Pith. sign in

REVIEW 3 major objections 4 minor 56 references

ABOPD claims that supervising a diffusion antibody generator on its own reverse-denoising states, using a frozen teacher with privileged native backbone geometry, reduces CDR-H3 RMSD from 2.37 Å to 1.95 Å and improves all six CDRs—outperfor

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

ABOPD cuts CDR-H3 backbone RMSD from 2.37 Å to 1.95 Å via on-policy coordinate distillation from a teacher with privileged native geometry.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid controlled application of on-policy distillation to antibody CDR design with real RMSD gains; missing ablation against native-coordinate targets leaves the distillation attribution unproven. the 3 major comments →

arxiv 2607.18835 v1 pith:ZEMM5YSW submitted 2026-07-21 cs.LG cs.AI

ABOPD: Antibody CDR Design via On-Policy Distillation

Zhuo Yang , Jiaying He , Jiaqing Xie , Daolang Wang , Xipeng Qiu , Yuxin Wang , Tianfan Fu , Beilun Wang This is my paper
classification cs.LG cs.AI
keywords antibody designCDR-H3on-policy distillationdiffusion modelprotein structure generationpost-trainingprivileged informationstructural recovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the standard training distribution for diffusion-based antibody CDR design—noise added to native structures—mismatches the recursive, self-generated states seen at inference, allowing backbone errors to accumulate in flexible loops like CDR-H3. ABOPD addresses this by distilling a frozen, backbone-aware teacher's coordinate predictions on states visited along the student's own reverse-denoising trajectory, while retaining the original denoising objective as an offline anchor. If correct, this reduces RAbD CDR-H3 RMSD from 2.37 Å to 1.95 Å and improves RMSD across all six CDRs in simultaneous redesign, beating supervised fine-tuning and offline distillation controls that share the same initialization. The practical objective is L_anchor + 0.6 L_OPD_pos, where L_OPD_pos matches student and teacher coordinate-noise predictions on detached rollout states.

Core claim

The paper's central claim is that the state-distribution mismatch in diffusion-based antibody CDR design—training on noise derived from native structures but generating recursively from the model's own intermediates—causes accumulated backbone error, and that supervising the model on its own rollout states fixes it. ABOPD does this by distilling a frozen backbone-aware teacher's Cα coordinate predictions on student-visited states, combined with the original denoising loss as an anchor. The result is a 0.42 Å reduction in RAbD CDR-H3 RMSD (2.37→1.95 Å) and improved RMSD across all six CDRs in simultaneous redesign, beating supervised fine-tuning and offline distillation controls that share th

What carries the argument

The central mechanism is the on-policy coordinate-transition distillation loss L_OPD_pos (Eq. 11): at selected timesteps along the student's own reverse-denoising trajectory, the student's predicted coordinate noise is regressed to the frozen teacher's prediction on the same state, uniformly reweighted, with gradients detached through the rollout. The teacher is a hybrid-pretrained DiffAb model augmented with privileged native backbone descriptors (Cα-centered N/C/Cα/O/Cβ coordinates and pair-distance features) injected through lightweight residual adapters. The final objective is L_ABOPD = L_anchor + 0.6 L_OPD_pos, where L_anchor is the original denoising loss on reference-derived states.

Load-bearing premise

The teacher is trained only on reference-derived states (Eq. 6) but is evaluated on student rollout states during ABOPD, and the paper assumes its fixed coordinate targets remain accurate and are better than direct regression to native coordinates on those same states.

What would settle it

Train ABOPD with the same anchor and rollout states but replace the teacher target in L_OPD_pos with the native-directed transition mean (the posterior target to x0 used in the paper's own diagnostic). If that variant achieves RMSD equal to or below 1.95 Å, the teacher is unnecessary and the gain comes from on-policy native supervision rather than distillation. Alternatively, track the frozen teacher's transition error on student rollout states through post-training; growing error would show the teacher is a stale target.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Existing antibody diffusion generators can be improved by post-training on their own rollout states, without changing architecture or sampling.
  • Coordinate-only distillation indicates backbone Cα geometry is the key bottleneck in CDR loop recovery; sequence and orientation targets add little.
  • The teacher advantage and student improvement concentrate at late denoising timesteps, suggesting timestep-adaptive weighting could yield further gains.
  • The same on-policy distillation recipe could apply to other diffusion-based protein design tasks where recursive generation drifts from the training distribution.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper never compares teacher targets against native-coordinate targets on rollout states; a direct regression to x0 could plausibly match or beat the teacher, which would reframe the contribution as on-policy native supervision rather than distillation.
  • H-DiffAb is pretrained with roughly 8× the budget of DiffAb; although post-training controls share this initialization, part of the absolute gain over DiffAb may reflect scale rather than the ABOPD objective.
  • The fidelity–diversity trade-off (lower IMP and higher JSDsc) suggests that dense trajectory supervision narrows the sampled design space; diversity-aware or all-atom extensions could recover exploration while keeping the backbone gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes ABOPD, a post-training procedure for antigen-conditioned antibody CDR diffusion models. ABOPD first pretrains a single hybrid model H-DiffAb across single-, multi-, and all-six-CDR masks; then adapts a 'backbone-aware teacher' that receives privileged native-side geometry through two small adapters; and finally distills the frozen teacher's C-alpha coordinate-transition predictions on states sampled from the student's own reverse denoising trajectory, while retaining the original denoising loss as an anchor. The reported headline results are a reduction in RAbD CDR-H3 RMSD from 2.37 Å to 1.95 Å (Table 3), improvements across all six CDRs in simultaneous redesign (Table 2), and superiority over SFT and offline-distillation controls (Fig. 4a). The paper includes teacher validation (Table 1), a timestep-resolved mechanism analysis (Fig. 4b-d), a target ablation (Fig. 5), and an EMA analysis (Appendix A.5).

Significance. Structural post-training for diffusion-based protein design is an important and underexplored area, and the reported gains on CDR-H3 are practically meaningful if supported. The paper's strengths are the controlled initialization across all post-training variants, three-seed confidence intervals, public code and model releases, and transparent discussion of side-chain limitations (Table 7). However, the central attribution of the gains to 'on-policy distillation' is not yet established: the method has not been compared against direct regression to the native-directed transition on the same student-visited states, and the headline comparison embeds an EMA asymmetry. These are fixable with additional ablations rather than fundamental design errors.

major comments (3)
  1. [Sec. 3.3 / Eq. (11), Sec. A.3] The teacher target in Eq. (11) is never compared with the native-directed transition target \hat{\epsilon}^* = (x_t - sqrt(\bar{\alpha}_t) x_0)/sqrt(1-\bar{\alpha}_t) on the same rollout states. Because the teacher is conditioned on the privileged descriptor B, an equally plausible explanation is that any native-informed target on student-visited states improves regression, and the teacher is unnecessary. \hat{\epsilon}^* is defined in A.3 but used only for diagnostics (Fig. 4c, Fig. 8). Please add an ablation replacing \hat{\epsilon}_T with \hat{\epsilon}^* in Eq. (11), keeping rollouts, L_anchor, EMA, and beta fixed. If that ablation matches ABOPD's RMSD, the distillation attribution is unsupported; if not, the teacher mechanism is confirmed. Fig. 4(c) does not settle this because it is measured on H-DiffAb rollouts, not on the student's evolving rollout distribution.
  2. [Sec. 3.3, Fig. 4(a), Eq. (15)] The reported ABOPD numbers use the EMA student, whereas SFT and offline distillation use directly optimized parameters; Table 6 shows EMA contributes about 0.045 Å on RAbD (1.9954 vs 1.9500). This does not erase the gain, but the headline comparisons are not parameterization-matched. Please report EMA versions of the controls or use the non-EMA student in the main tables/figures. In addition, the offline-distillation control (Eq. 15) omits L_anchor, so the ABOPD-vs-offline comparison confounds the state distribution (on-policy vs reference-derived) with the presence of the anchor. An 'offline distillation + anchor' control is needed to isolate the on-policy-state effect.
  3. [Sec. 3.3 / Table 8] The main hyperparameters—beta=0.6, rollout timestep set T={80,...,5}, rollout start K=80, EMA decay rho=0.9995, and mask distribution p_hyb—are set without sensitivity analysis. Since Eq. (13) balances anchor and OPD via beta and the OPD signal is concentrated at selected timesteps, the robustness of the reported gains to these choices is unknown. The target ablation (Fig. 5) does not address this. Please provide at least a beta sweep and a reduced-rollout-set sensitivity check.
minor comments (4)
  1. [Sec. 3.3, Eqs. (10)-(11)] Dropping the KL prefactor b_t^2/(2\sigma_t^2) and assigning uniform timestep weights means Eq. (11) is no longer the KL objective of Eq. (8). This is a design choice that should be justified more explicitly, or the reverse-KL derivation softened.
  2. [Table 2] The ddG confidence intervals are extremely wide (e.g., DiffAb 27.727±25.835). Consider reporting a median or a success-rate metric to make interface-energy comparisons interpretable.
  3. [Fig. 3(a)] The t-SNE shows ABOPD versus training CDRs but not H-DiffAb; adding the baseline distribution would better support the claim that ABOPD's sequences are not simply closer to training data.
  4. [Appendix D / Table 7] ABOPD's IMP is below H-DiffAb and its JSDsc is higher. This is acknowledged in the limitations, but the abstract's 'higher-fidelity' wording should be qualified or this caveat should appear in the main text.

Circularity Check

0 steps flagged

No construction-level circularity; the teacher-target vs native-target ablation gap is an attribution issue, not an equation-level reduction.

full rationale

ABOPD's objective (Eq. 13) combines an offline denoising anchor (Eq. 12) with an on-policy coordinate regression term (Eq. 11) that matches the student's coordinate-noise prediction to the frozen teacher's prediction on detached student-visited rollout states. The teacher is trained separately on reference-derived states with privileged native-geometry conditioning B (Eq. 6), so its rollout-state outputs are model predictions rather than quantities defined in terms of the student or the final RMSD metric. Equation (11) does not reduce to Eq. (12) or to direct native-coordinate regression by construction; the native-directed target \hat\epsilon* is introduced only as a diagnostic reference in A.3 and is not used as the OPD target. Thus the derivation chain is not circular. The paper's main weakness is an omitted control comparing teacher targets against direct native-directed regression on the same rollout states; its teacher-advantage analysis (Fig. 4c) measures the teacher against that native-directed target and is therefore partly self-supporting, but this is an ablation/attribution gap rather than a definitional reduction. The only overlapping-author citation is [34] (MolAct), a general molecular-RL reference that is not load-bearing. External RAbD benchmarks and same-initialization controls ground the empirical claim independently. Score 2 reflects the minor non-load-bearing self-citation and the self-supporting teacher-selection diagnostic, not a circular derivation.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The privileged descriptor B is an input representation, not an entity. The training method rests on hand-set hyperparameters (beta, timestep set, EMA) and on the assumption that a teacher trained on reference-derived states provides good targets on student rollouts. The missing direct-native-target control leaves the teacher component's necessity untested.

free parameters (8)
  • beta (OPD weight) = 0.6
    Eq. (13); hand-set, no sensitivity analysis reported.
  • Rollout timestep cache T = {80,70,60,50,40,30,20,10,5}
    Sec. 3.3; selected by hand for distillation, distinct from the 8-step mechanism grid.
  • KL-prefactor / equal timestep weighting = 1 for each t in T
    Eq. (10) to Eq. (11); dropped b_t^2/(2 sigma_t^2) and used uniform weights for implementation simplicity.
  • Rollout start index K = 80
    Rollouts begin at 80 on the original 100-step schedule (Table 8).
  • EMA decay rho = 0.9995
    Applied to the student; all main ABOPD results use the EMA student (Appendix A.5).
  • Target-mask distribution p_hyb = (single,multi,all)=(0.25,0.40,0.35)
    Hybrid pretraining mask distribution reused in teacher and student stages (Table 8).
  • H-DiffAb pretraining budget = 8x default DiffAb examples
    Acknowledged as a compute choice affecting absolute DiffAb comparisons (Sec. 3.2).
  • Teacher descriptor normalization = Calpha-centered, /10 A; pair width 10 A
    Hand-designed descriptor encoding in Appendix B.
axioms (6)
  • standard math DDPM Gaussian forward/reverse transition and the KL identity for shared-covariance Gaussians (Eqs. 8-10).
    Taken from Ho et al. [40] and DiffAb [4]; used to motivate coordinate-transition regression.
  • standard math Categorical and SO(3) diffusion posteriors from DiffAb.
    Used in the anchor loss; standard.
  • ad hoc to paper The reverse-KL OPD direction from DiffusionOPD [30] is appropriate for student-visited diffusion states.
    The paper adopts the reverse-KL convention but then drops the KL prefactor, so this is a motivating rather than derived choice.
  • domain assumption Teacher predictions on detached student rollouts remain useful as the student changes.
    Eq. (6) trains the teacher on reference-derived states only; it is then queried on student rollouts during ABOPD training.
  • ad hoc to paper Sparse cached timestep set with uniform weights is an adequate surrogate for full-trajectory supervision.
    The chosen T and equal weights are not derived from any optimality criterion.
  • domain assumption SAbDab split and RAbD benchmark are representative for evaluating antibody CDR design quality.
    Standard benchmark assumption; test sets are held out by CDR-H3 sequence identity.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ABOPD: Antibody CDR Design via On-Policy Distillation." pith.science (2026). https://pith.science/paper/ZEMM5YSW

@misc{pith2026260718835,
  author       = {Pith},
  title        = {Pith review of: ABOPD: Antibody CDR Design via On-Policy Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZEMM5YSW}},
  note         = {Machine review of arXiv:2607.18835}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Antibodies are essential therapeutic molecules, and their complementarity-determining regions (CDRs) form the primary antigen-recognition interface. Recent protein generative models have demonstrated broad capabilities in biomolecular design, yet post-training strategies for downstream objectives remain limited. Standard denoising training operates on noisy states obtained by perturbing native structures, whereas recursive generation proceeds through model-generated intermediate states. For flexible antibody CDR loops such as CDR-H3, this mismatch can allow backbone deviations to accumulate along the denoising trajectory and compromise antigen-facing loop geometry. We introduce ABOPD, an antibody design framework based on on-policy distillation that leverages privileged native geometry during training to supervise states visited along the model's own denoising trajectories. With this fine-grained structural supervision, ABOPD substantially improves structural recovery on RAbD CDR-H3 generation, reducing RMSD by 0.42 {\AA} (from 2.37 {\AA} to 1.95 {\AA}) and outperforming supervised fine-tuning and offline distillation controls, offering a path to higher-fidelity protein design.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 11 canonical work pages

  1. [1]

    R.J.PantazesandC.D.Maranas. OptCDR:ageneralcomputationalmethodforthedesignofantibodycomplementar- itydeterminingregionsfortargetedepitopebinding.ProteinEngineeringDesignandSelection,23(11):849–858,Septem- ber 2010. ISSN 1741-0134. doi: 10.1093/protein/gzq061. URLhttp://dx.doi.org/10.1093/protein/gzq061

    work page doi:10.1093/protein/gzq061 2010

  2. [2]

    RosettaAntibodyDesign(RAbD):Ageneralframework for computational antibody design.PLOS Computational Biology, 14(4):e1006112, April 2018

    JaredAdolf-Bryfogle,OleksKalyuzhniy,MichaelKubitz,etal. RosettaAntibodyDesign(RAbD):Ageneralframework for computational antibody design.PLOS Computational Biology, 14(4):e1006112, April 2018. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1006112. URLhttp://dx.doi.org/10.1371/journal.pcbi.1006112

    doi:10.1371/journal.pcbi.1006112 2018

  3. [3]

    Constantin Schneider, Matthew I. J. Raybould, and Charlotte M. Deane. SAbDab in the age of biotherapeutics: updates including SAbDab-nano, the nanobody structure tracker.Nucleic Acids Research, 50(D1):D1368–D1372, 2022. doi: 10.1093/nar/gkab1050

    work page doi:10.1093/nar/gkab1050 2022

  4. [4]

    Antigen-specific antibody design andoptimizationwithdiffusion-basedgenerativemodelsforproteinstructures

    Shitong Luo, Yufeng Su, Xingang Peng, Sheng Wang, Jian Peng, and Jianzhu Ma. Antigen-specific antibody design andoptimizationwithdiffusion-basedgenerativemodelsforproteinstructures. InProceedingsofthe36thInternational Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, USA, 2022. Curran Associates Inc. ISBN 9781713871088

    2022

  5. [5]

    End-to-end full-atom antibody design

    Xiangzhe Kong, Wenbing Huang, and Yang Liu. End-to-end full-atom antibody design. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023

    2023

  6. [6]

    GeoAB:towardsrealisticantibodydesignandreliableaffinitymaturation

    HaitaoLin, LirongWu, YufeiHuang, et al. GeoAB:towardsrealisticantibodydesignandreliableaffinitymaturation. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  7. [7]

    AbDiffuser: full-atom generation of in vitro functioning antibodies

    Karolis Martinkus, Jan Ludwiczak, Kyunghyun Cho, et al. AbDiffuser: full-atom generation of in vitro functioning antibodies. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA, 2023. Curran Associates Inc

    2023

  8. [8]

    Greene, et al

    Ali Madani, Ben Krause, Eric R. Greene, et al. Large language models generate functional protein sequences across diverse families.Nature Biotechnology, 41:1099–1106, 2023. doi: 10.1038/s41587-022-01618-2

    doi:10.1038/s41587-022-01618-2 2023

  9. [9]

    Antigen-specific antibody design via direct energy-based preference optimization

    Xiangxin Zhou, Dongyu Xue, Ruizhe Chen, Zaixiang Zheng, Liang Wang, and Quanquan Gu. Antigen-specific antibody design via direct energy-based preference optimization. InProceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY, USA, 2024. ISBN 9798331314385

    2024

  10. [10]

    Input perturbation reduces exposure bias in diffusion models

    Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Input perturbation reduces exposure bias in diffusion models. InProceedings of the 40th International Conference on Machine Learning, volume 202, pages 26245–26265. PMLR, 2023

    2023

  11. [11]

    Shuai, Jeffrey A

    Richard W. Shuai, Jeffrey A. Ruffolo, and Jeffrey J. Gray. IgLM: Infilling language modeling for antibody sequence design.Cell Systems, 14(11):979–989.e4, November 2023. ISSN 2405-4712. doi: 10.1016/j.cels.2023.10.001

    doi:10.1016/j.cels.2023.10.001 2023

  12. [12]

    Fleishman

    Rosalie Lipsh-Sokolik, Dina Listov, and Sarel J. Fleishman. The AbDesign computational pipeline for modular backboneassemblyanddesignofbindersandenzymes.ProteinScience,30(1):151–159,October2020. ISSN1469-896X. doi: 10.1002/pro.3970. URLhttp://dx.doi.org/10.1002/pro.3970

    work page doi:10.1002/pro.3970

  13. [13]

    Conditional antibody design as 3D equivariant graph translation

    Xiangzhe Kong, Wenbing Huang, and Yang Liu. Conditional antibody design as 3D equivariant graph translation. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  14. [14]

    Jaakkola

    Wengong Jin, Jeremy Wohlwend, Regina Barzilay, and Tommi S. Jaakkola. Iterative refinement graph neural network for antibody sequence-structure co-design. InInternational Conference on Learning Representations, 2022

    2022

  15. [15]

    Bennett, Joseph L

    Nathaniel R. Bennett, Joseph L. Watson, Robert J. Ragotte, et al. Atomically accurate de novo design of antibodies with RFdiffusion.Nature, 649(8095):183–193, November 2025. ISSN 1476-4687. doi: 10.1038/s41586-025-09721-5. URLhttp://dx.doi.org/10.1038/s41586-025-09721-5

    doi:10.1038/s41586-025-09721-5 2025

  16. [16]

    IgGM: A generative model for functional antibody and nanobody design

    Rubo Wang, Fandi Wu, Xingyu Gao, Jiaxiang Wu, Peilin Zhao, and Jianhua Yao. IgGM: A generative model for functional antibody and nanobody design. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  17. [17]

    AbFlow: End-to-endparatope-centricantibodydesign by interaction enhanced flow matching

    WendaWang,YangZhang,ZheweiWei,andWenbingHuang. AbFlow: End-to-endparatope-centricantibodydesign by interaction enhanced flow matching. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery 13 and Data Mining V.1, KDD ’26, page 1471–1482, New York, NY, USA, 2026. Association for Computing Machinery. ISBN 9798400722585. doi: 10.1145/3770854...

    arXiv 2026

  18. [18]

    ODesign: A world model for biomolecular interaction design, 2025

    Odin Zhang, Xujun Zhang, Haitao Lin, et al. ODesign: A world model for biomolecular interaction design, 2025

    2025

  19. [19]

    High-affinity protein binder design via flow matching and in silico maturation.bioRxiv, January 2026

    Qilin Yu, Liangyue Guo, Xiayan Qin, et al. High-affinity protein binder design via flow matching and in silico maturation.bioRxiv, January 2026. doi: 10.64898/2026.01.19.700484. URLhttp://dx.doi.org/10.64898/2026. 01.19.700484

    work page doi:10.64898/2026.01.19.700484 2026

  20. [20]

    Pellock, Tamuka M

    Ivan Anishchenko, Samuel J. Pellock, Tamuka M. Chidyausiku, et al. De novo protein design by deep network hallucination.Nature, 600(7889):547–552, December 2021. ISSN 1476-4687. doi: 10.1038/s41586-021-04184-w. URL http://dx.doi.org/10.1038/s41586-021-04184-w

    work page doi:10.1038/s41586-021-04184-w 2021

  21. [21]

    Dauparas, I

    J. Dauparas, I. Anishchenko, N. Bennett, et al. Robust deep learning-based protein sequence design using ProteinMPNN.Science, 378(6615):49–56, October 2022. ISSN 1095-9203. doi: 10.1126/science.add2187. URL http://dx.doi.org/10.1126/science.add2187

    doi:10.1126/science.add2187 2022

  22. [22]

    Watson, David Juergens, Nathaniel R

    Joseph L. Watson, David Juergens, Nathaniel R. Bennett, et al. De novo design of protein structure and function with RFdiffusion.Nature, 620(7976):1089–1100, July 2023. ISSN 1476-4687. doi: 10.1038/s41586-023-06415-8. URL http://dx.doi.org/10.1038/s41586-023-06415-8

    doi:10.1038/s41586-023-06415-8 2023

  23. [23]

    Ingraham, Max Baranov, Zak Costello, et al

    John B. Ingraham, Max Baranov, Zak Costello, et al. Illuminating protein space with a programmable generative model.Nature, 623(7989):1070–1078, November 2023. ISSN 1476-4687. doi: 10.1038/s41586-023-06728-8. URL http://dx.doi.org/10.1038/s41586-023-06728-8

    doi:10.1038/s41586-023-06728-8 2023

  24. [24]

    Trippe, Valentin De Bortoli, et al

    Jason Yim, Brian L. Trippe, Valentin De Bortoli, et al. SE(3) diffusion model with application to protein backbone generation. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023

    2023

  25. [25]

    FastproteinbackbonegenerationwithSE(3)flowmatching

    JasonYim,AndrewCampbell,AndrewY.K.Foong,etal. FastproteinbackbonegenerationwithSE(3)flowmatching. InNeurIPS 2023 Workshop on Machine Learning in Structural Biology, 2023

    2023

  26. [26]

    SE(3)-stochastic flow matching for protein backbone generation

    Joey Bose, Tara Akhound-Sadegh, Guillaume Huguet, et al. SE(3)-stochastic flow matching for protein backbone generation. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  27. [27]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, et al. On-policy distillation of language models: Learning from self-generated mistakes. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  28. [28]

    D-OPSD: On-policy self-distillation for continuously tuning step-distilled diffusion models, 2026

    Dengyang Jiang, Xin Jin, Dongyang Liu, et al. D-OPSD: On-policy self-distillation for continuously tuning step-distilled diffusion models, 2026

    2026

  29. [29]

    Flow-OPD: On-policy distillation for flow matching models, 2026

    Zhen Fang, Wenxuan Huang, Yu Zeng, et al. Flow-OPD: On-policy distillation for flow matching models, 2026

    2026

  30. [30]

    DiffusionOPD:Aunifiedperspectiveofon-policydistillationindiffusion models, 2026

    QuanhaoLi,JunqiuYu,KaixunJiang,etal. DiffusionOPD:Aunifiedperspectiveofon-policydistillationindiffusion models, 2026

    2026

  31. [31]

    Molecular de-novo design through deep reinforcement learning.Journal of Cheminformatics, 9(1):48, 2017

    Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. Molecular de-novo design through deep reinforcement learning.Journal of Cheminformatics, 9(1):48, 2017. doi: 10.1186/s13321-017-0235-x. URL https://doi.org/10.1186/s13321-017-0235-x

    doi:10.1186/s13321-017-0235-x 2017

  32. [32]

    Graph convolutional policy network for goal-directed molecular graph generation

    Jiaxuan You, Bowen Liu, Rex Ying, Vijay Pande, and Jure Leskovec. Graph convolutional policy network for goal-directed molecular graph generation. InAdvances in Neural Information Processing Systems, volume 31, 2018

    2018

  33. [33]

    Zare, and Patrick Riley

    Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N. Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning.Scientific Reports, 9(1):10752, 2019. doi: 10.1038/s41598-019-47148-x. URLhttps: //doi.org/10.1038/s41598-019-47148-x

    doi:10.1038/s41598-019-47148-x 2019

  34. [34]

    MolAct: An agentic RL framework for molecular editing and property optimization, 2025

    Zhuo Yang, Yeyun Chen, Jiaqing Xie, et al. MolAct: An agentic RL framework for molecular editing and property optimization, 2025

    2025

  35. [35]

    ProteinRL: Reinforcement learning with generative protein language models for property-directed sequence design

    Matt Sternke and Joel Karpiak. ProteinRL: Reinforcement learning with generative protein language models for property-directed sequence design. InNeurIPS 2023 Generative AI and Biology (GenBio) Workshop, 2023

    2023

  36. [36]

    Mark Kim, and Armen Mkrtchyan

    Yasha Ektefaie, Olivia Viessmann, Siddharth Narayanan, Drew Dresser, J. Mark Kim, and Armen Mkrtchyan. Reinforcement learning on structure-conditioned categorical diffusion for protein inverse folding, 2024. 14

    2024

  37. [37]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. URLhttps://openreview.net/forum?id=qw8AKxfYbI

    2021

  38. [38]

    Learning using privileged information: similarity control and knowledge transfer.J

    Vladimir Vapnik and Rauf Izmailov. Learning using privileged information: similarity control and knowledge transfer.J. Mach. Learn. Res., 16(1):2023–2049, January 2015. ISSN 1532-4435

    2023

  39. [39]

    Unifying distillation and privileged information

    David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. InInternational Conference on Learning Representations, 2016

    2016

  40. [40]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InProceedings of the 34th InternationalConferenceonNeuralInformationProcessingSystems,NIPS’20,RedHook,NY,USA,2020.CurranAssociates Inc. ISBN 9781713829546

    2020

  41. [41]

    SAbDab: the structural antibody database.Nucleic Acids Research, 42(D1):D1140–D1146, 2014

    James Dunbar, Konrad Krawczyk, Jinwoo Leem, et al. SAbDab: the structural antibody database.Nucleic Acids Research, 42(D1):D1140–D1146, 2014. ISSN 1362-4962. doi: 10.1093/nar/gkt1043

    work page doi:10.1093/nar/gkt1043 2014

  42. [42]

    Full-atompeptidedesignwithgeometriclatentdiffusion

    XiangzheKong,YinjunJia, WenbingHuang,andYangLiu. Full-atompeptidedesignwithgeometriclatentdiffusion. InProceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY, USA, 2024. Curran Associates Inc. ISBN 9798331314385

    2024

  43. [43]

    Proteo-R1: Reasoning foundation models for de novo protein design

    Fang Wu, Weihao Xuan, Heli Qi, et al. Proteo-R1: Reasoning foundation models for de novo protein design. In Proceedings of the 43rd International Conference on Machine Learning, 2026

    2026

  44. [44]

    Lewis, et al

    Andrew Leaver-Fay, Michael Tyka, Steven M. Lewis, et al. ROSETTA3: an object-oriented software suite for the sim- ulation and design of macromolecules.Methods in Enzymology, 487:545–574, 2011. doi: 10.1016/B978-0-12-381270-4. 00019-6

    work page doi:10.1016/b978-0-12-381270-4 2011

  45. [45]

    Alford, Andrew Leaver-Fay, Jeliazko R

    Rebecca F. Alford, Andrew Leaver-Fay, Jeliazko R. Jeliazkov, et al. The Rosetta all-atom energy function for macromolecular modeling and design.Journal of Chemical Theory and Computation, 13(6):3031–3048, 2017. ISSN 1549-9626. doi: 10.1021/acs.jctc.7b00125

    doi:10.1021/acs.jctc.7b00125 2017

  46. [46]

    DockQ: A quality measure for protein-protein docking models.PLOS ONE, 11(8): e0161879, 2016

    Sankar Basu and Björn Wallner. DockQ: A quality measure for protein-protein docking models.PLOS ONE, 11(8): e0161879, 2016. doi: 10.1371/journal.pone.0161879

    work page doi:10.1371/journal.pone.0161879 2016

  47. [47]

    UniMoMo: Unified generative modeling of 3D molecules for de novo binder design

    Xiangzhe Kong, Zishen Zhang, Ziting Zhang, et al. UniMoMo: Unified generative modeling of 3D molecules for de novo binder design. InForty-second International Conference on Machine Learning, 2025

    2025

  48. [48]

    Antibody design using a score-based diffusion model guided by evolutionary,physicalandgeometricconstraints

    Tian Zhu, Milong Ren, and Haicang Zhang. Antibody design using a score-based diffusion model guided by evolutionary,physicalandgeometricconstraints. InProceedingsofthe41stInternationalConferenceonMachineLearning, volume 235, pages 62531–62548, 21–27 Jul 2024

    2024

  49. [49]

    Ruffolo, Lee-Shin Chu, Sayan P

    Jeffrey A. Ruffolo, Lee-Shin Chu, Sayan P. Mahajan, and Jeffrey J. Gray. IgFold: Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies.Nature Communications, 14:2389, 2023. doi: 10.1038/s41467-023-38063-x

    doi:10.1038/s41467-023-38063-x 2023

  50. [50]

    The HDOCK server for integrated protein–protein docking.Nature Protocols, 15:1829–1852, 2020

    Yumeng Yan, Huanyu Tao, Jiahua He, and Sheng-You Huang. The HDOCK server for integrated protein–protein docking.Nature Protocols, 15:1829–1852, 2020. doi: 10.1038/s41596-020-0312-x

    work page doi:10.1038/s41596-020-0312-x 2020

  51. [51]

    Dynamic geometric equivariant network for full-atom antibodydesign.ProceedingsoftheAAAIConferenceonArtificialIntelligence,40(1):444–452,March2026

    Weihong Huang, Feng Yang, Qiang Zhang, and Juan Liu. Dynamic geometric equivariant network for full-atom antibodydesign.ProceedingsoftheAAAIConferenceonArtificialIntelligence,40(1):444–452,March2026. ISSN2159-5399. doi: 10.1609/aaai.v40i1.37007. URLhttp://dx.doi.org/10.1609/aaai.v40i1.37007

    work page doi:10.1609/aaai.v40i1.37007

  52. [52]

    BoltzGen: Toward universal binder design.bioRxiv, November

    Hannes Stark, Felix Faltings, MinGyu Choi, et al. BoltzGen: Toward universal binder design.bioRxiv, November

  53. [53]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. InAdvances in Neural Information Processing Systems, volume 37, 2024

    2024

  54. [54]

    Token perturbation guidance for diffusion models

    Javad Rajabi, Soroush Mehraban, Seyedmorteza Sadat, and Babak Taati. Token perturbation guidance for diffusion models. InAdvances in Neural Information Processing Systems, volume 38, 2025

    2025

  55. [55]

    reference-derived

    Dongxu Zhang, Zhichao Yang, Sepehr Janghorbani, et al. Fast and effective on-policy distillation from reasoning prefixes. InFindings of the Association for Computational Linguistics: ACL 2026, pages 25553–25569. Association for Computational Linguistics, 2026. doi: 10.18653/v1/2026.findings-acl.1276. 15 A Supplementary Experimental Results A.1 Additional ...

    work page doi:10.18653/v1/2026.findings-acl.1276 2026

  56. [2025]

    URLhttp://dx.doi.org/10.1101/2025.11.20.689494

    doi: 10.1101/2025.11.20.689494. URLhttp://dx.doi.org/10.1101/2025.11.20.689494

    doi:10.1101/2025.11.20.689494 2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.