pith. sign in

arxiv: 2606.06356 · v2 · pith:U4I3SG5Qnew · submitted 2026-06-04 · 💻 cs.AI

Where Should Knowledge Enter? A Layered Framework for Knowledge Infusion in Multimodal Iterative Generative Model

Renjith Prasad , Chathurangi Shyalika , Anushka Pawar , Aahan Rathod , Amit Sheth This is my paper

Pith reviewed 2026-06-28 01:26 UTC · model grok-4.3

classification 💻 cs.AI
keywords knowledge infusionlayered frameworkmultimodal generative modelsdiffusion modelssafety alignmentknowledge graphsiterative generation
0
0 comments X

The pith

Knowledge enters generative models most effectively when infused at four distinct layers of the iterative process rather than through isolated techniques.

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

The paper claims that knowledge infusion into multimodal iterative generators is best understood as an intervention-layer problem rather than a choice among techniques like prompting or fine-tuning. It identifies four structurally distinct layers where knowledge can act: surface interventions at the input/output boundary, trajectory interventions on the transition function, latent interventions on intermediate states, and parametric interventions on the model parameters. In a controlled experiment with diffusion models and a multimodal knowledge graph for safety alignment, the authors implement three layers cumulatively and show that each new layer reaches failure classes missed by the previous ones, yielding a 70.97 percent drop in knowledge-violating outputs. A sympathetic reader would care because the framework supplies design principles for composing methods that current ad-hoc approaches lack, making reliable knowledge adherence more systematic in safety-critical or domain-specific generation.

Core claim

We argue that knowledge infusion in iterative generative models is fundamentally an intervention-layer problem. Since the generative process unfolds as a trajectory of internal states, knowledge can act on four structurally distinct components of this process: the input/output boundary, the transition function, the intermediate state, and the model parameters. This maps to four intervention layers: surface, trajectory, latent, and parametric infusion. We instantiate the framework in diffusion models, map representative methods to all four layers, and derive design principles for multi-layer composition.

What carries the argument

The four intervention layers (surface, trajectory, latent, parametric) that map to distinct components of the generative trajectory and allow complementary knowledge interventions.

If this is right

  • Each additional layer addresses failure classes unreachable by prior layers alone.
  • Cumulative use of three layers produces a 70.97 percent reduction in knowledge-violating outputs relative to vanilla generation.
  • Representative existing methods can be mapped onto each of the four layers.
  • Design principles for composing multiple layers can be derived from the structural distinctions among the components they modify.

Where Pith is reading between the lines

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

  • Testing the parametric layer in the same safety-alignment setup could reveal whether further reductions beyond 70.97 percent are attainable.
  • The framework offers a diagnostic lens for why single-technique knowledge infusion sometimes fails in practice.
  • The layering idea may extend to other iterative generators such as autoregressive models without requiring diffusion-specific mechanics.

Load-bearing premise

The concrete implementations chosen for surface, trajectory, and latent interventions in the experiment map directly onto the abstract layers without substantial overlap or other factors that could produce the observed complementarity.

What would settle it

An experiment that reassigns the same intervention methods to different layers or introduces controlled overlap and finds that the reduction in knowledge-violating outputs no longer increases reliably with each added layer.

Figures

Figures reproduced from arXiv: 2606.06356 by Aahan Rathod, Amit Sheth, Anushka Pawar, Chathurangi Shyalika, Renjith Prasad.

Figure 1
Figure 1. Figure 1: Four intervention layers for knowledge infusion in iterative generative models. External knowledge acts on four structurally dis [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ontology-guided knowledge infusion for process simulation in a rocket assembly pipeline. The process ontology serves as [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-layer knowledge infusion for safety alignment in text-to-image diffusion using a multimodal knowledge graph (MMKG). [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Multimodal generative models produce fluent outputs but remain unreliable when generation must respect structured, domain-specific, or safety-critical knowledge. Existing methods incorporate knowledge through mechanisms such as prompt augmentation, guidance, latent editing, or fine-tuning, yet they are typically categorized by technique rather than by the component of the generative process they modify. We argue that knowledge infusion in iterative generative models is fundamentally anintervention-layer problem. Since thegenerative process unfolds as a trajectory of internal states, knowledge can act on four structurally distinct components of this process: the input/output boundary, the transition function, the intermediate state, and the model parameters. This maps to four intervention layers: surface, trajectory, latent, and parametric infusion. We instantiate the framework in diffusion models, map representative methods to all four layers, and derive design principles for multi-layer composition. In a controlled safety-alignment experiment using a multimodal knowledge graph with two diffusion backbones, we implement three of the four layers cumulatively, surface (input-side and output-side) and trajectory--latent (mid-generation). We show empirically that each additional layer addresses failure classes that prior layers cannot reach, reducing knowledge-violating outputs by 70.97% compared to vanilla generation and empirically confirming the framework's complementarity prediction.

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 proposes a four-layer framework (surface, trajectory, latent, parametric) for knowledge infusion in iterative multimodal generative models, arguing that interventions act on distinct components of the generative trajectory. It maps existing methods to the layers, derives composition principles, and reports a controlled safety-alignment experiment on diffusion models using a multimodal knowledge graph and two backbones. In the experiment, cumulative application of surface (input/output) and combined trajectory-latent (mid-generation) interventions reduces knowledge-violating outputs by 70.97% relative to vanilla generation, with each added layer claimed to address failure classes unreachable by prior layers.

Significance. If the empirical complementarity result holds after addressing controls for intervention strength and failure-class overlap, the framework supplies a useful organizing taxonomy that could improve systematic design of knowledge-infused systems in safety-critical settings. The use of multiple diffusion backbones and a structured knowledge graph strengthens the experiment's scope relative to single-model studies.

major comments (2)
  1. [Experiment description] Abstract and experiment description: the claim that 'each additional layer addresses failure classes that prior layers cannot reach' is load-bearing for the complementarity prediction, yet the manuscript reports only the cumulative 70.97% reduction without a per-layer breakdown of addressed failure classes (e.g., via table or figure showing incremental error types). This omission prevents verification that observed gains arise from structural layer distinctions rather than additive intervention effects.
  2. [Methods] Methods (trajectory-latent bundling): combining trajectory and latent interventions into a single mid-generation step creates potential functional overlap, since surface-level prompt changes can already alter denoising trajectories in diffusion models; without strength-matched ablations or separate controls for compute/intervention dosage, the design does not isolate whether incremental gains confirm the four-layer structure or simply reflect more total interventions.
minor comments (2)
  1. [Abstract] Abstract contains a typographical error ('Since thegenerative process') that should be corrected for clarity.
  2. [Framework section] Notation for the four layers is introduced conceptually but would benefit from an explicit summary table mapping representative methods to each layer to aid reader navigation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key gaps in the experimental reporting and controls. We agree that both points require additional analysis and will revise the manuscript accordingly to strengthen the evidence for the framework.

read point-by-point responses

  1. Referee: [Experiment description] Abstract and experiment description: the claim that 'each additional layer addresses failure classes that prior layers cannot reach' is load-bearing for the complementarity prediction, yet the manuscript reports only the cumulative 70.97% reduction without a per-layer breakdown of addressed failure classes (e.g., via table or figure showing incremental error types). This omission prevents verification that observed gains arise from structural layer distinctions rather than additive intervention effects.

    Authors: We agree that the manuscript lacks the per-layer breakdown needed to verify the complementarity claim. The current results report only the cumulative reduction. In revision we will add a table (or figure) that enumerates the distinct failure classes reduced at each cumulative stage, allowing direct inspection of whether each added layer targets error types unreachable by prior layers. revision: yes

  2. Referee: [Methods] Methods (trajectory-latent bundling): combining trajectory and latent interventions into a single mid-generation step creates potential functional overlap, since surface-level prompt changes can already alter denoising trajectories in diffusion models; without strength-matched ablations or separate controls for compute/intervention dosage, the design does not isolate whether incremental gains confirm the four-layer structure or simply reflect more total interventions.

    Authors: The referee is correct that bundling trajectory and latent interventions risks confounding additive dosage with layer-specific effects. The manuscript does not include strength-matched ablations or dosage controls. We will add separate ablations that apply trajectory and latent interventions independently (where feasible) together with matched intervention-strength and compute controls to better isolate the contribution of each layer. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is a structural categorization of interventions; experiment tests cumulative effects without reducing claims to fitted inputs or self-citations.

full rationale

The paper defines four intervention layers by the component of the generative trajectory they target (input/output boundary, transition function, intermediate state, model parameters). It maps existing methods to these layers, derives composition principles, and reports an experiment implementing three layers cumulatively on diffusion models. The central empirical result (70.97% reduction and distinct failure-class coverage) is presented as confirmation of a derived prediction rather than a fit. No equations, parameter-fitting steps, or self-citations are shown that would make any claim equivalent to its inputs by construction. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that the generative process can be decomposed into four structurally distinct intervention points; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption The generative process unfolds as a trajectory of internal states with four structurally distinct components where knowledge can intervene.
    This premise directly enables the mapping to surface, trajectory, latent, and parametric layers as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5776 in / 1239 out tokens · 24818 ms · 2026-06-28T01:26:37.250968+00:00 · methodology

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

34 extracted references · 7 canonical work pages · 3 internal anchors

  1. [1]

    Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models.ACM transactions on Graphics (TOG), 42 (4):1–10, 2023

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models.ACM transactions on Graphics (TOG), 42 (4):1–10, 2023. 5

    2023

  2. [2]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022. 4, 5, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  3. [3]

    Diffusion models beat gans on image synthesis, 2021

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 1, 4, 6

    2021

  4. [4]

    Erasing concepts from dif- fusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto- Kaufman, and David Bau. Erasing concepts from dif- fusion models. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 2426– 2436, 2023. 4, 5, 6

    2023

  5. [5]

    Retrieval-augmented generation for large language models: A survey, 2024

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey, 2024. 1

    2024

  6. [6]

    Analog and multi-modal manufacturing datasets acquired on the future factories platform.arXiv preprint arXiv:2401.15544, 2024

    Ramy Harik, Fadi El Kalach, Jad Samaha, Devon Clark, Drew Sander, Philip Samaha, Liam Burns, Ibrahim Yousif, Victor Gadow, Theodros Tarekegne, et al. Analog and multi-modal manufacturing datasets acquired on the future factories platform.arXiv preprint arXiv:2401.15544, 2024. 5

    work page arXiv 2024

  7. [7]

    Prompt-to-Prompt Image Editing with Cross Attention Control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aber- man, Yael Pritch, and Daniel Cohen-Or. Prompt- to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 4, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Classifier-free diffu- sion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffu- sion guidance, 2022. 1, 4, 6

    2022

  9. [9]

    Denoising diffusion probabilistic models.Advances in neural in- formation processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural in- formation processing systems, 33:6840–6851, 2020. 2

    2020

  10. [10]

    Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Clau- dia D’amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, Jos ´e Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel-Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann. Knowledge graphs.ACM Computing Sur...

    2021

  11. [11]

    Lexically constrained decoding for sequence generation using grid beam search

    Chris Hokamp and Qun Liu. Lexically constrained decoding for sequence generation using grid beam search. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 1535–1546, 2017. 4, 6

    2017

  12. [12]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022. 1, 4, 5, 6

    2022

  13. [13]

    A survey on hallucination in large language mod- els: Principles, taxonomy, challenges, and open ques- tions.ACM Transactions on Information Systems, 43 (2):1–55, 2025

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Wei- hua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language mod- els: Principles, taxonomy, challenges, and open ques- tions.ACM Transactions on Information Systems, 43 (2):1–55, 2025. 1

    2025

  14. [14]

    Survey of hallucination in natural language generation.ACM Computing Sur- veys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM Computing Sur- veys, 55(12):1–38, 2023. 1, 4

    2023

  15. [15]

    Retrieval- augmented generation for knowledge-intensive nlp tasks, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen tau Yih, Tim Rockt ¨aschel, Sebastian Riedel, and Douwe Kiela. Retrieval- augmented generation for knowledge-intensive nlp tasks, 2021. 1, 4, 6

    2021

  16. [16]

    Inference-time inter- vention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

    Kenneth Li, Oam Patel, Fernanda Vi ´egas, Hanspeter Pfister, and Martin Wattenberg. Inference-time inter- vention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023. 4, 6

    2023

  17. [17]

    Learn to explain: Multi- modal reasoning via thought chains for science ques- tion answering, 2022

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai- Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multi- modal reasoning via thought chains for science ques- tion answering, 2022. 1

    2022

  18. [18]

    Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022. 1, 4, 6

    2022

  19. [19]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InPro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6038–6047, 2023. 4, 6

    2023

  20. [20]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agar- wal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haim- ing Bao, Mohammad Bavarian, Jeff Belgum, Ir- wan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christop...

    2024

  21. [21]

    Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36 (7):3580–3599, 2024

    Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Ji- apu Wang, and Xindong Wu. Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36 (7):3580–3599, 2024. 1

    2024

  22. [22]

    Assemai: Interpretable image-based anomaly detection for manufacturing pipelines

    Renjith Prasad, Chathurangi Shyalika, Fadi El Kalach, Revathy Venkataramanan, Ramtin Zand, Ramy Harik, and Amit Sheth. Assemai: Interpretable image-based anomaly detection for manufacturing pipelines. In 2024 International Conference on Machine Learning and Applications (ICMLA), pages 1720–1727. IEEE,

    2024

  23. [23]

    Detonate: A benchmark for text-to-image alignment and kernel- ized direct preference optimization.arXiv preprint arXiv:2506.14903, 2025

    Renjith Prasad, Abhilekh Borah, Hasnat Md Abdul- lah, Chathurangi Shyalika, Gurpreet Singh, Ritvik Garimella, Rajarshi Roy, Harshul Surana, Nasrin Imanpour, Suranjana Trivedy, et al. Detonate: A benchmark for text-to-image alignment and kernel- ized direct preference optimization.arXiv preprint arXiv:2506.14903, 2025. 7

    work page arXiv 2025

  24. [24]

    Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610,

    Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tram`er. Red-teaming the stable dif- fusion safety filter.arXiv preprint arXiv:2210.04610,

    work page arXiv

  25. [25]

    High- resolution image synthesis with latent diffusion mod- els, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion mod- els, 2022. 1, 2

    2022

  26. [26]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500–22510, 2023. 1, 4, 6

    2023

  27. [27]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam- yar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1

    2022

  28. [28]

    Safe latent diffu- sion: Mitigating inappropriate degeneration in diffu- sion models

    Patrick Schramowski, Manuel Brack, Bj ¨orn Deis- eroth, and Kristian Kersting. Safe latent diffu- sion: Mitigating inappropriate degeneration in diffu- sion models. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 22522–22531, 2023. 7

    2023

  29. [29]

    Shades of knowledge-infused learning for enhancing deep learning.IEEE Internet Computing, 23(6):54–63, 2020

    Amit Sheth, Manas Gaur, Ugur Kursuncu, and Ruwan Wickramarachchi. Shades of knowledge-infused learning for enhancing deep learning.IEEE Internet Computing, 23(6):54–63, 2020. 2

    2020

  30. [30]

    Nsf-map: neurosymbolic multimodal fusion for robust and interpretable anomaly prediction in assembly pipelines

    Chathurangi Shyalika, Renjith Prasad, Fadi El Kalach, Revathy Venkataramanan, Ramtin Zand, Ramy Harik, and Amit Sheth. Nsf-map: neurosymbolic multimodal fusion for robust and interpretable anomaly prediction in assembly pipelines. InProceedings of the Thirty- Fourth International Joint Conference on Artificial In- telligence, pages 9330–9338, 2025. 5

    2025

  31. [31]

    Safree: Training-free and adaptive guard for safe text-to-image and video gen- eration.arXiv preprint arXiv:2410.12761, 2024

    Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video gen- eration.arXiv preprint arXiv:2410.12761, 2024. 7

    work page arXiv 2024

  32. [32]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 3836– 3847, 2023. 1, 4, 6

    2023

  33. [33]

    Knowgpt: Knowl- edge graph based prompting for large language mod- els.Advances in neural information processing sys- tems, 37:6052–6080, 2024

    Qinggang Zhang, Junnan Dong, Hao Chen, Daochen Zha, Zailiang Yu, and Xiao Huang. Knowgpt: Knowl- edge graph based prompting for large language mod- els.Advances in neural information processing sys- tems, 37:6052–6080, 2024. 4, 6

    2024

  34. [34]

    Representation Engineering: A Top-Down Approach to AI Transparency

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down ap- proach to ai transparency, 2023.URL https://arxiv. org/abs/2310.01405, 97, 2022. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023