acceptodds
Under review as a conference paper at ICLR 2027

CAMS-Diff: Coupled Adaptive Mixed-Space Diffusion for Molecular Translation

25 views

Abstract

Diffusion models are increasingly used for molecular translation, including molecule captioning, text-guided generation, forward reaction prediction, and retrosynthesis. Continuous sequence diffusion provides a common formulation across these tasks, with molecules represented through SMILES. However, two limitations remain. (1) At low noise levels, continuous representations can retain substantial local token information, reducing reliance on broader sequence context and the input condition. (2) During inference, uniform timestep reduction can also skip informative regions of a non-uniform denoising trajectory. We introduce Coupled Adaptive Mixed-Space Diffusion (CAMS-Diff). During training, CAMS-Diff couples token-wise adaptive Gaussian noising with a discrete representation-masking process governed by the same learned schedule, selectively masking direct local token evidence. During inference, it profiles reconstruction changes in token-wise log-SNR space and uses the resulting trajectory importance to select denoising timesteps under a fixed sampling budget. Across five molecular translation and generalization settings, CAMS-Diff improves molecular and semantic recovery; on ChEBI-20, it reaches 0.916 MACCS compared with 0.883 for BiMol-Diff. Controlled ablations isolate both mechanisms: coupling outperforms adaptive noising, masking, and their uncoupled combination, while trajectory-informed timestep selection outperforms uniform subsampling at matched budgets. Two-step captioning, using two denoiser evaluations instead of 2000, achieves quality comparable to full-step BiMol-Diff with an approximately 245x wall-clock speedup. CAMS-Diff also transfers zero-shot to PCDes, improving MACCS from 0.763 to 0.810.

Primary area:
generative models
Keywords:
Molecular Diffusion Models, Coupled Mixed-Space Diffusion, Adaptive Noising, Trajectory-Informed Decoding, Molecular Translation, Text-to-Molecule Generation, Reaction Prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_M5K1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAMS-Diff, a coupled adaptive mixed-space diffusion framework for molecular translation. The method introduces two main components: (1) coupling token-wise adaptive Gaussian noising with discrete representation masking under a shared learned schedule, and (2) trajectory-informed timestep selection for efficient reduced-step inference. The framework is evaluated across molecule captioning, text-guided molecule generation, forward reaction prediction, retrosynthesis, and zero-shot generalization. Experimental results demonstrate improvements in molecular similarity and semantic reconstruction over several baselines. Ablation studies further support the effectiveness of the proposed coupling mechanism and adaptive timestep allocation.

Critical strengths
  1. Effective integration of complementary mechanisms. The coupling of adaptive Gaussian noising and discrete masking is well motivated. The ablation results in Table 5 demonstrate consistent improvements over individual and uncoupled variants, supporting the effectiveness of the proposed design.
  2. Comprehensive evaluation and promising efficiency. The method is evaluated across multiple molecular translation tasks and zero-shot transfer. It achieves competitive improvements in molecular similarity and demonstrates substantial inference acceleration through trajectory-informed reduced-step decoding.
Critical weaknesses
  1. Reduced molecular validity. CAMS-Diff achieves lower SMILES validity than several baselines, despite improvements in molecular similarity.
  2. Model capacity differences. Some baseline comparisons use different parameter counts, potentially affecting the fairness of performance comparisons.
  3. Additional profiling overhead. Trajectory-informed timestep selection requires full-step validation profiling, and this computational cost is excluded from the reported inference speedups.
  4. Limited theoretical justification. The timestep allocation theory relies on a continuous approximation, while the practical finite-step scheduling strategy remains heuristic.
Questions for authors

Please see the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to W1

W1. Reduced Molecular Validity


We thank the reviewer for highlighting this concern. We acknowledge the lower raw SMILES validity and have conducted additional error analysis and deterministic repair experiments, detailed in our response to Reviewer CBtW (W1).

As shown in Figure 10 (Appendix G), invalid generations are dominated by SMILES serialization errors, including unpaired ring labels (43.95%), unclosed branches (41.35%), and extra closing parentheses (37.86%), whereas only 4.87% involve invalid atomic valence (categories may overlap).

Importantly, a simple, deterministic correction of missing closing parentheses recovers 248 of 3,300 predictions, improving validity from 0.748 to 0.823 (+7.5 percentage points) without retraining. This exceeds the uncorrected diffusion baselines in Table 3, while retaining strong MACCS similarity (0.899) and unchanged exact match (0.232).

Furthermore, as discussed in Reviewer CBtW (W1), validity alone does not establish correct molecular generation. CAMS-Diff achieves higher exact match on retrosynthesis, forward prediction, and zero-shot generation (Tables 4 and 6), supporting its downstream effectiveness despite lower raw validity.

We refer the reviewer to our detailed response to CBtW (W1) for the matched-validity evaluation, error analysis, and repair results, which we will incorporate into the revised manuscript.

AuthorsOpenReview ↗

Response to W2

W2. Model Capacity Differences


To address the fairness concern, we performed additional capacity-controlled experiments, reducing CAMS-Diff from 180M to 125M parameters for text-guided molecule generation and from 91M to 60M for molecule captioning.

(1) Text-guided molecule generation (ChEBI-20).

We reduced CAMS-Diff to 125M parameters, matching UTGDiff and TGM-DLM. We additionally evaluated both CAMS-Diff configurations with the deterministic SMILES repair described in W1.

Metric ↑BiMol-Diff (180M)CAMS-Diff (125M)125M + repairCAMS-Diff (180M)180M + repair
MACCS0.8830.9060.8890.9160.899
RDK0.7850.8030.8080.8110.807
Morgan0.7580.7820.7720.7900.780
Validity0.8120.7480.8230.7480.823

Despite reducing parameter count by approximately 31%, the 125M configuration retains strong molecular similarity, outperforming the reproduced 180M BiMol-Diff baseline across all three fingerprint metrics. Moreover, SMILES repair increases validity to 0.823 for both CAMS-Diff configurations, exceeding BiMol-Diff's 0.812, while preserving superior fingerprint similarity.

(2) Molecule captioning (M3-20M).

We further reduced CAMS-Diff from 91M to 60M parameters, closely matching the reproduced 63M-parameter BiMol-Diff baseline.

Metric ↑BiMol-Diff (63M)CAMS-Diff (60M)CAMS-Diff (91M)
BLEU0.5570.6000.610
chrF++0.7100.7090.705
METEOR0.6320.6900.740
BERTScore-F10.8380.8490.861
MAUVE0.9130.9180.927

Despite having fewer parameters than BiMol-Diff, the 60M CAMS-Diff configuration outperforms it on four of the five captioning metrics, while achieving nearly identical chrF++ performance.

These capacity-controlled results complement our existing same-architecture component ablations, which isolate the effects of adaptive noising and coupled masking without changing parameter count (see our response to Reviewer CBtW, W2/Q1).

Together, these experiments demonstrate that CAMS-Diff retains its principal improvements at comparable or smaller model capacities, supporting the methodological contributions independently of model size. We will incorporate these experiments and include capacity-controlled comparisons for the remaining tasks in the revised manuscript.

AuthorsOpenReview ↗

Response to W3

W3. Additional Profiling Overhead


We thank the reviewer for raising this concern. Trajectory profiling is a one-time, offline procedure performed after training; it does not contribute to test-time generation latency. We quantify both its cost and the quality it buys.

(1) Profiling reuses the existing validation pass. Measuring denoising progress requires a full-step reverse pass over the validation set. This is the same pass already performed after training to assess generation quality; we additionally log position-wise reconstruction errors at each retained timestep, requiring no additional denoiser evaluations. Errors are accumulated as described in Appendix C.1.

(2) Remaining stages require no denoiser evaluations. Let VV be the validation-set size, T=2000T = 2000 the full budget, ∣A∣|\mathcal{A}| the retained target positions, K≪TK \ll T the reduced budget, LL the output length, dd the prediction dimension, and CDC_D the cost of one denoiser evaluation.

StageOperationTime complexity
1. Denoising progressFull-step validation and error accumulationO(VTCD+VTLd)\mathcal{O}(VTC_D + VTLd), shared with validation
2. Local importanceFit error curves, compute local importanceO(∣A∣T)\mathcal{O}(\lvert\mathcal{A}\rvert T)
3. Shared importanceAggregate into p(t)p(t)O(∣A∣T)\mathcal{O}(\lvert\mathcal{A}\rvert T)
4. Timestep allocationSelect K−1K-1 interior steps from ≈ 2K{\approx}\,2K binsO(T+Klog⁡K)\mathcal{O}(T + K\log K)

Once validation statistics are available, the incremental scheduling cost is

Cschedule(K)=O ⁣(∣A∣T+T+Klog⁡K).C_{\mathrm{schedule}}(K) = \mathcal{O}\!\left(\lvert\mathcal{A}\rvert T + T + K\log K\right).

(3) Measured overhead (K=4K=4, single NVIDIA A100).

TaskPositionsStage 2Stage 3Stage 4Total
Captioning64104.1 ms0.6 ms0.1 ms0.105 s
Text-guided generation256406.2 ms2.3 ms0.1 ms0.409 s
Forward prediction128204.8 ms1.2 ms0.1 ms0.206 s
Retrosynthesis128204.8 ms1.2 ms0.1 ms0.206 s

(4) Overhead relative to generation cost, and the quality it buys. The scheduling cost is incurred once per task and does not scale with the number of samples generated. We report it against the cost of generating 100 samples, alongside the quality obtained at the same denoising budget.

TaskSampling / sample/ 100 samplesOverheadShare at 100Share at 1,000Quality gain at K=4K=4
Captioning0.0091 s0.91 s0.105 s11.5%1.2%+1.00 BLEU pp (DDIM +0.80)
Text-guided generation0.079 s7.90 s0.409 s5.2%0.5%+0.67 MACCS pp (DDIM +0.40)
Forward prediction0.102 s10.20 s0.206 s2.0%0.2%+0.50 validity pp (DDIM +1.50)
Retrosynthesis0.080 s8.00 s0.206 s2.6%0.3%+1.90 validity pp (DDIM +1.80)

Importantly, the two arms consume identical sampling compute: both perform exactly KK denoiser evaluations per sample and differ only in where those evaluations are placed along the reverse trajectory. The overhead is therefore not an inference-time latency trade-off, but a one-time offline cost after which every generated sample is improved at no recurring expense. It amortizes to below 1.5% within 1,000 samples on all four tasks, and the same importance profile is reused across sampling budgets KK, with the constructed schedule fixed across all test examples.

(5) On the reported speedups. These measure test-time generation latency. Since schedule construction is performed once offline and the retained timesteps are fixed thereafter, inference requires exactly KK denoiser evaluations per sample with no recurring profiling cost. We will state this distinction, the stage-wise complexity, and the measured overhead explicitly in Appendix C.1 of the revised manuscript.

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

The paper proposes CAMS-Diff, a continuous sequence diffusion framework for molecular translation tasks. The model makes two major modifications to the conventional diffusion framework: (1) apply discrete masking to encourage context learning during training (2) define and use more "informative" timesteps, i.e. those that show major changes in clean data reconstruction error, during inference under limited budget. Benchmarking shows improved performance across five translation tasks and good balance between sampling efficiency and generation quality.

Critical strengths
  1. The paper identifies and addresses two critical limitations in continuous sequence diffusion models for molecular-related tasks. The methodology and experiments are clearly explained and well-motivated, Both proposed components produce notable benefits across a broad range of tasks and datasets.
Critical weaknesses
  1. CAMS-Diff shows notably weakened molecular validity in several tasks, such as text-guided generation on ChEBI-20 (Table 3) and retrosysnthesis (Table 4). This poses concerns on how the model can be reliably deployed in real downstream tasks.
  2. Despite improved performances in several downstream metrics, it remains to be justified whether certain design choices achieve the stated goals. For example, whether the coupled masking really improve the context reliance needs to be demonstrated.
Questions for authors
  • How to verify that the coupled masking really encourage the reliance on contextual information?
  • On ChEBI-20, why does the overall higher molecular similarity on ChEBI-20 not translate into higher validity?
  • Does the trajectory-informed scheduling improve diffusion-based models in general as well?
6 comments
AuthorsOpenReview ↗

Response to W1

W1. CAMS-Diff shows notably weakened molecular validity in several tasks, such as text-guided generation on ChEBI-20 (Table 3) and retrosysnthesis (Table 4). This poses concerns on how the model can be reliably deployed in real downstream tasks.


We agree that molecular validity is important. However, RDKit-based validity measures SMILES parseability rather than exact recoverability of the intended molecule. Notably, even when the model predicts an empty string, RDKit considers it valid under a parser-based validity check. Therefore, exact match is a stricter and more task-relevant measure of successful target generation, since a valid but incorrect molecule does not satisfy the intended prediction.

(1) Higher exact-match performance across downstream tasks.

CAMS-Diff achieves higher exact match on retrosynthesis (0.541 vs. 0.462; Table 4), forward reaction prediction (0.830 vs. 0.828; Table 4), and zero-shot generation (0.427 vs. 0.386; Table 6) compared with UTGDiff, despite lower validity. For text-guided generation, it remains competitive in exact match (0.232 vs. 0.227 for UTGDiff and 0.242 for BiMol-Diff), while achieving the highest molecular similarities (MACCS 0.916, RDK 0.811, Morgan 0.790; Table 3).

Additional matched-validity evaluation. To isolate exact-match performance from differences in validity, we evaluate both models on the same test instances where both generate valid SMILES. Starting from the full test sets (3,300 ChEBI-20, 1,000 forward, and 1,000 retrosynthesis instances; Table 9), we identify their jointly valid subsets, containing 1,848, 879, and 795 instances, respectively. Thus, both models achieve 100% validity on these identical subsets, enabling a direct comparison of exact match and Levenshtein distance.

Task (jointly valid instances)UTGDiff Exact ↑CAMS-Diff Exact ↑UTGDiff Levenshtein ↓CAMS-Diff Levenshtein ↓
ChEBI-20 (n=1,848)0.3560.3917.644.33
Forward (n=879)0.89760.92492.431.14
Retro (n=795)0.55720.64539.695.59

Under identical validity conditions, CAMS-Diff consistently achieves higher exact match and lower Levenshtein distance, demonstrating stronger target recovery among jointly valid predictions. This analysis complements the full-test results in Tables 3–4.

(2) Understanding why validity is lower and how it can be improved.

Despite achieving higher molecular similarity, CAMS-Diff exhibits lower validity primarily due to SMILES serialization errors rather than invalid atomic valence. As shown in Figure 10 (Appendix G), invalid generations frequently involve unpaired ring labels (43.95%; e.g., C1CC, with an unmatched ring index), unclosed branches (41.35%; e.g., C(C, with a missing closing parenthesis), and extra closing parentheses (37.86%). In contrast, only 4.87% involve invalid atomic valence (categories may overlap).

Importantly, some of these errors can be corrected through inexpensive post-processing. For example, adding missing closing parentheses to generations with unclosed branches, through deterministic, reference-independent correction without retraining, recovers 248 of 3,300 predictions (7.5%). This improves validity from 0.748 to 0.823, exceeding the other diffusion baselines in Table 3 (UTGDiff: 0.812; BiMol-Diff: 0.812; TGM-DLM: 0.789), while retaining strong MACCS similarity (0.899).

Regarding reliable downstream deployment, higher validity alone does not guarantee generation of the correct molecular target. CAMS-Diff's higher exact match on retrosynthesis, forward prediction, and zero-shot generation makes it a stronger choice for these downstream tasks, while its competitive text-guided exact match and higher molecular similarity further support its utility. We will incorporate these additional analyses into the revision.

AuthorsOpenReview ↗

Response to W2 and Q1

[W2 & Q1]. (W1) Despite improved performances.....? and (Q1) How to verify that the coupled...?


We thank the reviewer for this suggestion. We clarify how coupled masking encourages reliance on contextual information and provide additional experiments to verify whether generated molecules follow the supplied conditioning instructions.

(1) How coupled masking encourages contextual reliance.

As discussed in Section 4.1, continuous diffusion can retain substantial local token information, particularly at low noise levels. This allows the denoiser to recover tokens directly from their noisy representations without sufficiently using surrounding context or conditioning information.

CAMS-Diff addresses this through two complementary mechanisms:

  • Removing local token information (Eq. 3): Masking replaces the noisy token representation with a learned mask embedding. Since the original token information is no longer directly available at that position, the denoiser is encouraged to use surrounding tokens and the conditioning input (Eq. 5).
  • Coupling masking with adaptive noising (Eq. 2): Rather than applying masking independently, CAMS-Diff determines the masking probability using the learned token-wise Gaussian noise schedule. Thus, masking is coordinated with the noise level at each position, encouraging the model to recover missing information using the available context.

Existing results support this explanation. In Table 5, the uncoupled baseline uses the same overall masking rate but applies masking independently of the learned schedule. Coupling improves captioning BLEU from 0.597 to 0.610, BERTScore-F1 from 0.841 to 0.861, and molecular generation MACCS from 0.909 to 0.916. Furthermore, Figure 4c and Appendix G show increasingly large exact-match improvements for longer instructions, consistent with better contextual utilization. However, these results alone do not directly establish reliance on conditioning information.

(2) Verifying contextual reliance through instruction substitution.

To directly verify whether the model follows the supplied conditioning information, we select 500 ChEBI-20 test examples and replace each original molecular description XX with a different description X′X'. The model then generates a new molecule Y^′\hat{Y}'. If the model utilizes contextual information, the generated molecule should align with X′X' rather than the original description XX.

We evaluate this using Text2Mol (EMNLP 2021), a cross-modal metric that measures semantic alignment between a molecule and its natural-language description.

Masking strategyAlignment with original XX ↓Alignment with substituted X′X' ↑
Uniform0.1740.579
Adaptive0.1700.598
Uncoupled0.1680.617
Coupled (CAMS-Diff)0.1610.643

CAMS-Diff achieves the highest alignment with the substituted instruction (0.643) and the lowest alignment with the original instruction (0.161). This demonstrates that its generated molecules are more consistent with the newly supplied conditioning information than with the original description.

Furthermore, compared with uncoupled masking, CAMS-Diff achieves higher alignment with the new instruction (0.643 vs. 0.617) and lower alignment with the original instruction (0.161 vs. 0.168). These results provide direct evidence of sensitivity to conditioning changes and support more effective contextual utilization through coupling.

(3) Additional Text2Mol evaluation on the full test set.

We further evaluate Text2Mol on the entire ChEBI-20 test set using the original conditioning instructions.

Masking strategyText2Mol ↑
Uniform0.584
Adaptive0.589
Uncoupled0.591
Coupled (CAMS-Diff)0.599

Coupled masking achieves the highest alignment, outperforming uniform (+0.015), adaptive (+0.010), and uncoupled (+0.008) masking. It also performs better compared to other SoTA baselines: MolT5-Large (0.554), UTGDiff (0.587), and TGM-DLM without corrector (0.589). These results provide complementary evidence of improved molecule–text semantic alignment across the full test set.

Overall, these findings connect the methodology to the observed behavior: coupled masking removes direct local token information (Eq. 3), improves performance over independently applied masking at a matched overall masking rate (Table 5), and generates molecules that align more strongly with newly supplied instructions while showing less alignment with the original instructions. Together with the full-test Text2Mol improvements, these results provide a mechanistic explanation and empirical support for more effective contextual utilization through coupled masking. We will incorporate these analyses into the revision.

AuthorsOpenReview ↗

Response to Q2

Q2. Why does higher molecular similarity not translate into higher validity?


As discussed in W1, molecular similarity and validity measure different properties. Fingerprint similarity captures structural agreement with the target molecule, whereas validity requires the generated SMILES to satisfy molecular parsing and serialization constraints.

Despite achieving the highest molecular similarities on ChEBI-20 (MACCS 0.916, RDK 0.811, Morgan 0.790; Table 3), CAMS-Diff exhibits lower validity primarily due to SMILES serialization errors, including unpaired ring labels (43.95%) and unclosed branches (41.35%), while only 4.87% of invalid generations involve atomic valence violations (Figure 10, Appendix G).

As demonstrated in W1, inexpensive deterministic correction of these serialization errors recovers 248 of 3,300 predictions, improving validity from 0.748 to 0.823 while retaining strong MACCS similarity (0.899) and unchanged exact match (0.232).

Thus, higher molecular similarity does not necessarily translate into higher validity because structurally similar predictions can still violate SMILES syntax constraints. The additional repair experiment demonstrates that a portion of this gap can be addressed without retraining.

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

CAMS-Diff is a continuous SMILES/text diffusion model for molecular translation. However, its seems to be that this is rather a BiMol-Diff++ rather than a new method. It adds a per-token mask following a learned noise schedule (which has already been presented in previous works such as MELD), and a few-step decoding pipeline (which also has a wide range of prior approaches). The versatility and range of conducted experiments seem to be the major strengths of this paper, rather than the method itself.

Critical strengths

Broad empirical analysis, with one formulation evaluated both on captioning, text-to-molecule, forward, retrosynthesis, and zero-shot transfer. Multi-seed also helps with deciphering real gains from seed specialization.

Critical weaknesses

Limited novelty: the backbone and the token-wise adaptive schedule are taken from BiMol-Diff, the masking only reuses that schedule as a mask rate (close to MELD https://arxiv.org/abs/2505.16790 learned per-element masking, not cited), and the few-step scheduler is a post-hoc sampler change. It seems to me that the novelty itself is quite thin, at least in the methodological perspective.

Gains over BiMol-Diff are not attributable to the method. CAMS-Diff seems to be larger in terms of parameters (180M vs 125M, 91M vs 63M) with no parameter-matched comparisons. Furthermore, the 245× speedup is measured against 2000-step BiMol-Diff rather than BiMol-Diff with the same few-step solver. It would be much appreciated if the authors provide a fair comparison with the baselines, at LEAST with BiMol-Diff.

Questions for authors

Refer to above. Clear explanation of novelty and gains over BiMol-Diff are necessary

7 comments
AuthorsOpenReview ↗

Response to W1

W1. Limited Methodological Novelty


We thank the reviewer for highlighting MELD, which we will cite and discuss in the updated manuscript. CAMS-Diff introduces two methodological contributions distinct from BiMol-Diff: (1) coupled adaptive mixed-space diffusion and (2) trajectory-informed timestep allocation.

(1) Coupled adaptive mixed-space diffusion.

Both BiMol-Diff (ACL 2026) and MELD (ICLR 2026) demonstrate that uniform corruption is suboptimal for molecular modeling, but operate in a single diffusion space: BiMol-Diff uses reconstruction-difficulty-adaptive Gaussian corruption, while MELD learns element-wise masking trajectories in discrete graph diffusion.

DiffuSeq-v2 (EMNLP 2023) on the other hand demonstrated that combining continuous Gaussian diffusion with discrete masking improves (i) conditional reconstruction and (ii) sampling efficiency. However, simply combining these mechanisms does not ensure improvements for molecules. In Table 5, adding masking to uniform Gaussian diffusion improves captioning BLEU (0.523 → 0.557) but slightly reduces generation MACCS (0.874 → 0.870).

Motivated by this, CAMS-Diff introduces coupled adaptive mixed-space diffusion, coordinating Gaussian corruption and discrete masking through a shared learned token-wise schedule: ϖti=η1−αˉti.\varpi_t^i=\eta\sqrt{1-\bar{\alpha}_t^i}. Since SNRti=αˉti/(1−αˉti)\mathrm{SNR}_t^i=\bar{\alpha}_t^i/(1-\bar{\alpha}_t^i), the square-root factor is the per-token Gaussian noise standard deviation. Masking therefore follows the same learned signal degradation, without an independently learned masking-schedule network. At η=0\eta=0, BiMol-Diff's Gaussian-only corruption process is recovered.

Evidence (i): Conditional reconstruction. Relative to non-adaptive mixed-space corruption, CAMS-Diff improves captioning BLEU 0.557 → 0.610 and generation MACCS 0.870 → 0.916 (Table 5). Importantly, coupling outperforms independently combined adaptive noising and masking at a matched masking rate (BLEU 0.597 → 0.610; MACCS 0.909 → 0.916). Instruction-substitution experiments further show stronger conditioning alignment (Text2Mol 0.643 vs. 0.617; Reviewer CBtW, W2/Q1).

Evidence (ii): Quality retention under reduced step sampling. Mixed-space training exposes the denoiser to missing local token evidence, encouraging reconstruction from context and conditioning.

ModelStepsCaptioning BLEU ↑
BiMol-Diff20000.557
BiMol-Diff1000.312
BiMol-Diff + DPM-Solver++20.303
BiMol-Diff + our allocation20.319
CAMS-Diff20000.610
CAMS-Diff + our allocation20.568

BiMol-Diff loses approximately 47% of its full-step BLEU under two-step sampling, whereas CAMS-Diff retains approximately 93%, achieving quality comparable to full-step BiMol-Diff with only two evaluations. This substantial difference cannot be explained by timestep allocation alone; controlled coupling ablations (Table 5) and capacity-matched experiments (W2) further support the distinct training formulation.

(2) Trajectory-informed timestep allocation.

Our second contribution is independent of mixed-space training. Rather than introducing another solver, we derive a reconstruction-informed principle for selecting denoising evaluations under a fixed budget (Proposition 1): pi∗(λ)∝[−eλdEi(λ)dλ]+.p_i^*(\lambda)\propto\sqrt{\left[-e^\lambda\frac{dE_i(\lambda)}{d\lambda}\right]_+}. This principle also works on BiMol-Diff without retraining. At K=2K=2 with DPM-Solver++, it improves BiMol-Diff captioning BLEU 0.303 → 0.319 and generation MACCS 0.870 → 0.881. On CAMS-Diff, it improves two-step captioning BLEU by 1.8 points over uniform allocation. Independently trained uniform and adaptive diffusion models additionally gain 1.5–2.1 MACCS points across DDIM and DPM-Solver++ (Reviewer CBtW, Q3).

Thus, CAMS-Diff differs methodologically from BiMol-Diff at two levels: its training formulation couples continuous and discrete corruption rather than relying solely on Gaussian noising, while its inference formulation uses reconstruction-informed timestep allocation rather than conventional timestep selection. The latter also improves BiMol-Diff itself, demonstrating that the two contributions are independently applicable rather than simply modifications to an existing sampler.

AuthorsOpenReview ↗

Response to W2

W2. Attribution of Gains and Fair Comparison with BiMol-Diff


We provide (1) parameter-matched and (2) reduced-step comparisons.

(1) Parameter-matched comparisons.

First, the 125M-parameter model in Table 3 is UTGDiff, not BiMol-Diff. For molecule generation, both BiMol-Diff and CAMS-Diff use 180M parameters (MACCS: 0.883 vs. 0.916).

We also evaluated smaller CAMS-Diff configurations to isolate the effect of model capacity.

Molecule captioning (M3-20M):

ModelParametersBLEU ↑BERTScore-F1 ↑
BiMol-Diff63M0.5570.838
CAMS-Diff (smaller)60M0.6000.849
CAMS-Diff (original)91M0.6100.861

Text-guided molecule generation (ChEBI-20):

ModelParametersMACCS ↑
UTGDiff125M0.867
BiMol-Diff180M0.883
CAMS-Diff (smaller)125M0.906
CAMS-Diff (original)180M0.916

CAMS-Diff outperforms BiMol-Diff even with fewer parameters, demonstrating that the improvements are not simply due to model size.

(2) Fair reduced-step comparison.

We agree that the reported 245× speedup compares two-step CAMS-Diff with full 2000-step BiMol-Diff, rather than BiMol-Diff using the same reduced-step solver. To address this, we additionally evaluated both models using DPM-Solver++ at the same two-step budget.

ModelStepsCaptioning BLEU ↑Generation MACCS ↑
BiMol-Diff (full-step)20000.5570.883
BiMol-Diff + DPM-Solver++20.3030.870
BiMol-Diff + our allocation20.3190.881
CAMS-Diff (full-step)20000.6100.916
CAMS-Diff + DPM-Solver++20.5500.895
CAMS-Diff + our allocation20.5680.913

At the same solver and sampling budget, CAMS-Diff achieves substantially higher captioning BLEU (0.550 vs. 0.303) even without trajectory-informed allocation. Applying our allocation further improves both models, demonstrating that its benefit is independent of the training formulation.

Together, the parameter-controlled and matched-solver experiments show that CAMS-Diff's improvements are not explained by additional parameters or an unfair sampling-budget comparison. The controlled ablations in Table 5 further isolate the contribution of coupled adaptive mixed-space corruption. We will include these additional results in the revised manuscript.

AuthorsOpenReview ↗

Additional Clarification on W1

Additional Clarification on W1: Distinction from BiMol-Diff


We further clarify three distinctions between CAMS-Diff and BiMol-Diff, focusing on the coupled mixed-space formulation.

(1) Adaptive coupling formulation. BiMol-Diff [1] employs token-wise adaptive Gaussian noising, but the noisy token representations can still retain local information, particularly at low noise levels [2]. CAMS-Diff, on the other hand, jointly controls Gaussian noising and masking through the same learned token-wise schedule (Eq. 2, Section 4.1). This allows the denoiser to learn from both partially informative Gaussian-noised representations and masked positions where direct local evidence is unavailable. Unlike MELD's discrete-space masking [3], masking is coupled directly to continuous Gaussian signal degradation. The importance of coupling is demonstrated by the controlled ablation: compared with independently applied masking at a matched overall masking rate, CAMS-Diff improves captioning BLEU (0.597 → 0.610) and generation MACCS (0.909 → 0.916) (Table 5).

(2) Increased reliance on sequence context and conditioning. Unlike BiMol-Diff's Gaussian-only formulation, CAMS-Diff selectively masks local token evidence, encouraging reconstruction from broader sequence context and the conditioning input (Eqs. 3–5). On ChEBI-20, CAMS-Diff achieves higher Text2Mol alignment than BiMol-Diff (0.599 vs. 0.586). Instruction-substitution experiments further show stronger alignment with the substituted condition than uncoupled masking (0.643 vs. 0.617; Reviewer CBtW, W2/Q1).

(3) Better quality–efficiency trade-off under few-step decoding. Under the same two-step DPM-Solver++ setting, without trajectory-informed timestep allocation, CAMS-Diff retains 90.2% of its full-step BLEU (0.610 → 0.550), compared with 54.4% for BiMol-Diff (0.557 → 0.303). CAMS-Diff therefore achieves substantially higher reconstruction quality at the same two-evaluation budget, independently of the proposed timestep-allocation method (Section 4.2).

Together, these results distinguish CAMS-Diff from BiMol-Diff through (1) adaptive coupling of Gaussian noising and masking, (2) increased reliance on sequence context and conditioning, and (3) a stronger quality–efficiency trade-off under few-step decoding.


References

[1] BiMol-Diff: A Unified Diffusion Framework for Molecular Generation and Captioning. ACL 2026.

[2] Diffusion-LM Improves Controllable Text Generation. NeurIPS 2022.

[3] Learning Flexible Forward Trajectories for Masked Molecular Diffusion (MELD). ICLR 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Consolidated Rebuttal Summary and Additional Experimental Evidence

Dear Area Chair,

We thank you and the reviewers for the constructive feedback. During the rebuttal, we added experiments, clarified the contributions, and strengthened baseline comparisons. The main concerns and evidence are summarized below.

Summary of Rebuttal and Additional Evidence

ConcernEvidence (rebuttal reference)
1. Novelty (EveE W1)Two contributions: coupled adaptive mixed-space diffusion and trajectory-informed timestep allocation. Coupling Gaussian noising and masking through a shared learned schedule improves BLEU 0.597 → 0.610 and MACCS 0.909 → 0.916 over uncoupled masking at a matched rate (Table 5).
2. Capacity and fairness (EveE W2, M5K1 W2)Capacity-matched: 180M CAMS-Diff vs. BiMol-Diff, MACCS 0.916 vs. 0.883; 125M vs. UTGDiff, 0.906 vs. 0.867; 60M vs. 63M BiMol-Diff (captioning), BLEU 0.600 vs. 0.557. Gains are not due to capacity alone.
3. Few-step comparison (EveE W2)Same two-step DPM-Solver++, both without our allocation: BLEU 0.550 vs. 0.303; with allocation applied to both: 0.568 vs. 0.319. The 245× speedup compares two-step CAMS-Diff with 2000-step BiMol-Diff, not matched steps (§5, Fig. 3).
4. Contextual reliance (CBtW W2/Q1)Under instruction substitution, Text2Mol 0.643 vs. 0.617 for uncoupled masking; full-test Text2Mol 0.591 → 0.599 and BERTScore-F1 0.841 → 0.861 (Table 5).
5. Validity (CBtW W1/Q2, M5K1 W1)Invalid SMILES are mainly unpaired ring labels (43.95%) and unclosed branches (41.35%). Deterministic repair recovers 248/3,300 predictions, raising validity 0.748 → 0.823 (vs. 0.812 for uncorrected diffusion baselines), with MACCS 0.899 and unchanged exact match (App. G).
6. Matched-validity recovery (CBtW W1)On subsets where both models are valid, exact match vs. UTGDiff improves on ChEBI-20 (0.356 → 0.391), forward prediction (0.898 → 0.925), and retrosynthesis (0.557 → 0.645), with lower Levenshtein distance.
7. Generality of allocation (CBtW Q3, M5K1 W4)On independently trained uniform and adaptive diffusion models, at matched budgets: +1.5 to +2.1 MACCS (DPM-Solver++) and +1.9 (DDIM), across all four tasks (App. C.3).
8. Theory (M5K1 W4)Clarified Proposition 1 (App. C.2) and the finite-step approximation (Eq. 9, App. C.4); fixed early/middle/late schedules do not consistently match our allocation (App. H, Fig. 11).
9. Profiling overhead (M5K1 W3)Profiling reuses the full-step validation pass with no extra denoiser evaluations; scheduling takes 0.105–0.409 s per task (<1.5% amortized over 1,000 samples) and is not repeated at inference (App. C.1).

Reviewer Reassessment

ReviewerInitialCurrentOutcome
EveE13Raised rating, noting robust additional gains.
CBtW33Confirmed questions were adequately addressed.
M5K133Maintained positive rating, but did not engage in rebuttal.

Average rating increased from 2.33 to 3.00; all reviewers now recommend acceptance.

We would be grateful if the Area Chair would consider the additional evidence and the reviewers' reassessments in the final recommendation.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.