Semantic-aware Graph-guided Behavior Sequences Generation with Large Language Models for Smart Homes

Dan Zhao; Jingyu Xiao; Qing Li; Qingsong Zou; Yong Jiang; Yuhang Wang; Zhiyao Xu

REVIEW 4 major objections 3 minor 65 references

Semantic-aware Graph-guided Behavior Sequences Generation with Large Language Models for Smart Homes

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LLM-generated behavior data restores smart home performance after drift

desk verdict The submission file is the wrong paper, so the claimed 85%/70% gains are untestable; the SmartGen abstract alone reads plausible but cannot justify peer review. read the letter →

arxiv 2508.03484 v1 pith:CX3OC2U7 submitted 2025-08-05 cs.AI

Zhiyao Xu , Dan Zhao , Qingsong Zou , Qing Li , Yong Jiang , Yuhang Wang , Jingyu Xiao This is my paper

classification cs.AI

keywords smarthomesbehavioraldriftlargelanguagemodelssyntheticdatagenerationanomalydetectionbehaviorpredictioncontinualadaptationsequences

verification ladder T0 review T1 audit T2 compute T3 formal

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 introduces SmartGen, an LLM-based pipeline that manufactures synthetic behavior sequences for smart homes so that downstream models can keep working when people's routines change. Behavioral drift—seasonal changes, new habits, shifting schedules—makes models trained on static logs brittle, while collecting fresh real data is slow, costly, and privacy-sensitive. SmartGen attempts to show that generated, filtered behavior data can stand in for that fresh data. On three real-world datasets the authors report average gains of 85.43% on anomaly detection and 70.51% on behavior prediction after drift. The claim, if right, would let smart home systems adapt continually without new field collection.

What carries the argument

The carrying mechanism is graph-guided sequence synthesis: a behavior relationship graph built from the original logs captures frequent transitions, and those transitions are encoded into prompts so the LLM generates new sequences that preserve core patterns while reflecting contextual change. Around it sit three supporting components: a split module that cuts long sequences into semantically coherent chunks under dual time-span constraints, a compression step that clusters behavior mappings in latent space to shorten inputs without losing meaning, and a two-stage outlier filter that removes implausible or semantically inconsistent outputs.

What would settle it

Inspect the evaluation protocol for the three datasets: if the drifted test sequences are produced by the same SmartGen pipeline rather than collected after a real change point, retrain a downstream model on generated data and test it on genuine post-drift logs; the reported transfer gains would fail if the retrained model performs no better than the unadapted baseline on actually collected drifted data.

Watch

Extended reading notes

Core claim

SmartGen's central claim is that four stages—time-and-semantic splitting of long sequences, semantic-aware compression by clustering behavior mappings in latent space, graph-guided synthesis where a behavior relationship graph encodes frequent transitions into prompts, and a two-stage outlier filter—can make an LLM produce behavior sequences whose statistical structure is close enough to real drifted behavior that retraining on them restores model performance. The reported numbers (85.43% improvement on anomaly detection, 70.51% on behavior prediction, averaged over three real-world datasets) are the evidence offered for that transfer. The intended consequence is that continual adaptation can be driven by synthetic data instead of slow, costly, privacy-risky recollection.

Load-bearing premise

The load-bearing premise is the transfer premise: LLM-generated sequences, after clustering and filtering, faithfully reproduce the statistics of real behavior under drift—frequencies, transitions, and anomaly semantics—so models trained on them generalize to real data; the abstract states this goal but does not confirm whether the drift test set is real held-out data or synthetic output of the same pipeline.

Editorial extensions

If this is right

If generated data transfers, smart home models can be retrained on demand after drift instead of waiting for months of fresh real data.
Anomaly detection and behavior prediction can share the same synthetic corpus, since both tasks improved on it.
Privacy pressure lessens, because synthetic logs could be produced and shared without exposing raw household behavior.
The four-component pipeline gives a concrete recipe for using LLMs in continual adaptation settings beyond smart homes, wherever drift outpaces data collection.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The supplied full text does not contain the SmartGen experiments, so the architecture details and dataset protocol rest entirely on the abstract; a reader should verify the evaluation protocol before relying on the 85.43% and 70.51% figures.
The transfer claim depends on the drift test set being real held-out data; if it is generated by the same SmartGen pipeline, the reported gains would partly measure self-consistency rather than transfer, and the abstract does not settle this.
A direct testable extension is to compare models retrained on SmartGen output against models retrained on an equal-sized sample of real drifted data, holding downstream architecture and compute constant.
The privacy benefit is double-edged: if synthetic data is close enough to real behavior to retrain models, it may also be close enough to reveal household patterns, so a membership-inference audit would be a useful follow-up.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First thing you should know: the PDF attached to this submission is not the SmartGen paper. It's LRQ-DiT, a quantization paper for diffusion transformers. So I could only review the SmartGen abstract. That is not a minor formatting issue; it's the entire empirical content walking out of the room.

What the abstract describes is genuinely reasonable. The problem is real: smart-home models degrade under behavioral drift, and collecting fresh data is slow, expensive, and privacy-sensitive. The proposed pipeline—time-semantic splitting, latent clustering compression, graph-guided transition prompting, and a two-stage outlier filter—is a sensible combination of known ideas. I would have believed the system works as a data-augmentation tool before seeing any numbers. The writing is clear, and the four components are described without obvious hand-waving.

The soft spots are proportionate to the fact that we only have the abstract. The headline numbers (85.43% anomaly detection, 70.51% behavior prediction) are bare averages with no named baselines, no dataset details, no variance, and no protocol. The abstract says 'three real-world datasets' but doesn't say whether the drift test segments are real held-out data or outputs of the same generator. If they're generator outputs, some of the gain probably comes from matching the generator's own distribution rather than transferring to real drift. That's not an accusation—it's an open question that the full paper would need to answer. And the method has at least four free parameters (span constraints, clustering granularity, transition frequency cutoff, filter thresholds) that could be tuned on the test set.

I can't check the math, the data, or the citation pattern because none of it is in front of me. The code link exists but that's not enough.

My recommendation: do not send this to peer review in this state. The correct course is to return the submission to the authors for the correct manuscript, then evaluate that. If the actual SmartGen paper is as clear as its abstract and resolves the held-out-data question, it deserves a serious referee—the topic is useful, and the framework is worth engaging with. But as it stands, reviewing it would mean reviewing an abstract.

Take it to the reading group? No, not until there's a real paper.

Referee Report

4 major / 3 minor

Summary. The paper, arXiv:2508.03484, proposes SmartGen, an LLM-based framework for synthesizing context-aware smart-home behavior sequences to support continual adaptation under behavioral drift. The abstract describes four components: Time and Semantic-aware Split, Semantic-aware Sequence Compression, Graph-guided Sequence Synthesis, and Two-stage Outlier Filter, and reports that training on generated data improves anomaly detection by 85.43% and behavior prediction by 70.51% on average across three real-world datasets. The submitted full text, however, is not the SmartGen paper: it is the LRQ-DiT post-training quantization paper for diffusion transformers, with no mention of smart homes, behavioral drift, SmartGen, or the reported experiments. Consequently, the technical content and experimental evidence behind the abstract's claims are not available in the submitted manuscript.

Significance. If the headline claims hold, SmartGen would address a practically important problem: static smart-home models degrade under behavioral drift, and recollecting real data is slow, costly, and privacy-sensitive. A pipeline that synthesizes training data good enough to restore downstream model performance would be valuable, and the graph-guided prompting and outlier filtering ideas are plausible design choices. The paper also deserves credit for committing to public code and for reporting concrete effect sizes, conditional on those numbers being verifiable. However, the evaluation must establish transfer from generated sequences to real held-out drift data, and the current submission does not supply the protocol needed to assess this. The significance of the claimed result cannot be evaluated from the abstract alone, especially given that the attached full text is an unrelated manuscript.

major comments (4)

[Full text] The submitted full text is the LRQ-DiT quantization paper (arXiv:2508.03485v3), not the SmartGen paper under review. None of the four SmartGen components, the three real-world datasets, or the anomaly-detection and behavior-prediction experiments appear in the body. The central claims of the paper are therefore unsupported by the submitted document, and the technical evaluation cannot be reproduced or checked. The authors should resubmit with the correct full text before any further review.
[Abstract] The headline improvements (85.43% for anomaly detection, 70.51% for behavior prediction) are reported as bare averages. No baselines are named, no dataset details are given, no variance or confidence intervals are reported, and no evaluation protocol (train/test split, drift simulation, number of runs) is described. As stated, the numbers are unverifiable even from the abstract.
[Abstract] The abstract does not state whether the drift evaluation segments are real held-out data or sequences generated by SmartGen itself. If the test set is produced by the same generation pipeline, then retraining on SmartGen outputs and testing on SmartGen outputs measures distributional alignment with the generator, not transfer to genuine behavioral drift. The authors should specify the provenance of the test data and include a protocol with real held-out drift segments and a no-generation retraining baseline.
[Abstract] The four components rely on several adjustable thresholds and granularities: dual time-span split constraints, latent clustering granularity, graph transition frequency cutoff, and outlier filter thresholds. The abstract provides no sensitivity analysis or justification for these choices. Because these parameters control what data is generated and retained, their settings are load-bearing for the claimed improvements; the full text should report ablations and robustness checks.

minor comments (3)

[Abstract] The term 'behavioral drift' is used without a formal definition; the paper should state what distributional changes are assumed, such as shifts in activity frequencies, transition probabilities, or anomaly semantics.
[Abstract] The percentages '85.43%' and '70.51%' should be accompanied by standard deviations or confidence intervals, especially given the phrase 'on average'.
[Abstract / GitHub link] The abstract states that code is available at https://github.com/horizonsinzqs/SmartGen, but the submitted manuscript does not allow the reviewer to verify that this link corresponds to the SmartGen implementation; the unrelated full text should be corrected so the link and the described system are consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrable circularity in the provided document; the SmartGen abstract cannot be checked because the supplied body is a different paper, LRQ-DiT.

full rationale

The submitted full text is not the SmartGen paper described in the abstract: it is LRQ-DiT, a post-training quantization method for diffusion transformers, with different authors and no mention of smart homes, behavioral drift, anomaly detection, or the claimed 85.43%/70.51% improvements. This is a serious completeness defect: the central SmartGen claim cannot be verified from the provided document. However, a missing or mismatched full text is not itself circularity. The LRQ-DiT text that is actually present is a self-contained engineering contribution. Its components, Twin-Log Quantization (TLQ) and Adaptive Rotation Scheme (ARS), are motivated by distributional observations about DiT weights and activations, and they are evaluated against external baselines (SmoothQuant, QuaRot, DuQuant, ViDiT-Q) on external models and prompt sets with ablation studies (Table 5). The theoretical analysis in Section 4.3 derives an error-reduction ratio from stated assumptions about the logarithmic quantizer and orthogonal rotations; it is a mathematical derivation, not a prediction forced by construction. The only potential circularity flagged by the reader's take, namely that SmartGen's drift test data might be synthetic outputs of the same generation pipeline, cannot be evaluated because the SmartGen full text is absent. Under the hard rule that circularity must be exhibited by quoting the paper and showing the reduction, ambiguity about the test protocol is insufficient. Therefore the honest finding is no significant circularity in the provided document, with the caveat that the SmartGen claim is unsupported by the supplied text rather than independently confirmed.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework introduces no new physical entities; it is an algorithmic pipeline. The central claim rests on three domain assumptions about transfer validity and evaluation protocol, plus four families of tunable parameters whose values are not reported in the abstract. Because the attached full text is a different paper, none of these choices can be audited here.

free parameters (4)

Dual time-span split constraints = not reported in abstract
The Time and Semantic-aware Split module divides long behavior sequences under dual time-span constraints; the specific constraint values are free choices that are not given in the abstract.
Latent clustering granularity = not reported in abstract
Semantic-aware Sequence Compression clusters behavior mappings in latent space; the number of clusters or compression ratio is a hand-chosen detail absent from the abstract.
Graph transition frequency cutoff = not reported in abstract
Graph-guided Sequence Synthesis encodes frequent transitions into prompts; the cutoff that defines 'frequent' shapes the generated distribution and is not specified.
Outlier filter thresholds = not reported in abstract
The Two-stage Outlier Filter removes implausible or inconsistent outputs; its acceptance thresholds are free parameters with no values in the abstract.

assumptions (3)

domain assumption Generated sequences preserve the statistical structure of real behavior under drift
SmartGen's usefulness depends on synthetic sequences matching real transition frequencies and anomaly semantics; the abstract asserts this goal but provides no distributional validation.
domain assumption Drift evaluation uses real held-out data
The abstract reports gains under behavioral drift on three real-world datasets but does not say whether the test segments are real or produced by the same pipeline; this determines whether the gains reflect transfer.
domain assumption The three datasets are representative of smart home drift
Generalization to real deployments rests on dataset representativeness; the datasets are not named or characterized in the abstract.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic-aware Graph-guided Behavior Sequences Generation with Large Language Models for Smart Homes." pith.science (2026). https://pith.science/paper/CX3OC2U7

@misc{pith2026250803484,
  author       = {Pith},
  title        = {Pith review of: Semantic-aware Graph-guided Behavior Sequences Generation with Large Language Models for Smart Homes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CX3OC2U7}},
  note         = {Machine review of arXiv:2508.03484}
}

read the original abstract

As smart homes become increasingly prevalent, intelligent models are widely used for tasks such as anomaly detection and behavior prediction. These models are typically trained on static datasets, making them brittle to behavioral drift caused by seasonal changes, lifestyle shifts, or evolving routines. However, collecting new behavior data for retraining is often impractical due to its slow pace, high cost, and privacy concerns. In this paper, we propose SmartGen, an LLM-based framework that synthesizes context-aware user behavior data to support continual adaptation of downstream smart home models. SmartGen consists of four key components. First, we design a Time and Semantic-aware Split module to divide long behavior sequences into manageable, semantically coherent subsequences under dual time-span constraints. Second, we propose Semantic-aware Sequence Compression to reduce input length while preserving representative semantics by clustering behavior mapping in latent space. Third, we introduce Graph-guided Sequence Synthesis, which constructs a behavior relationship graph and encodes frequent transitions into prompts, guiding the LLM to generate data aligned with contextual changes while retaining core behavior patterns. Finally, we design a Two-stage Outlier Filter to identify and remove implausible or semantically inconsistent outputs, aiming to improve the factual coherence and behavioral validity of the generated sequences. Experiments on three real-world datasets demonstrate that SmartGen significantly enhances model performance on anomaly detection and behavior prediction tasks under behavioral drift, with anomaly detection improving by 85.43% and behavior prediction by 70.51% on average. The code is available at https://github.com/horizonsinzqs/SmartGen.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 63 canonical work pages

[1]

Quarot: Outlier-free 4-bit inference in rotated llms.�� , 2024

Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms.�� , 2024

work page 2024
[2]

Quip: 2-bit quantiza- tion of large language models with guarantees.�� , 36:4396–4429, 2023

Jerry Chee, Yaohui Cai, V olodymyr Kuleshov, and Christopher M De Sa. Quip: 2-bit quantiza- tion of large language models with guarantees.�� , 36:4396–4429, 2023

work page 2023
[3]

Conceptcraft: One-shot personalized text-to-image generation via object-background disentanglement.�� , 2025

Haibo Chen, Zhiwen Zuo, Lei Zhao, Jun Li, and Jian Yang. Conceptcraft: One-shot personalized text-to-image generation via object-background disentanglement.�� , 2025

work page 2025
[4]

Pixart-� : Fast training of diffusion transformer for photorealistic text-to-image synthesis.�� , 2023

Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-� : Fast training of diffusion transformer for photorealistic text-to-image synthesis.�� , 2023

work page 2023
[5]

Pixart- �: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- �: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In�� , pages 74–91. Springer, 2024

work page 2024
[6]

Q-dit: Accurate post-training quantization for diffusion transformers

Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. In�� , pages 28306–28315, 2025

work page 2025
[7]

Microsoft coco captions: Data collection and evaluation server.�� , 2015

Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server.�� , 2015

work page 2015
[8]

Llm.int8(): 8-bit matrix multiplication for transformers at scale

Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm.int8(): 8-bit matrix multiplication for transformers at scale. In�� , 2022

work page 2022

Show all 65 references

[9]

Scaling rectified flow trans- formers for high-resolution image synthesis

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. In�� ...

2024
[10]

Q-vdit: Towards accurate quantization and distillation of video-generation diffusion transformers.�� , 2025

Weilun Feng, Chuanguang Yang, Haotong Qin, Xiangqi Li, Yu Wang, Zhulin An, Libo Huang, Boyu Diao, Zixiang Zhao, Yongjun Xu, et al. Q-vdit: Towards accurate quantization and distillation of video-generation diffusion transformers.�� , 2025

2025
[11]

Shiftquant: Towards accurate and efficient sub-8-bit integer training.�� , 2025

Wenjin Guo, Donglai Liu, Weiying Xie, Yunsong Li, Xuefei Ning, Zihan Meng, Shulin Zeng, Jie Lei, Zhenman Fang, and Yu Wang. Shiftquant: Towards accurate and efficient sub-8-bit integer training.�� , 2025

2025
[12]

Ptqd: Accurate post-training quantization for diffusion models.�� , 2023

Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models.�� , 2023

2023
[13]

Gans trained by a two time-scale update rule converge to a local nash equilibrium.�� , 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.�� , 30, 2017

2017
[14]

Tfmq-dm: Temporal feature maintenance quantization for diffusion models

Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. In�� , pages 7362–7371, 2024

2024
[15]

Image-level memorization detection via inversion-based inference perturbation

Yue Jiang, Haokun Lin, Yang Bai, Bo Peng, Zhili Liu, Yueming Lyu, Yong Yang, Jing Dong, et al. Image-level memorization detection via inversion-based inference perturbation. In�� , 2024. 14

2024
[16]

Hunyuanvideo: A systematic framework for large video generative models.�� , 2024

Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.�� , 2024

2024
[17]

Leveraging vision-language models to select trustworthy super-resolution samples generated by diffusion models.�� , 2025

Cansu Korkmaz, A Murat Tekalp, and Zafer Do ˘gan. Leveraging vision-language models to select trustworthy super-resolution samples generated by diffusion models.�� , 2025

2025
[18]

Flux.��, 2024

Black Forest Labs. Flux.��, 2024

2024
[19]

Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

2024
[20]

Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.�� , 2024

Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.�� , 2024

2024
[21]

Q-diffusion: Quantizing diffusion models

Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In�� , pages 17535–17545, 2023

2023
[22]

Repq-vit: Scale reparameterization for post-training quantization of vision transformers

Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repq-vit: Scale reparameterization for post-training quantization of vision transformers. In�� , pages 17227–17236, 2023

2023
[23]

Dvd-quant: Data-free video diffusion transformers quantization.�� , 2025

Zhiteng Li, Hanxuan Li, Junyi Wu, Kai Liu, Linghe Kong, Guihai Chen, Yulun Zhang, and Xiaokang Yang. Dvd-quant: Data-free video diffusion transformers quantization.�� , 2025

2025
[24]

Mope-clip: Structured pruning for efficient vision-language models with module-wise pruning error metric

Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. Mope-clip: Structured pruning for efficient vision-language models with module-wise pruning error metric. In�� ...

2024
[25]

Duquant: Distributing outliers via dual transformation makes stronger quantized llms.�� , 37:87766–87800, 2024

Haokun Lin, Haobo Xu, Yichen Wu, Jingzhi Cui, Yingtao Zhang, Linzhan Mou, Linqi Song, Zhenan Sun, and Ying Wei. Duquant: Distributing outliers via dual transformation makes stronger quantized llms.�� , 37:87766–87800, 2024

2024
[26]

Toklip: Marry visual tokens to clip for multimodal comprehension and generation.�� , 2025

Haokun Lin, Teng Wang, Yixiao Ge, Yuying Ge, Zhichao Lu, Ying Wei, Qingfu Zhang, Zhenan Sun, and Ying Shan. Toklip: Marry visual tokens to clip for multimodal comprehension and generation.�� , 2025

2025
[27]

Fq-vit: Post-training quantization for fully quantized vision transformer

Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In Luc De Raedt, editor,�� , pages 1173–1179, 2022

2022
[28]

Intactkv: Improving large language model quantization by keeping pivot tokens intact.�� , 2024

Ruikang Liu, Haoli Bai, Haokun Lin, Yuening Li, Han Gao, Zhengzhuo Xu, Lu Hou, Jun Yao, and Chun Yuan. Intactkv: Improving large language model quantization by keeping pivot tokens intact.�� , 2024

2024
[29]

Diffvein: A unified diffusion network for finger vein segmentation and authentication.�� , 2024

Yanjun Liu, Wenming Yang, and Qingmin Liao. Diffvein: A unified diffusion network for finger vein segmentation and authentication.�� , 2024

2024
[30]

Sora: A review on background, technology, limitations, and opportunities of large vision models.�� , 2024

Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models.�� , 2024

2024
[31]

Spinquant: Llm quantization with learned rotations.�� , 2024

Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. Spinquant: Llm quantization with learned rotations.�� , 2024. 15

2024
[32]

Latte: Latent diffusion transformer for video generation.�� , 2024

Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.�� , 2024

2024
[33]

Sdxl: Improving latent diffusion models for high-resolution image synthesis.�� , 2023

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.�� , 2023

2023
[34]

High- resolution image synthesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In�� , pages 10684–10695, 2022

2022
[35]

Post-training quantization on diffusion models

Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In�� , pages 1972–1981, 2023

1972
[36]

Omniquant: Omnidirectionally calibrated quantiza- tion for large language models.�� , 2023

Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqian Li, Kaipeng Zhang, Peng Gao, Yu Qiao, and Ping Luo. Omniquant: Omnidirectionally calibrated quantiza- tion for large language models.�� , 2023

2023
[37]

Exploring pruning-based efficient object tracking via hybrid knowledge distillation.�� , 2025

Yidong Song, Shilei Wang, Zhaochuan Zeng, Jikai Zheng, Zhenhua Wang, and Jifeng Ning. Exploring pruning-based efficient object tracking via hybrid knowledge distillation.�� , 2025

2025
[38]

Massive activations in large language models.�� , 2024

Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. Massive activations in large language models.�� , 2024

2024
[39]

Quip#: Even better llm quantization with hadamard incoherence and lattice codebooks.�� , 2024

Albert Tseng, Jerry Chee, Qingyao Sun, V olodymyr Kuleshov, and Christopher De Sa. Quip#: Even better llm quantization with hadamard incoherence and lattice codebooks.�� , 2024

2024
[40]

A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions

Jack Urbanek, Florian Bordes, Pietro Astolfi, Mary Williamson, Vasu Sharma, and Adriana Romero-Soriano. A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In�� ...

2024
[41]

Wan: Open and advanced large-scale video generative models.�� , 2025

Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.�� , 2025

2025
[42]

Quest: Low-bit diffusion model quantization via efficient selective finetuning.�� , 2024

Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning.�� , 2024

2024
[43]

Ptq4dit: Post-training quantization for diffusion transformers.�� , 37: 62732–62755, 2024

Junyi Wu, Haoxuan Wang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Ptq4dit: Post-training quantization for diffusion transformers.�� , 37: 62732–62755, 2024

2024
[44]

Adalog: Post- training quantization for vision transformers with adaptive logarithm quantizer

Zhuguanyu Wu, Jiaxin Chen, Hanwen Zhong, Di Huang, and Yunhong Wang. Adalog: Post- training quantization for vision transformers with adaptive logarithm quantizer. In�� , pages 411–427. Springer, 2024

2024
[45]

Smoothquant: Accurate and efficient post-training quantization for large language models

Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In�� , pages 38087–38099. PMLR, 2023

2023
[46]

Patch-wise mixed-precision quantization of vision transformer

Junrui Xiao, Zhikai Li, Lianwei Yang, and Qingyi Gu. Patch-wise mixed-precision quantization of vision transformer. In�� , pages 1–7. IEEE, 2023

2023
[47]

Binaryvit: Towards efficient and accurate binary vision transformers.�� , 2024

Junrui Xiao, Zhikai Li, Jianquan Li, Lianwei Yang, and Qingyi Gu. Binaryvit: Towards efficient and accurate binary vision transformers.�� , 2024. 16

2024
[48]

Ttaq: Towards stable post-training quantization in continuous domain adaptation.�� , 2024

Junrui Xiao, Zhikai Li, Lianwei Yang, Yiduo Mei, and Qingyi Gu. Ttaq: Towards stable post-training quantization in continuous domain adaptation.�� , 2024

2024
[49]

Sana: Efficient high-resolution image synthesis with linear diffusion transformers.�� , 2024

Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers.�� , 2024

2024
[50]

Imagereward: Learning and evaluating human preferences for text-to-image generation

Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. �� , 36:15903–15935, 2023

2023
[51]

Improving extreme low-bit quantization with soft threshold.�� , 33(4):1549–1563, 2022

Weixiang Xu, Fanrong Li, Yingying Jiang, Andrew Yong, Xiangyu He, Peisong Wang, and Jian Cheng. Improving extreme low-bit quantization with soft threshold.�� , 33(4):1549–1563, 2022

2022
[52]

Dopq-vit: Towards distribution-friendly and outlier-aware post-training quantization for vision transform- ers.�� , 2024

Lianwei Yang, Haisong Gong, Haokun Lin, Yichen Wu, Zhenan Sun, and Qingyi Gu. Dopq-vit: Towards distribution-friendly and outlier-aware post-training quantization for vision transform- ers.�� , 2024

2024
[53]

Mgrq: Post-training quantization for vision transformer with mixed granularity reconstruction.�� , 2024

Lianwei Yang, Zhikai Li, Junrui Xiao, Haisong Gong, and Qingyi Gu. Mgrq: Post-training quantization for vision transformer with mixed granularity reconstruction.�� , 2024

2024
[54]

Cogvideox: Text-to-video diffusion models with an expert transformer.�� , 2024

Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.�� , 2024

2024
[55]

Zeroquant: Efficient and affordable post-training quantization for large-scale transformers

Zhewei Yao, Reza Yazdani Aminabadi, Minjia Zhang, Xiaoxia Wu, Conglong Li, and Yuxiong He. Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. �� , 35:27168–27183, 2022

2022
[56]

Diff-transformer: Heterogeneous feature fusion network for multisource remote sensing classification.�� , 2025

Zhihao Ying, Jie Guo, Yunsong Li, Yu’e Gao, and Chenyu Li. Diff-transformer: Heterogeneous feature fusion network for multisource remote sensing classification.�� , 2025

2025
[57]

Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization

Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In�� , pages 191–207. Springer, 2022

2022
[58]

Spatial-spectral aggregation transformer with diffusion prior for hyperspectral image super-resolution.�� , 2024

Mingyang Zhang, Xiangyu Wang, Shuang Wu, Zhaoyang Wang, Maoguo Gong, Yu Zhou, Fenlong Jiang, and Yue Wu. Spatial-spectral aggregation transformer with diffusion prior for hyperspectral image super-resolution.�� , 2024

2024
[59]

Tvg: A training-free transition video generation method with diffusion models.�� , 2025

Rui Zhang, Yaosen Chen, Yuegen Liu, Wei Wang, Xuming Wen, and Hongxia Wang. Tvg: A training-free transition video generation method with diffusion models.�� , 2025

2025
[60]

Dfmc: feature-driven data-free knowledge distillation.�� , 2025

Zherui Zhang, Rongtao Xu, Changwei Wang, Wenhao Xu, Shunpeng Chen, Shibiao Xu, Guangyuan Xu, and Li Guo. Dfmc: feature-driven data-free knowledge distillation.�� , 2025

2025
[61]

Exploiting channel similarity for network pruning.�� , 33(9):5049–5061, 2023

Chenglong Zhao, Yunxiang Zhang, and Bingbing Ni. Exploiting channel similarity for network pruning.�� , 33(9):5049–5061, 2023

2023
[62]

Vidit-q: Efficient and accu- rate quantization of diffusion transformers for image and video generation.�� , 2024

Tianchen Zhao, Tongcheng Fang, Haofeng Huang, Enshu Liu, Rui Wan, Widyadewi Soedar- madji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, et al. Vidit-q: Efficient and accu- rate quantization of diffusion transformers for image and video generation.�� ...

2024
[63]

Open-sora: Democratizing efficient video production for all

Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. �� , 2024

2024
[64]

I&s-vit: An inclusive & stable method for pushing the limit of post-training vits quantization.�� , 2023

Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&s-vit: An inclusive & stable method for pushing the limit of post-training vits quantization.�� , 2023

2023
[65]

Doge: Towards versatile visual document grounding and referring.�� , 2024

Yinan Zhou, Yuxin Chen, Haokun Lin, Shuyu Yang, Li Zhu, Zhongang Qi, Chen Ma, and Ying Shan. Doge: Towards versatile visual document grounding and referring.�� , 2024. 18 A Visual Comparison We provide more visual comparison results, as follows: • F...

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Quarot: Outlier-free 4-bit inference in rotated llms.�� , 2024

Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms.�� , 2024

work page 2024

[2] [2]

Quip: 2-bit quantiza- tion of large language models with guarantees.�� , 36:4396–4429, 2023

Jerry Chee, Yaohui Cai, V olodymyr Kuleshov, and Christopher M De Sa. Quip: 2-bit quantiza- tion of large language models with guarantees.�� , 36:4396–4429, 2023

work page 2023

[3] [3]

Conceptcraft: One-shot personalized text-to-image generation via object-background disentanglement.�� , 2025

Haibo Chen, Zhiwen Zuo, Lei Zhao, Jun Li, and Jian Yang. Conceptcraft: One-shot personalized text-to-image generation via object-background disentanglement.�� , 2025

work page 2025

[4] [4]

Pixart-� : Fast training of diffusion transformer for photorealistic text-to-image synthesis.�� , 2023

Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-� : Fast training of diffusion transformer for photorealistic text-to-image synthesis.�� , 2023

work page 2023

[5] [5]

Pixart- �: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- �: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In�� , pages 74–91. Springer, 2024

work page 2024

[6] [6]

Q-dit: Accurate post-training quantization for diffusion transformers

Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. In�� , pages 28306–28315, 2025

work page 2025

[7] [7]

Microsoft coco captions: Data collection and evaluation server.�� , 2015

Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server.�� , 2015

work page 2015

[8] [8]

Llm.int8(): 8-bit matrix multiplication for transformers at scale

Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm.int8(): 8-bit matrix multiplication for transformers at scale. In�� , 2022

work page 2022