pith. sign in

arxiv: 2602.04883 · v2 · pith:OOZ62HB3new · submitted 2026-02-04 · 💻 cs.LG · cs.AI· q-bio.BM· q-bio.QM

Protein Autoregressive Modeling via Multiscale Structure Generation

Yanru Qu , Cheng-Yen Hsieh , Zaixiang Zheng , Ge Liu , Quanquan Gu This is my paper

Pith reviewed 2026-05-21 13:16 UTC · model grok-4.3

classification 💻 cs.LG cs.AIq-bio.BMq-bio.QM
keywords protein structure generationautoregressive modelingmulti-scale generationbackbone designconditional generationmotif scaffoldingtransformer modelflow-based decoder
0
0 comments X

The pith

A multi-scale autoregressive model generates protein backbones by predicting from coarse topology to fine details.

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

The paper presents PAR as a new way to generate protein structures autoregressively across multiple scales. It starts with a coarse representation of the protein and refines it step by step using a transformer that learns to predict the next finer scale. This method allows the model to handle conditional tasks like motif scaffolding without any additional training, while also performing well on generating diverse and high-quality structures from scratch. A sympathetic reader would care because it offers a flexible framework that could speed up protein design by mimicking how one might build a structure gradually.

Core claim

PAR is the first multi-scale autoregressive framework for protein backbone generation via coarse-to-fine next-scale prediction. The framework consists of multi-scale downsampling to represent structures at different scales, an autoregressive transformer that encodes multi-scale information and produces conditional embeddings, and a flow-based backbone decoder that generates the atoms conditioned on those embeddings. The model uses noisy context learning and scheduled sampling to mitigate exposure bias. It demonstrates strong zero-shot generalization for conditional generation and motif scaffolding without fine-tuning, high design quality on unconditional benchmarks, and favorable scaling.

What carries the argument

The autoregressive transformer that encodes multi-scale information from downsampled protein structures and produces conditional embeddings to guide the flow-based backbone decoder in generating structures scale by scale.

If this is right

  • Supports flexible human-prompted conditional generation without fine-tuning
  • Performs motif scaffolding directly in zero-shot manner
  • Generates high design quality backbones on unconditional tasks
  • Shows favorable scaling behavior as model capacity increases

Where Pith is reading between the lines

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

  • This coarse-to-fine approach could be applied to generating other complex hierarchical structures beyond proteins.
  • Interactive design tools might allow users to prompt at different scales for more control over protein engineering.
  • Combining this with experimental validation could test if the generated structures fold as predicted.

Load-bearing premise

The hierarchical nature of proteins can be captured by multi-scale downsampling operations that preserve enough structural information for the autoregressive model to learn across scales.

What would settle it

Observing that generated backbones do not achieve high design quality scores or that zero-shot motif scaffolding fails to produce valid structures on standard benchmarks would falsify the claim of strong performance and generalization.

read the original abstract

We present protein autoregressive modeling (PAR), the first multi-scale autoregressive framework for protein backbone generation via coarse-to-fine next-scale prediction. Using the hierarchical nature of proteins, PAR generates structures that mimic sculpting a statue, forming a coarse topology and refining structural details over scales. To achieve this, PAR consists of three key components: (i) multi-scale downsampling operations that represent protein structures across multiple scales during training; (ii) an autoregressive transformer that encodes multi-scale information and produces conditional embeddings to guide structure generation; (iii) a flow-based backbone decoder that generates backbone atoms conditioned on these embeddings. Moreover, autoregressive models suffer from exposure bias, caused by the training and the generation procedure mismatch, and substantially degrades structure generation quality. We effectively alleviate this issue by adopting noisy context learning and scheduled sampling, enabling robust backbone generation. Notably, PAR exhibits strong zero-shot generalization, supporting flexible human-prompted conditional generation and motif scaffolding without requiring fine-tuning. On the unconditional generation benchmark, PAR effectively learns protein distributions and produces backbones of high design quality, and exhibits favorable scaling behavior. Together, these properties establish PAR as a promising framework for protein structure generation.

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.

Referee Report

2 major / 2 minor

Summary. The paper presents Protein Autoregressive Modeling (PAR), the first multi-scale autoregressive framework for protein backbone generation using coarse-to-fine next-scale prediction. It consists of multi-scale downsampling operations to represent structures across scales, an autoregressive transformer that encodes multi-scale information and produces conditional embeddings, and a flow-based backbone decoder that generates atoms conditioned on those embeddings. Noisy context learning and scheduled sampling are used to mitigate exposure bias. The authors claim strong zero-shot generalization to conditional generation and motif scaffolding without fine-tuning, high design quality on unconditional benchmarks, and favorable scaling behavior.

Significance. If the empirical claims are substantiated with rigorous controls, this would be a meaningful contribution to protein structure generation. The multi-scale autoregressive formulation that explicitly exploits protein hierarchy, combined with zero-shot generalization to motif scaffolding and conditional tasks, addresses a practical need in design workflows. The incorporation of flow-based decoding and exposure-bias mitigation techniques is technically sound and could influence subsequent autoregressive models in structural biology.

major comments (2)
  1. [§3.2] §3.2 (Multi-scale downsampling): The central modeling assumption—that the chosen downsampling operations preserve sufficient geometric and topological information for the autoregressive transformer to learn useful cross-scale conditionals—is load-bearing, yet the manuscript provides no quantitative verification (e.g., per-scale reconstruction RMSD, secondary-structure retention rates, or mutual information between coarse and fine representations). Without such diagnostics, it remains unclear whether the learned p(structure_{k+1} | structure_k) actually captures the hierarchical statistics of proteins or simply fits under-constrained distributions.
  2. [§5.1 and Table 4] §5.1 and Table 4 (zero-shot motif scaffolding): The reported success rates for motif scaffolding are presented without error bars across multiple random seeds or explicit comparison to fine-tuned baselines of comparable capacity. Because the zero-shot claim is a primary selling point, the absence of these controls makes it difficult to judge whether the multi-scale architecture itself, rather than dataset scale or decoder choice, drives the observed generalization.
minor comments (2)
  1. [Abstract] The abstract states performance claims without citing the specific metrics, tables, or figures that support them; adding one-sentence references to the relevant results would improve readability.
  2. [Methods] Notation for scale indices and conditional distributions is introduced inconsistently between the methods equations and the results text; a single consolidated notation table would reduce ambiguity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive feedback on our manuscript. We address each of the major comments below, providing clarifications and outlining the revisions we will make to improve the paper.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Multi-scale downsampling): The central modeling assumption—that the chosen downsampling operations preserve sufficient geometric and topological information for the autoregressive transformer to learn useful cross-scale conditionals—is load-bearing, yet the manuscript provides no quantitative verification (e.g., per-scale reconstruction RMSD, secondary-structure retention rates, or mutual information between coarse and fine representations). Without such diagnostics, it remains unclear whether the learned p(structure_{k+1} | structure_k) actually captures the hierarchical statistics of proteins or simply fits under-constrained distributions.

    Authors: We agree that additional quantitative diagnostics would strengthen the support for our central modeling assumption. Although the end-to-end results demonstrate the utility of the multi-scale approach, we will incorporate per-scale reconstruction RMSD and secondary-structure retention rates in the revised manuscript to verify that the downsampling operations preserve sufficient geometric and topological information. This will help confirm that the autoregressive transformer learns meaningful cross-scale conditionals. revision: yes

  2. Referee: [§5.1 and Table 4] §5.1 and Table 4 (zero-shot motif scaffolding): The reported success rates for motif scaffolding are presented without error bars across multiple random seeds or explicit comparison to fine-tuned baselines of comparable capacity. Because the zero-shot claim is a primary selling point, the absence of these controls makes it difficult to judge whether the multi-scale architecture itself, rather than dataset scale or decoder choice, drives the observed generalization.

    Authors: We acknowledge the importance of these controls for substantiating the zero-shot generalization claims. In the revision, we will report error bars for the success rates across multiple random seeds in Table 4 and the associated text. Furthermore, we will add explicit comparisons to fine-tuned baselines of comparable capacity to better isolate the contribution of the multi-scale autoregressive architecture versus other factors such as dataset scale or decoder design. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is trained on external data with independent empirical claims

full rationale

The paper presents PAR as a trained multi-scale autoregressive model using downsampling, transformer, and flow decoder components on protein structure data. No equations, predictions, or uniqueness theorems are shown reducing claimed performance to fitted inputs or self-citations by construction. Claims rest on external benchmarks and zero-shot generalization, making the derivation self-contained against the provided abstract and context.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions from machine learning and structural biology rather than new invented entities or heavily fitted parameters.

axioms (1)
  • domain assumption Protein structures possess a hierarchical organization that can be represented meaningfully at multiple resolution scales.
    Invoked to justify the multi-scale downsampling and coarse-to-fine prediction strategy.

pith-pipeline@v0.9.0 · 5757 in / 1244 out tokens · 56804 ms · 2026-05-21T13:16:40.825649+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 7 internal anchors

  1. [1]

    GPT-4 Technical Report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    Why exposure bias matters: An imitation learning perspective of error accumulation in language generation.arXiv preprint arXiv:2204.01171, 2022

    Kushal Arora, Layla El Asri, Hareesh Bahuleyan, and Jackie Chi Kit Cheung. Why exposure bias matters: An imitation learning perspective of error accumulation in language generation.arXiv preprint arXiv:2204.01171, 2022

    work page arXiv 2022

  4. [4]

    Scheduled sampling for sequence prediction with recurrent neural networks.Advancesin neural information processing systems, 28, 2015

    Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks.Advancesin neural information processing systems, 28, 2015

    work page 2015

  5. [5]

    Se (3)-stochastic flow matching for protein backbone generation.arXiv preprint arXiv:2310.02391, 2023

    Avishek Joey Bose, Tara Akhound-Sadegh, Guillaume Huguet, Kilian Fatras, Jarrid Rector-Brooks, Cheng-Hao Liu, Andrei Cristian Nica, Maksym Korablyov, Michael Bronstein, and Alexander Tong. Se (3)-stochastic flow matching for protein backbone generation.arXiv preprint arXiv:2310.02391, 2023

    work page arXiv 2023

  6. [6]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advancesin neural information processing systems, 33:1877–1901, 2020

    work page 1901

  7. [7]

    Gen- erative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design,

    Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design.arXiv preprint arXiv:2402.04997, 2024

    work page arXiv 2024

  8. [8]

    arXiv preprint arXiv:2504.07963 (2025)

    Shoufa Chen, Chongjian Ge, Shilong Zhang, Peize Sun, and Ping Luo. Pixelflow: Pixel-space generative models with flow.arXiv preprint arXiv:2504.07963, 2025

    work page arXiv 2025

  9. [9]

    Analog bits: Generating discrete data using diffusion models with self-conditioning

    Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022

    work page arXiv 2022

  10. [10]

    An all-atom protein generative model.Proceedings of the National Academy of Sciences, 121(27):e2311500121, 2024

    Alexander E Chu, Jinho Kim, Lucy Cheng, Gina El Nesr, Minkai Xu, Richard W Shuai, and Po-Ssu Huang. An all-atom protein generative model.Proceedings of the National Academy of Sciences, 121(27):e2311500121, 2024

    work page 2024

  11. [11]

    Robust deep learning–based protein sequence design using proteinmpnn

    Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas, Neville Bethel, et al. Robust deep learning–based protein sequence design using proteinmpnn. Science, 378(6615):49–56, 2022

    work page 2022

  12. [12]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021

    work page 2021

  13. [13]

    Learning the language of protein structure.arXiv preprint arXiv:2405.15840, 2024

    Benoit Gaujac, Jérémie Donà, Liviu Copoiu, Timothy Atkinson, Thomas Pierrot, and Thomas D Barrett. Learning the language of protein structure.arXiv preprint arXiv:2405.15840, 2024

    work page arXiv 2024

  14. [14]

    Proteina: Scaling flow-based protein structure generative models

    Tomas Geffner, Kieran Didi, Zuobai Zhang, Danny Reidenbach, Zhonglin Cao, Jason Yim, Mario Geiger, Christian Dallago, Emine Kucukbenli, Arash Vahdat, et al. Proteina: Scaling flow-based protein structure generative models. arXiv preprint arXiv:2503.00710, 2025

    work page arXiv 2025

  15. [15]

    Simulating 500 million years of evolution with a language model

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, et al. Simulating 500 million years of evolution with a language model. Science, 387(6736):850–858, 2025

    work page 2025

  16. [16]

    Exposure bias versus self-recovery: Are distortions really incremental for autoregressive text generation?arXiv preprint arXiv:1905.10617, 2019

    Tianxing He, Jingzhao Zhang, Zhiming Zhou, and James Glass. Exposure bias versus self-recovery: Are distortions really incremental for autoregressive text generation?arXiv preprint arXiv:1905.10617, 2019

    work page arXiv 1905

  17. [17]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advancesin neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advancesin neural information processing systems, 30, 2017

    work page 2017

  18. [18]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 13

    work page 2020

  19. [19]

    Elucidating the design space of multimodal protein language models.arXiv preprint arXiv:2504.11454, 2025

    Cheng-Yen Hsieh, Xinyou Wang, Daiheng Zhang, Dongyu Xue, Fei Ye, Shujian Huang, Zaixiang Zheng, and Quan- quan Gu. Elucidating the design space of multimodal protein language models.arXiv preprint arXiv:2504.11454, 2025

    work page arXiv 2025

  20. [20]

    Riemannian diffusion models

    Chin-Wei Huang, Milad Aghajohari, Joey Bose, Prakash Panangaden, and Aaron C Courville. Riemannian diffusion models. Advancesin Neural Information Processing Systems, 35:2750–2761, 2022

    work page 2022

  21. [21]

    The coming of age of de novo protein design.Nature, 537 (7620):320–327, 2016

    Po-Ssu Huang, Scott E Boyken, and David Baker. The coming of age of de novo protein design.Nature, 537 (7620):320–327, 2016

    work page 2016

  22. [22]

    Illuminatingprotein space with a programmable generative model

    John B Ingraham, Max Baranov, Zak Costello, Karl W Barber, Wujie Wang, Ahmed Ismail, Vincent Frappier, Dana M Lord, Christopher Ng-Thow-Hing, Erik RVan Vlack, et al. Illuminatingprotein space with a programmable generative model. Nature, 623(7989):1070–1078, 2023

    work page 2023

  23. [23]

    Highly accurate protein structure prediction with alphafold

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589, 2021

    work page 2021

  24. [24]

    Scaling Laws for Neural Language Models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprintarXiv:2001.08361, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2001

  25. [25]

    Advances in protein structure prediction and design.Nature reviews molecular cell biology, 20(11):681–697, 2019

    Brian Kuhlman and Philip Bradley. Advances in protein structure prediction and design.Nature reviews molecular cell biology, 20(11):681–697, 2019

    work page 2019

  26. [26]

    Biotite: a unifying open source computational biology framework in python

    Patrick Kunzmann and Kay Hamacher. Biotite: a unifying open source computational biology framework in python. BMC bioinformatics, 19(1):346, 2018

    work page 2018

  27. [27]

    P-sea: a new efficient assignment of secondary structure from cαtrace of proteins

    Gilles Labesse, N Colloc’h, Joël Pothier, and J-P Mornon. P-sea: a new efficient assignment of secondary structure from cαtrace of proteins. Bioinformatics, 13(3):291–295, 1997

    work page 1997

  28. [28]

    Onecat: Decoder-only auto-regressive model for unified understanding and generation.arXiv preprint arXiv:2509.03498, 2025

    Han Li, Xinyu Peng, Yaoming Wang, Zelin Peng, Xin Chen, Rongxiang Weng, Jingang Wang, Xunliang Cai, Wenrui Dai, and Hongkai Xiong. Onecat: Decoder-only auto-regressive model for unified understanding and generation. arXiv preprint arXiv:2509.03498, 2025

    work page arXiv 2025

  29. [29]

    Back to Basics: Let Denoising Generative Models Denoise

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise. arXiv preprint arXiv:2511.13720, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Autoregressive image generation without vector quantization.Advancesin Neural Information Processing Systems, 37:56424–56445, 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization.Advancesin Neural Information Processing Systems, 37:56424–56445, 2024

    work page 2024

  31. [31]

    Generating novel, designable, and diverse protein structures by equivari- antly diffusing oriented residue clouds.arXiv preprint arXiv:2301.12485, 2023

    Yeqing Lin and Mohammed AlQuraishi. Generating novel, designable, and diverse protein structures by equivari- antly diffusing oriented residue clouds.arXiv preprint arXiv:2301.12485, 2023

    work page arXiv 2023

  32. [32]

    Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2.arXiv preprint arXiv:2405.15489, 2024

    Yeqing Lin, Minji Lee, Zhao Zhang, and Mohammed AlQuraishi. Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2.arXiv preprint arXiv:2405.15489, 2024

    work page arXiv 2024

  33. [33]

    Evolutionary-scale prediction of atomic-level protein structure with a language model

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1130, 2023

    work page 2023

  34. [34]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  35. [35]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, pages 23–40. Springer, 2024

    work page 2024

  36. [36]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    work page 2023

  37. [37]

    P (all-atom) is unlocking new path for protein design

    Wei Qu, Jiawei Guan, Rui Ma, Ke Zhai, Weikun Wu, and Haobo Wang. P (all-atom) is unlocking new path for protein design. bioRxiv, pages 2024–08, 2024

    work page 2024

  38. [38]

    Beyond next-token: Next-x prediction for autoregressive visual generation.arXiv preprint arXiv:2502.20388, 2025

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Beyond next-token: Next-x prediction for autoregressive visual generation.arXiv preprint arXiv:2502.20388, 2025. 14

    work page arXiv 2025

  39. [39]

    Improved techniques for training gans.Advancesin neural information processing systems, 29, 2016

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans.Advancesin neural information processing systems, 29, 2016

    work page 2016

  40. [40]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024

    work page 2024

  41. [41]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    Attention is all you need.Advancesin neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advancesin neural information processing systems, 30, 2017

    work page 2017

  43. [43]

    arXiv preprint arXiv:2410.13782 , year=

    Xinyou Wang, Zaixiang Zheng, Fei Ye, Dongyu Xue, Shujian Huang, and Quanquan Gu. Dplm-2: A multimodal diffusion protein language model.arXiv preprint arXiv:2410.13782, 2024

    work page arXiv 2024

  44. [44]

    Zero-shot image restora- tion using denoising diffusion null-space model.arXiv preprint arXiv:2212.00490,

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022

    work page arXiv 2022

  45. [45]

    De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach, Woody Ahern, Andrew J Borst, Robert J Ragotte, Lukas F Milles, et al. De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    work page 2023

  46. [46]

    A learning algorithm for continually running fully recurrent neural networks

    Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989

    work page 1989

  47. [47]

    Fast protein backbone generation with se (3) flow matching,

    Jason Yim, Andrew Campbell, Andrew YK Foong, Michael Gastegger, José Jiménez-Luna, Sarah Lewis, Vic- tor Garcia Satorras, Bastiaan S Veeling, Regina Barzilay, Tommi Jaakkola, et al. Fast protein backbone generation with se (3) flow matching.arXiv preprint arXiv:2310.05297, 2023

    work page arXiv 2023

  48. [48]

    Se (3) diffusion model with application to protein backbone generation

    Jason Yim, Brian L Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. Se (3) diffusion model with application to protein backbone generation. InInternational Conference on Machine Learning, pages 40001–40039. PMLR, 2023

    work page 2023

  49. [49]

    Diffusion Transformers with Representation Autoencoders

    Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representation autoencoders. arXiv preprint arXiv:2510.11690, 2025. 15 Appendix A Implementation and Evaluation Details WefollowtheimplementationofProteina[ 14]fortrainingPAR,usingthesamearchitectureandhyperparameter setup. Training is conducted on 8 H100 GPUs, with a bat...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  50. [50]

    Downsample the coordinate sequence fromRL×3 toR size(i)×3 for each scale i

    work page

  51. [51]

    Spatial relationships in 3D space after downsampling.We quantify this using the pairwise distance map calculated from the full-resolution structure:

    We compute pairwise distance maps using the downsampled sequence, leading to asize(i) ×size (i) map. Spatial relationships in 3D space after downsampling.We quantify this using the pairwise distance map calculated from the full-resolution structure:

    work page

  52. [52]

    Calculate the pairwise distance map of the structure, producing aL×Lmap

    work page

  53. [53]

    We downsample pairwise map this using theF.interpolate(mode=’bicubic’) operation, resulting in asize(i)×size(i)map. Does sequence-based downsampling preserve spatial relationships? We select all samples from the testing set, and calculate the RMSE and LDDT between the aforementioned two size(i) ×size (i)pairwise maps for each sample. As expected, rmse sli...

    work page