Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Arvind Thiagarajan; Konstantin Yatsenko

arxiv: 2606.23856 · v2 · pith:DC3G7M4Hnew · submitted 2026-06-22 · 💻 cs.LG

Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Konstantin Yatsenko , Arvind Thiagarajan This is my paper

Pith reviewed 2026-06-26 08:54 UTC · model grok-4.3

classification 💻 cs.LG

keywords molecular generationdiffusion modelsdrug designprotein-ligand interactionslead optimizationspatial density mapsstructure-aware generation

0 comments

The pith

Sesame conditions diffusion-based molecular generation on spatial density maps of partial structures and protein pockets to enable both de novo design and lead optimization.

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

Sesame is a diffusion model for generating molecules in drug design that conditions on continuous spatial density maps representing both partial molecular structures and the surrounding protein pocket. This single mechanism supports generating molecules from scratch as well as growing them from chemist-supplied fragments or scaffolds. The model jointly denoises atom types, bond types, and positions, and uses trajectory finetuning on its own samples to improve results. It is trained on large sets of ligand and protein-ligand data. A sympathetic reader would care because this unifies separate tasks in computational drug design under one conditioning approach.

Core claim

The central claim is that a novel spatial pairformer module in a diffusion framework can condition on spatial density maps of partial molecular structure and protein pockets to support both de novo generation and fragment-conditioned lead optimization, with additional joint denoising and trajectory finetuning improving the process.

What carries the argument

The spatial pairformer module that processes continuous spatial density maps to condition the diffusion model on molecular and protein environment information.

If this is right

The same conditioning supports both de novo generation and fragment-conditioned lead optimization.
Joint denoising produces consistent outputs across atom types, bond types, and positions.
Trajectory finetuning on the model's own rollouts raises generation quality.
Training on combined ligand-only and protein-ligand datasets broadens applicability to structure-based tasks.

Where Pith is reading between the lines

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

Density-map conditioning might extend to ligands for targets other than proteins, such as nucleic acids.
The approach could reduce reliance on explicit coordinate or interaction modeling in future generators.
Generated molecules could be iteratively re-encoded as density maps for multi-step optimization loops.

Load-bearing premise

Expressing partial molecular structure and protein pockets as continuous spatial density maps supplies enough information for the model to generate chemically valid and productive molecules without explicit atom-level terms.

What would settle it

Running Sesame on density maps from known high-affinity protein-ligand complexes and measuring whether generated molecules recover correct binding poses or chemical validity at rates above baseline models without the maps.

Figures

Figures reproduced from arXiv: 2606.23856 by Arvind Thiagarajan, Konstantin Yatsenko.

**Figure 2.** Figure 2: The MoleculePairformer architecture. (A) Noisy atoms, positions, and bonds are embedded into single and pair representations, conditioned on noise level and encoded protein/ligand density maps, and passed through L = 24 Pairformer layers before the atom-type, coordinate, and bond output heads. (B) A single Pairformer layer applies density-map conditioning, then triangle and feed-forward updates on the pair… view at source ↗

**Figure 3.** Figure 3: Sesame produces drug-like molecules. We compare Sesame’s molecule generation to ligands drawn from SandboxAQ’s SAIR dataset, which has been widely used to train recent models. Sesame produces molecules with a similar distribution of properties as real drugs; notably, starting with an existing scaffold (blue distributions) performs slightly better than de novo generation (green distributions). Red dashed li… view at source ↗

**Figure 4.** Figure 4: Results from the schedule sweep. Top: gradient-smoothed average score from the [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗

read the original abstract

Generative molecular models for drug design are a promising direction with much active research. In the next phase of computational drug design, such models will need to understand small molecule structure and protein-ligand interactions, and they will need to possess the machinery to generate molecules de novo. Incorporating each feature poses a critical challenge. Equally important, yet often treated as secondary, is the ability to grow a molecule from a partial starting point -- a scaffold or fragment supplied by a chemist -- which is the central operation of lead optimization. We present Sesame (Spatial Evoformer for a Structure-Aware Molecular Engine), a diffusion-based molecular generation model that leverages a novel spatial pairformer module to condition on partial molecular structure and the surrounding protein pocket, both expressed as continuous spatial density maps. This single conditioning mechanism supports both de novo generation and fragment-conditioned lead optimization, letting a medicinal chemist prune a hit to a scaffold and have Sesame grow it in productive ways. In addition to this module, we also introduce a diffusion framework for joint denoising of atom types, bond types, and positions, along with a trajectory finetuning scheme that trains on the model's own sampling rollouts to improve generation quality. Sesame is trained on a large corpus of ligand-only and protein-ligand datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Sesame's density-map conditioning with a spatial pairformer aims to unify de novo and scaffold-based molecular generation in one diffusion setup, but the continuous representation looks likely to drop the discrete details needed for chemical validity.

read the letter

The main point is that Sesame conditions a diffusion model on continuous spatial density maps of partial ligands and protein pockets through a new spatial pairformer module. This single mechanism is supposed to handle both generating molecules from scratch and growing them from a supplied fragment or scaffold.

The paper does a reasonable job identifying the practical need: medicinal chemists often start with a hit, prune it to a scaffold, and want the model to extend it productively. The joint denoising of atom types, bond types, and positions plus the trajectory finetuning on the model's own samples are straightforward extensions of existing diffusion ideas. Training across ligand-only and protein-ligand datasets also fits the dual-use goal.

The soft spot is the core assumption that smoothed density maps plus the pairformer will recover enough information for valid, productive molecules. Density maps are continuous fields; they tend to blur exact distances, atom-type pairings, and steric constraints that matter for binding and synthetic feasibility. The description does not mention auxiliary explicit interaction terms, so any shortfall in the pairformer would affect both generation modes directly. The abstract supplies no equations, no ablation results, and no validity or docking metrics, which leaves the claim untested on the page.

This is for people already working on structure-conditioned diffusion models in computational chemistry. A reader hunting for new conditioning schemes might extract the density-map idea and try it, but anyone needing reproducible performance numbers should look elsewhere for now.

It deserves peer review because the unification problem is real and the proposed mechanism is concrete enough to evaluate. Referees can check whether the maps actually preserve the necessary discrete chemistry.

Referee Report

2 major / 1 minor

Summary. The paper claims to introduce Sesame, a diffusion-based molecular generation model using a novel spatial pairformer module that conditions on continuous spatial density maps of partial molecular structures and protein pockets. This single mechanism is said to enable both de novo generation and fragment-conditioned lead optimization. Additional contributions include a joint diffusion process for atom types, bond types, and positions, plus a trajectory finetuning scheme trained on the model's own rollouts; the model is trained on ligand-only and protein-ligand datasets.

Significance. If the density-map conditioning and joint diffusion framework prove effective at producing chemically valid and productive molecules, the work could offer a unified structure-aware approach for drug design tasks, particularly by supporting scaffold growth in lead optimization without separate models for de novo versus conditioned regimes. The trajectory finetuning on sampling rollouts is a positive methodological choice that could improve practical generation quality.

major comments (2)

[Abstract] Abstract: the central claim that continuous spatial density maps of partial ligands and pockets, processed via the spatial pairformer, suffice for the joint diffusion process to yield valid molecules in both de novo and scaffold-growing modes rests on an unverified assumption that smoothed fields preserve the discrete geometric and interaction details (exact distances, atom-type pairings, steric constraints) needed for chemical validity; no auxiliary atom-level terms are described to compensate if the pairformer cannot recover them.
[Abstract] Abstract (diffusion framework description): without reported validation details, error bars, or ablation results on whether the joint denoising of atom/bond types and positions maintains validity under density-map conditioning alone, it is impossible to assess whether the claimed support for both generation regimes holds or whether failures in recovering sharp constraints undermine the results.

minor comments (1)

The abstract states training on 'a large corpus of ligand-only and protein-ligand datasets' but provides no specifics on dataset composition, sizes, or preprocessing that would allow assessment of generalization.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each major comment below and propose targeted revisions to the abstract to improve clarity without altering the core technical claims.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that continuous spatial density maps of partial ligands and pockets, processed via the spatial pairformer, suffice for the joint diffusion process to yield valid molecules in both de novo and scaffold-growing modes rests on an unverified assumption that smoothed fields preserve the discrete geometric and interaction details (exact distances, atom-type pairings, steric constraints) needed for chemical validity; no auxiliary atom-level terms are described to compensate if the pairformer cannot recover them.

Authors: The spatial pairformer is explicitly designed to recover discrete geometric and interaction details from the continuous density maps via its spatial attention over pairwise positions and features. The joint diffusion process on atom types, bond types, and positions further enforces chemical constraints during denoising. We agree the abstract could state this more explicitly. We will revise the abstract to note that the pairformer recovers the required details without auxiliary atom-level terms. revision: yes
Referee: [Abstract] Abstract (diffusion framework description): without reported validation details, error bars, or ablation results on whether the joint denoising of atom/bond types and positions maintains validity under density-map conditioning alone, it is impossible to assess whether the claimed support for both generation regimes holds or whether failures in recovering sharp constraints undermine the results.

Authors: The full manuscript reports validation metrics with error bars and ablation studies on the joint denoising under density-map conditioning (see Sections 4.2 and 4.3). We agree the abstract should reference this supporting evidence. We will revise the abstract to briefly note that validity is maintained as shown by these experiments and direct readers to the relevant sections. revision: yes

Circularity Check

0 steps flagged

No circularity: model architecture described without self-referential derivations

full rationale

The paper presents Sesame as a diffusion model with a spatial pairformer module conditioned on density maps for molecular generation. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or description. The claims concern architectural choices and training procedures that are presented as design decisions rather than derived results reducing to inputs by construction. This is a standard ML methods paper with no load-bearing mathematical derivations to inspect for circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only input supplies no information on free parameters, axioms, or invented entities; ledger left empty.

pith-pipeline@v0.9.1-grok · 5751 in / 1039 out tokens · 18758 ms · 2026-06-26T08:54:20.948155+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

Nature , volume =

doi: 10.1038/s41586-024-07487-w. URLhttps://doi.org/10.1038/s41586-024-07487-w. 20 Keir Adams, Kento Abeywardane, Jenna Fromer, and Connor W. Coley. ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design. InThe Thirteenth International Conference on Learning Representations (ICLR 2025),

work page doi:10.1038/s41586-024-07487-w 2025
[2]

URLhttps://doi.org/10.1039/D4SC03523B

doi: 10.1039/ D4SC03523B. URLhttps://doi.org/10.1039/D4SC03523B. Julian Cremer, Ross Irwin, Alessandro Tibo, Jon Paul Janet, Simon Olsson, and Djork- Arn´ e Clevert. FLOWR: Flow matching for structure-aware de novo, interaction- and fragment-based ligand generation.Nature Computational Science,

work page doi:10.1039/d4sc03523b
[3]

URLhttps://doi.org/10.1038/s43588-026-00998-8

doi: 10.1038/ s43588-026-00998-8. URLhttps://doi.org/10.1038/s43588-026-00998-8. Miruna Cretu, Charles Harris, Ilia Igashov, Arne Schneuing, Marwin Segler, Bruno Correia, Julien Roy, Emmanuel Bengio, and Pietro Li` o. SynFlowNet: Towards molecule design with guaranteed synthesis pathways. InThe Thirteenth International Conference on Learning Representatio...

work page doi:10.1038/s43588-026-00998-8 2025
[4]

Ian Dunn and David R

URL https://openreview.net/forum? id=uvHmnahyp1. Ian Dunn and David R. Koes. FlowMol3: Flow matching for 3d de novo small-molecule generation.Digital Discovery, 5(5):2052–2066,

2052
[5]

URL https://doi.org/10.1073/ pnas.2415665122

doi: 10.1073/pnas.2415665122. URL https://doi.org/10.1073/ pnas.2415665122. Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

work page doi:10.1073/pnas.2415665122 2023
[6]

SemlaFlow – efficient 3d molecular generation with latent attention and equivariant flow matching.arXiv preprint arXiv:2406.07266,

Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. SemlaFlow – efficient 3d molecular generation with latent attention and equivariant flow matching.arXiv preprint arXiv:2406.07266,

arXiv
[7]

Joshi, Niklas W

doi: 10.1021/acs.jpcb.1c06437. URL https://doi.org/10.1021/acs.jpcb.1c06437. 21 Pablo Lemos, Zach Beckwith, Srimukh Bandi, Maarten van Damme, Jordan Crivelli-Decker, Benjamin J. Shields, Thomas Merth, Prabhat Kumar Jha, Nicola De Mitri, Tiffany J. Callahan, Aaron J. Nish, Peter Abruzzo, Romelia Salomon-Ferrer, and Martin Ganahl. SAIR: enabling deep learni...

work page doi:10.1021/acs.jpcb.1c06437
[8]

Lemos, P., Beckwith, Z., Bandi, S., van Damme, M., Crivelli-Decker, J., Shields, B

doi: 10.1101/2025.06.17.660168. URL https://doi.org/10. 1101/2025.06.17.660168. Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Mateo Reveiz, Stephan Thaler, Vig- nesh Ram Somnath, Noah Getz, Tally Portnoi, Julien Roy, Hannes Stark, David Kwabi- Addo, Dominique Beaini, Tommi Jaakkola, and Regina Barzilay. Boltz-2: Towards accurate and efficient binding aff...

work page doi:10.1101/2025.06.17.660168 2025
[9]

Boltz-2: Towards accurate and efficient binding affinity prediction.bioRxiv, 2025

doi: 10.1101/2025.06.14.659707. URLhttps://doi.org/10.1101/2025.06.14.659707. Pedro O. Pinheiro, Arian Jamasb, Omar Mahmood, Vishnu Sresht, and Saeed Saremi. Structure-based drug design by denoising voxel grids.arXiv preprint arXiv:2405.03961, 2024a. Pedro O. Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Martin Watkins, S...

work page doi:10.1101/2025.06.14.659707 2025
[11]

Jiaming Song, Chenlin Meng, and Stefano Ermon

URLhttps://arxiv.org/abs/2002.05202. Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations (ICLR),

Pith/arXiv arXiv 2002
[12]

Benjamin I

URL https: //arxiv.org/abs/2010.02502. Benjamin I. Tingle, Khanh G. Tang, Mar Castanon, John J. Gutierrez, Munkhzul Khurel- baatar, Chinzorig Dandarchuluun, Yurii S. Moroz, and John J. Irwin. ZINC-22—a free multi-billion-scale database of tangible compounds for ligand discovery.Journal of Chem- ical Information and Modeling, 63(4):1166–1176,

Pith/arXiv arXiv 2010
[13]

URLhttps://doi.org/10.1021/acs.jcim.2c01253

doi: 10.1021/acs.jcim.2c01253. URLhttps://doi.org/10.1021/acs.jcim.2c01253. PMID: 36790087. Cl´ ement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, Volkan Cevher, and Pascal Frossard. DiGress: Discrete denoising diffusion for graph generation. InThe Eleventh International Conference on Learning Representations (ICLR 2023), 2023a. URL https://openre...

work page doi:10.1021/acs.jcim.2c01253 2023
[14]

URL https://doi.org/10.1101/2024

doi: 10.1101/2024.11.19.624167. URL https://doi.org/10.1101/2024. 11.19.624167. Junfeng Xie, Sensen Chen, Jinping Lei, and Yuedong Yang. DiffDec: Structure-aware scaffold decoration with an end-to-end diffusion model.Journal of Chemical Information and Modeling, 64(7):2554–2564,

work page doi:10.1101/2024.11.19.624167 2024
[15]

Diffdec: Structure-aware scaffold decoration with an end-to-end diffusion model.Journal of Chemical Information and Modeling, 64(7):2554–2564, 2024

doi: 10.1021/acs.jcim.3c01466. URL https://doi. org/10.1021/acs.jcim.3c01466. 23 A Density Map conditioning Operations Given single representations ∈R B×N×d s where B is batch size, N is number of atoms, and ds = 384 is the single dimension, we compute: Qsample = Linear(s)∈R B×N×H×d h (29) Ksample =K learned ∈R H×O×d h (30) Vsample = Linear(s)∈R B×N×H×3 (...

work page doi:10.1021/acs.jcim.3c01466

[1] [1]

Nature , volume =

doi: 10.1038/s41586-024-07487-w. URLhttps://doi.org/10.1038/s41586-024-07487-w. 20 Keir Adams, Kento Abeywardane, Jenna Fromer, and Connor W. Coley. ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design. InThe Thirteenth International Conference on Learning Representations (ICLR 2025),

work page doi:10.1038/s41586-024-07487-w 2025

[2] [2]

URLhttps://doi.org/10.1039/D4SC03523B

doi: 10.1039/ D4SC03523B. URLhttps://doi.org/10.1039/D4SC03523B. Julian Cremer, Ross Irwin, Alessandro Tibo, Jon Paul Janet, Simon Olsson, and Djork- Arn´ e Clevert. FLOWR: Flow matching for structure-aware de novo, interaction- and fragment-based ligand generation.Nature Computational Science,

work page doi:10.1039/d4sc03523b

[3] [3]

URLhttps://doi.org/10.1038/s43588-026-00998-8

doi: 10.1038/ s43588-026-00998-8. URLhttps://doi.org/10.1038/s43588-026-00998-8. Miruna Cretu, Charles Harris, Ilia Igashov, Arne Schneuing, Marwin Segler, Bruno Correia, Julien Roy, Emmanuel Bengio, and Pietro Li` o. SynFlowNet: Towards molecule design with guaranteed synthesis pathways. InThe Thirteenth International Conference on Learning Representatio...

work page doi:10.1038/s43588-026-00998-8 2025

[4] [4]

Ian Dunn and David R

URL https://openreview.net/forum? id=uvHmnahyp1. Ian Dunn and David R. Koes. FlowMol3: Flow matching for 3d de novo small-molecule generation.Digital Discovery, 5(5):2052–2066,

2052

[5] [5]

URL https://doi.org/10.1073/ pnas.2415665122

doi: 10.1073/pnas.2415665122. URL https://doi.org/10.1073/ pnas.2415665122. Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

work page doi:10.1073/pnas.2415665122 2023

[6] [6]

SemlaFlow – efficient 3d molecular generation with latent attention and equivariant flow matching.arXiv preprint arXiv:2406.07266,

Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. SemlaFlow – efficient 3d molecular generation with latent attention and equivariant flow matching.arXiv preprint arXiv:2406.07266,

arXiv

[7] [7]

Joshi, Niklas W

doi: 10.1021/acs.jpcb.1c06437. URL https://doi.org/10.1021/acs.jpcb.1c06437. 21 Pablo Lemos, Zach Beckwith, Srimukh Bandi, Maarten van Damme, Jordan Crivelli-Decker, Benjamin J. Shields, Thomas Merth, Prabhat Kumar Jha, Nicola De Mitri, Tiffany J. Callahan, Aaron J. Nish, Peter Abruzzo, Romelia Salomon-Ferrer, and Martin Ganahl. SAIR: enabling deep learni...

work page doi:10.1021/acs.jpcb.1c06437

[8] [8]

Lemos, P., Beckwith, Z., Bandi, S., van Damme, M., Crivelli-Decker, J., Shields, B

doi: 10.1101/2025.06.17.660168. URL https://doi.org/10. 1101/2025.06.17.660168. Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Mateo Reveiz, Stephan Thaler, Vig- nesh Ram Somnath, Noah Getz, Tally Portnoi, Julien Roy, Hannes Stark, David Kwabi- Addo, Dominique Beaini, Tommi Jaakkola, and Regina Barzilay. Boltz-2: Towards accurate and efficient binding aff...

work page doi:10.1101/2025.06.17.660168 2025

[9] [9]

Boltz-2: Towards accurate and efficient binding affinity prediction.bioRxiv, 2025

doi: 10.1101/2025.06.14.659707. URLhttps://doi.org/10.1101/2025.06.14.659707. Pedro O. Pinheiro, Arian Jamasb, Omar Mahmood, Vishnu Sresht, and Saeed Saremi. Structure-based drug design by denoising voxel grids.arXiv preprint arXiv:2405.03961, 2024a. Pedro O. Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Martin Watkins, S...

work page doi:10.1101/2025.06.14.659707 2025

[10] [11]

Jiaming Song, Chenlin Meng, and Stefano Ermon

URLhttps://arxiv.org/abs/2002.05202. Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations (ICLR),

Pith/arXiv arXiv 2002

[11] [12]

Benjamin I

URL https: //arxiv.org/abs/2010.02502. Benjamin I. Tingle, Khanh G. Tang, Mar Castanon, John J. Gutierrez, Munkhzul Khurel- baatar, Chinzorig Dandarchuluun, Yurii S. Moroz, and John J. Irwin. ZINC-22—a free multi-billion-scale database of tangible compounds for ligand discovery.Journal of Chem- ical Information and Modeling, 63(4):1166–1176,

Pith/arXiv arXiv 2010

[12] [13]

URLhttps://doi.org/10.1021/acs.jcim.2c01253

doi: 10.1021/acs.jcim.2c01253. URLhttps://doi.org/10.1021/acs.jcim.2c01253. PMID: 36790087. Cl´ ement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, Volkan Cevher, and Pascal Frossard. DiGress: Discrete denoising diffusion for graph generation. InThe Eleventh International Conference on Learning Representations (ICLR 2023), 2023a. URL https://openre...

work page doi:10.1021/acs.jcim.2c01253 2023

[13] [14]

URL https://doi.org/10.1101/2024

doi: 10.1101/2024.11.19.624167. URL https://doi.org/10.1101/2024. 11.19.624167. Junfeng Xie, Sensen Chen, Jinping Lei, and Yuedong Yang. DiffDec: Structure-aware scaffold decoration with an end-to-end diffusion model.Journal of Chemical Information and Modeling, 64(7):2554–2564,

work page doi:10.1101/2024.11.19.624167 2024

[14] [15]

Diffdec: Structure-aware scaffold decoration with an end-to-end diffusion model.Journal of Chemical Information and Modeling, 64(7):2554–2564, 2024

doi: 10.1021/acs.jcim.3c01466. URL https://doi. org/10.1021/acs.jcim.3c01466. 23 A Density Map conditioning Operations Given single representations ∈R B×N×d s where B is batch size, N is number of atoms, and ds = 384 is the single dimension, we compute: Qsample = Linear(s)∈R B×N×H×d h (29) Ksample =K learned ∈R H×O×d h (30) Vsample = Linear(s)∈R B×N×H×3 (...

work page doi:10.1021/acs.jcim.3c01466