessment for reactions

QueryResponse (summarized)
Explain the Suzuki–Miyaura mechanism.Pd-catalyzed oxidative addition, transmetalation, and reductive elimination; bases activate organoboron reagents, while ligands modulate and stabilize the catalyst.
Assess 10 reactions with distinct inactive catalyst substitutes and randomly sampled accompanying conditions.10/10 classified as infeasible, consistently identifying the absence of an active catalyst needed to support the catalytic cycle.
AuthorsOpenReview ↗

Response to Reviewer 2 (2/5): Encoder Selection and Validation Protocol

Q1.2 : The final encoder combination was chosen because it gave the best overall balance across both benchmarks, which are the test benchmarks. The reported test numbers therefore partly reflect model selection on the test data ...

A1.2: Thank the reviewer for raising the concern. We would like to clarify that our encoder configuration was motivated by the distinct representation requirements of the three branches. Specifically, for the reaction branch, we selected DRFP over RxnFP and ChemT5-Base because it explicitly represents structural changes between reactants and products through differences in local molecular environments, providing a direct, training-independent description of reaction transformations. For the reagent branch, we adopted Morgan fingerprints rather than D-MPNN or ChemT5-Base to obtain consistent, training-independent representations of molecular substructures and local atomic environments. In contrast, the category branch requires semantic representations of functional roles and chemically defined inclusion and exclusion rules, motivating the use of a language-model-based encoder. These complementary representations provide a chemically motivated foundation for learning reaction–category–reagent associations. The encoder configuration was determined based on the aforementioned design rationales and finalized prior to test-set evaluation (Table 13).

To further clarify our encoder selection procedure, we provide the previously unreported validation results for nine encoder combinations on a separate validation set. As shown in Table R2.5, our selected configuration DRFP–Llama−31˙−8B–MorganDRFP–Llama-3\.1-8B–Morgan achieved the highest retrieval proxy score 879˙287\.92 among all evaluated combinations. The retrieval proxy is a composite metric that evaluates reagent ranking, category ranking, and the margin between positives and negatives.

Retrival-proxy=0.4×MRRg+0.3×Hitsg@10+0.15×MRRc+0.15×Marginp,gRetrival \text{-} proxy=0.4 \times MRR_g + 0.3 \times Hits_g@10 + 0.15 \times MRR_c + 0.15 \times Margin_{p,g}

These results provide additional empirical support for our encoder configuration without relying on test-set performance for encoder selection.

We explicitly confirm that no test-set samples were used for model training, parameter updates, or encoder configuration selection. We will clarify the a priori rationale behind our encoder choices and include the previously unreported validation results in the revised manuscript.

Table R2.5: Comparison of encoder configurations on validation data with retrieval proxy metric.

ConfigurationsRetrieval proxy ↑
drfp-llama-3-1-8b-morgan87.92
chemt5-llama-chemt583.55
chemt5-llama-morgan87.42
drfp-chemt5-morgan87.52
drfp-llama-chemt585.45
drfp-llama-dmpnn83.42
drfp-qwen-morgan87.84
drfp-qwenembed-morgan87.87
rxnfp-llama-morgan87.75
AuthorsOpenReview ↗

Response to Reviewer 2 (3/5): BCE Interactions and Supervised Baselines

Q2.1: As described, the BCE baseline scores a candidate with a sigmoid of a linear function of the concatenated query and candidate features. The query term is then constant across candidates within a category, so the ranking does not depend on the reaction, which explains its near-zero wMRR and means the comparison changes scoring capacity as well as the objective ...

A2.1: We thank the reviewer for the insightful suggestion. To further examine the role of query-candidate interactions in BCE-based ranking, we have supplemented an interaction-enhanced BEC baseline BCE−intBCE-int: p(q,z)=σ(wq⊤q+wg⊤z+wint⊤(q⊙z)+b)p(q,z)=σ( w_{q}^{⊤} q+w_{g}^ {⊤} z + wᵢₙₜ^{⊤}(q⊙ z) + b ) . This additonal interaction term enables the relative ranking among candidates to vary depending on the query. We have included the results for BCE-int in New Table 1 of the main text. Here, we present the comparison of the two BCE architectures and ChemVD after training:

New Table 1 (Excerpt): Retrieval performance of BCE, BCE-int, and ChemVD on the Seen and Unseen splits.

SplitModelwMRR ↑wMPR ↓wAPP ↑wAPC@1 ↑wAPC@3 ↑wAPC@5 ↑wAPC@10 ↑wAPC@20 ↑
SeenBCE2.77 ± 0.69143.3613 ± 6.725266.14 ± 37.940.02 ± 0.012.63 ± 4.125.15 ± 4.228.31 ± 0.279.39 ± 0.37
SeenBCE-int2.76 ± 0.75147.5120 ± 5.923676.43 ± 39.680.05 ± 0.012.68 ± 4.195.21 ± 4.108.32 ± 0.379.21 ± 0.20
SeenChemVD65.09 ± 0.167.7210 ± 0.070299.40 ± 0.0741.37 ± 0.0863.63 ± 0.1373.53 ± 0.3085.01 ± 0.3691.69 ± 0.23
UnseenBCE3.09 ± 0.49148.0613 ± 4.258865.96 ± 41.640.27 ± 0.082.34 ± 2.824.39 ± 2.877.10 ± 0.489.17 ± 0.49
UnseenBCE-int2.99 ± 0.45152.2604 ± 3.929175.11 ± 41.950.17 ± 0.122.43 ± 2.834.55 ± 2.857.01 ± 0.389.07 ± 0.38
UnseenChemVD48.80 ± 0.4331.8061 ± 1.124696.08 ± 0.4727.52 ± 0.4345.78 ± 0.7753.18 ± 0.8365.00 ± 0.3173.16 ± 0.60

As shown in New Table 1 (Excerpt), BCE-int achieves comparable wMRR and wMPR to the original BCE on both Seen and Unseen splits, despite improvements in mean wAPP. Importantly, BCE-int remains substantially below ChemVD across the reported ranking metrics. These findings suggest that introducing explicit query–candidate interactions alone is insufficient to close the performance gap between BCE and ChemVD.

...ReaCon and Label-mix are also below the training-frequency baseline, which suggests under-tuning.

We would like to clarify that the observed performance gap does not necessarily imply insufficient tuning. As discussed in Appendix A.1.2, supervised classification methods such as ReaCon and Label-Mix tend to learn reagent selection patterns from historical reaction records. These patterns often reflect reagent usage frequencies rather than their experimental efficacy or utility for new reactions. In contrast, the training-frequency baseline directly exploits these historical frequencies as a static ranking reference without optimization. Consequently, learning to fit historical recording probabilities does not inherently surpass a direct frequency-based reference, highlighting a fundamental objective mismatch between standard supervised classification and downstream variable-space design.

Furthermore, as reported in Table 1 (Excerpt), both ReaCon and Label-mix exhibit very small standard deviations across runs on the seen test set. This high stability indicates that their limited performance is not a symptom of underfitting optimization, but rather an intrinsic consequence of the limitations embedded within the supervised fitting objective itself.

Table 1 (Excerpt): Retrieval performance of ReaCon and Label-Mix on the Seen split.

ModelwMRR ↑wMPR ↓wAPP ↑wAPC@K @1 ↑wAPC@K @3 ↑wAPC@K @5 ↑wAPC@K @10 ↑wAPC@K @20 ↑
ReaCon17.09±0.0439.0698±0.001688.10±0.008.97±0.0612.95±0.0424.23±0.0031.10±0.0038.08±0.00
Label-mix50.86±0.6216.8131±0.062291.16±0.1427.66±0.1145.37±1.6655.03±1.4764.50±0.9175.00±0.32
Reviewer_PiNNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets a practical pain point in reaction-condition recommendation: candidate reagent libraries keep growing, while existing approaches (closed-set multi-label classification, autoregressive sequence generation, and directly prompting LLMs) are either limited by a fixed output vocabulary or must enumerate candidates into the prompt, making budget-constrained experimental design over an open candidate space difficult. The authors argue that the task is not to reproduce historical conditions but to construct a high-value condition subspace under a finite experimental budget, and that candidate suitability depends jointly on the task and the functional role, motivating explicit hierarchical modeling.

Critical strengths
  1. The problem reformulation is valuable and aligned with experimental practice: recasting condition recommendation as budget-constrained open-candidate retrieval and subspace construction, and using MAMY (the maximum achievable yield within the selected subspace) rather than top-k hit rate as the primary metric (Appendix A.4.2), matches the goal of real high-throughput screening. The formal argument that closed-set cross-entropy learns historical recording probabilities rather than experimental efficacy (Appendix A.1.2, Eqs. (12)–(16)) is more persuasive than a generic critique.
Critical weaknesses

The construction of HTE-OOD undermines the interpretability of MAMY. Of the 9,600 combinations, only 384 are measured and 9,216 are synthetic negatives assigned pseudo-yield zero, so subspace quality largely reflects the ability to exclude obviously inorganic acetate/halide salts, precisely the signal that Morgan fingerprints plus chemical priors capture most easily. Moreover, the budget retains a fraction within each category separately (at 20%, only 20 combinations, about 0.21% of the full space), making MAMY extremely sensitive to a handful of choices; Table 12 shows ChemVD's standard deviation at ±1.93 (20%) versus ±0.00 for Llama-3.3-70B (a single ranking reused), and this heteroscedasticity also affects comparability. Metrics computed only over the 384 measured combinations, or more realistic negative construction, should be reported.

Questions for authors
  1. On LLM baseline fairness: can the DeepSeek/Llama baselines be run under the same interface as ChemVD (e.g., reranking only the top-M candidates per category, or providing K-nearest reference conditions), and what MAMY do they achieve at a 20% budget? How should the "6.7 points" in the main text, the "20.44 percentage points" in the summary, and the three numbers in Appendix Table 12 be reconciled? Given that Llama-3.3-70B outperforms ChemVD at a 20% budget (72.16% vs. 68.78%), how do the authors believe the "parameter-efficiency advantage" should be stated?
8 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (1/7): Evaluation on experimentally measured combinations

We thank the reviewer for recognizing the practical value of budget-constrained subspace design. We also appreciate the constructive feedback on the HTE-OOD evaluation and the LLM comparisons. In response, we provide a complementary evaluation without the synthesized conditions, report a separate task-specific wet-lab study, and evaluate LLMs with retrieved reference conditions. We also clarify and adjust several definitions regarding some comparisons and claims of concern. Below, we provide a response to each comment.

W1.1: Metrics computed only over the 384 measured combinations, or more realistic negative construction, should be reported.

A W1.1: We thank the reviewer for raising this important suggestion. We conducted a complementary evaluation using only the original experimentally measured reaction conditions, excluding all artificially added combinations and their associated pseudo-yields.

As shown in Table R1.2 and Table R1.3, the overall results across the original HTE-OOD subsets demonstrate that ChemVD achieves competitive performance across different selection budgets, with particularly pronounced advantages in challenging subsets. Specifically, we highlighted the Sub3 subset because it represents a highly hard scenario due to the sparse distribution of high-yielding conditions (only 5.21% of conditions achieve yields above 60%, as shown in Table R2.1).

As illustrated in Table R1.3, ChemVD consistently achieves the highest achievable maximum yield across all evaluated selection budgets (20%–80%) compared to large language models (LLMs). This confirms that ChemVD's strong performance is not merely an artifact of binary catalyst filtering, but stems from robust hierarchical variable-space design across all reaction components.

Table R1.1: Yield-interval distribution of original measured combinations in Sub3

yield_interval_percent[0,20)[20,40)[40,60)[60,80)[80,100]
count2209846200
percent57.29%25.52%11.98%5.21%0.00%

Table R1.2: Mean achievable maximum yield of each method on the original measured overall space across different budgets

Budget (Overall)ChemVDPairwiseQwen3-30BQwen3.5-35BLLaMa-3.1-70BLLaMa-3.3-70BDeepSeek-v4-flashDeepSeek-v4-pro
20%68.7860.1565.4658.2971.2672.1645.3048.34
30%73.2662.5067.7261.5472.7174.0557.5159.48
40%77.7472.4776.2966.3175.5177.0871.7475.52
50%79.6274.1079.2069.7480.5781.2276.4479.95
60%81.5174.7380.7970.2083.1482.3979.9581.01
80%87.9384.5284.4677.5885.0985.1985.4784.58

Table R1.3: Achievable maximum yield of each method on the original measured Sub3 space across different budgets

Budget (Sub3)ChemVDPairwiseQwen3-30BQwen3.5-35BLlama-3.1-70BLlama-3.3-70BDeepSeek-v4-flashDeepSeek-v4-pro
20%67.5343.7137.865.3844.4757.6828.918.91
30%67.5743.8037.865.3844.4757.6847.9962.45
40%67.6157.7557.617.0944.4757.6853.1562.45
50%67.6167.5364.2210.2964.2257.6858.2762.45
60%67.6167.5364.2210.2967.4967.4962.4162.45
80%74.8469.9467.6145.0767.6167.6169.167.61
AuthorsOpenReview ↗

Response to Reviewer 1 (2/7): Evaluation Beyond Synthetic Negative Filtering

W1.2 ... Of the 9,600 combinations, only 384 are measured and 9,216 are synthetic negatives assigned pseudo-yield zero, so subspace quality largely reflects the ability to exclude obviously inorganic acetate/halide salts ...

A W1.2: We thank the reviewer for raising this concern. To assess whether ChemVD can distinguish experimentally effective reagents beyond merely excluding synthetic negatives, we conducted a separate evaluation using wet-laboratory data from an unpublished 1,4-diene asymmetric 1,3-difunctionalization reaction newly developed by our collaborators. As this work is unpublished and involves confidential information, we cannot currently disclose the specific reactions, complete reaction-condition records, or reagent structures. We therefore report only the training setup and selected key evaluation results below.

In this reaction, a total of 12 chiral ligands were experimentally tested, among which TY-Phos series ligands, especially TY2, TY3, TY4, and TY5, performed well in the wet-lab experiments, with TY5 as the best-performing ligand. Other TY-phos ligands (TY1, TY6-9) exhibited lower performance, and the remaining Ming-phos (M1), Xu-phos (Xu1), and Xiang-phos (X1) ligands could barely promote the reaction. However, all 12 ligands are plausible choices for chemists prior to experiments, which provides us with a challenging scenario to evaluate ChemVD.

For model training, we collected 177 experimental records from this new reaction system, all involving TY5 as the ligand, and jointly trained ChemVD with the original Suzuki dataset using the same model configuration as in the manuscript. For evaluation, we considered seven reactions, including three represented in the training data and four previously unseen substrate pairs. The inference candidate pool contained 79 reagents, of which 68 were present in the training data, and 11 were newly introduced ligands entirely unseen during training. Specifically, among the 12 candidate ligands, only TY5 appeared in the training data, whereas the remaining 11 ligands, including the experimentally effective TY2, TY3, and TY4, were unseen. This setting allows us to assess whether ChemVD could choose privileged ligands (TY2-TY5) over other ligands. ChemVD and the LLM baselines were evaluated using identical candidate pools and evaluation conditions.

We evaluated recommendation performance using Exact@1 and Family@K. Exact@1 measures whether the top-ranked recommendation matches the experimentally identified reference combination, while Family@K measures the proportion of the top-K recommendations belonging to the experimentally validated ligand family (TY2-TY5). As shown in Table R1.4 ChemVD achieves an Exact@1 of 0.77 and Family@1, Family@3, and Family@4 scores of 1.00, 0.81, and 0.72, respectively, consistently outperforming all evaluated LLM baselines.

These results provide additional evidence that ChemVD can distinguish chemically plausible reagents with different experimental performance, including promising ligands entirely unseen during training, rather than merely excluding obviously infeasible candidates.

Table R1.4: Reagent recommendation performance for 1,4-diene asymmetric 1,3-difunctionalization.

ModelExact@1Family@1Family@3Family@4
Qwen3.5-35B0.080.370.470.48
Qwen3-30B0.000.420.480.49
Llama3.1-70B0.080.480.520.52
Llama3.3-70B0.030.520.480.47
DeepSeek-v4-pro0.210.560.670.65
ChemVD0.771.000.810.72
AuthorsOpenReview ↗

Response to Reviewer 1 (3/7): Screening budget sensitivity

W1.3: Moreover, the budget retains a fraction within each category separately (at 20%, only 20 combinations, about 0.21% of the full space), making MAMY extremely sensitive to a handful of choices.

A W1.3: We thank the reviewer for highlighting the potential sensitivity of MAMY to the screening budget and candidate selection. We agree that MAMY can be sensitive to individual candidate selections under 20% budgets. To verify that our conclusions do not depend on a specific budget, we have evaluated all methods across a broad screening budget range from 20% to 80%. Furthermore, our complementary evaluation restricted to experimentally measured combinations confirms that ChemVD maintains competitive overall performance and consistently achieves the highest MAMY on the challenging Sub3 subset across all evaluated budgets (Table R1.3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chairs and All Reviewers

We sincerely thank the Area Chairs and Reviewers for their time and constructive feedback. To facilitate the evaluation of our submission, we provide a concise overview of the reviewers' main comments and the most significant revisions presented in our responses.

We appreciate the reviewers' recognition of our work:

  • All three reviewers (PiNN, DaPD, and ULRr) recognized the practical relevance and novelty of reformulating reaction condition recommendation as budget-constrained open-candidate retrieval for compact variable-space construction, a paradigm closely aligned with real-world high-throughput experimentation.

  • Reviewers DaPD and ULRr acknowledged our well-motivated learning objectives and strong experimental evidence for hierarchical modeling, particularly the Pairwise ablation and controlled XOR experiment.

  • Reviewers DaPD and ULRr appreciated the extensive baseline comparisons and ablation studies, as well as the careful experimental design and transparent reporting.

We also thank the reviewers for their constructive suggestions. The main concerns centered on the encoder selection procedure, the interpretation of HTE-OOD results, the fairness of baseline comparisons, and the evidence for generalization beyond familiar reagents and fixed role requirements. We have carefully addressed all comments in a detailed point-by-point response and updated our manuscript accordingly. Here, we summarize five core revisions, specifically the inclusion of key methodological clarifications, new experiments and corresponding analyses, as these changes most directly improve the soundness of the paper:

  1. Clarified the Encoder Selection Procedure and Experimental Chronology. We clarified that the final encoder configuration was determined strictly from chemically motivated representation design and preliminary comparisons on an independent validation set, prior to any evaluation on either test benchmark. We further provided timestamped experimental records documenting the chronology of preliminary encoder comparisons, final model selection, and subsequent benchmark evaluations. Importantly, we distinguished the original encoder selection procedure from the later robustness comparisons reported in Table 13 and revised the corresponding section of the Supplementary Information (originally Appendix A.5.5) to eliminate any ambiguity regarding test-set usage (Reviewer 2 response).

  2. Broadened Evaluation Across Role Requirements and Reaction Families. We expanded category-coverage evaluation to include an additional reaction family (Indole Synthesis) with differing role requirements. We further evaluated ChemVD on unpublished wet-laboratory data from a newly developed 1,3-difunctionalization reaction using task-specific candidate reagents. ChemVD successfully prioritizes previously unseen yet experimentally effective reagents and outperforms all evaluated LLM baselines. These results further support ChemVD’s generalization across reaction families and its ability to identify chemically plausible candidates.

  3. Evaluated Reagent Selection on Experimentally Measured Combinations. We repeated the evaluation on all experimentally measured combinations, removing the synthetic negatives. With the catalyst fixed, ChemVD remains competitive with the strongest zero-shot LLM baseline on the overall set and outperforms it on more challenging subsets where high-yield conditions are sparse.

  4. Strengthened Evaluation of Structural Generalization. We strengthened the structural-generalization evaluation of unseen reagents with a maximum Tanimoto similarity below 0.5 and 0.7 to any training reagent. ChemVD outperforms Pairwise across all evaluated retrieval metrics. (Reviewer 2 response, Table R2.6 and R2.7).

  5. Expanded Baseline Diagnostics and LLM Comparisons. We added two BCE baselines with explicit query–candidate interaction (BCE-int) and the same similarity scorer as ChemVD (BCE-scorer), and evaluated their behaviors on Reagent-Split. For LLM baselines, we supplemented the few-shot evaluation under the original interface with nearest training-set reference positive or negative conditions. (Reviewer 2 response, New Table 1 (Excerpt) and Table R2.10; Reviewer 1 response, Table R1.5; Reviewer 2 response, Table R2.9).

We hope that these additional experiments and clarifications will assist the reviewers and Area Chair in their final assessment of our work. We sincerely thank them for their time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.