ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection

Hongxiao Fei; Jun Long; Ke Ma; Liujie Hua; Yueyi Luo; Zhen Dai

REVIEW 3 major objections 4 minor 1 cited by

ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Co-designed adapters close CLIP's zero-shot anomaly gap

desk verdict The abstract is a clean pitch for a plausible incremental method, but the submitted full text is another paper entirely, so there is nothing to verify. read the letter →

arxiv 2508.07819 v6 pith:HUKZT7TP submitted 2025-08-11 cs.CV cs.AIcs.LG

Ke Ma , Jun Long , Hongxiao Fei , Liujie Hua , Zhen Dai , Yueyi Luo This is my paper

classification cs.CVcs.AIcs.LG

keywords zero-shotanomalydetectionvision-languagemodelsCLIPparameter-efficientfine-tuningconvolutionalLoRAcross-modalfusiondynamicgatewaydenseprediction

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 argues that pre-trained vision-language models such as CLIP fail at zero-shot anomaly detection because they lack the local inductive biases needed for dense prediction and because their text-image fusion is static. It claims that jointly co-designing the two sides fixes this: a parameter-efficient Convolutional Low-Rank Adaptation (Conv-LoRA) adapter adds local inductive biases to the visual encoder, and a Dynamic Fusion Gateway (DFG) lets visual context modulate text prompts. Together, the authors argue, this yields superior accuracy and robustness on industrial and medical anomaly detection benchmarks, and the synergy between the two components is what makes the difference.

What carries the argument

Conv-LoRA is a parameter-efficient convolutional low-rank adapter that adds local inductive biases to CLIP's visual encoder. The Dynamic Fusion Gateway (DFG) is a gating mechanism that uses visual context to adaptively modulate text prompts. The load-bearing idea is that these two modifications are synergistic: the improved local representation makes the learned fusion more effective, and the fusion in turn makes the representation better suited to anomaly detection.

What would settle it

Take one industrial and one medical anomaly benchmark, train CLIP with Conv-LoRA only, DFG only, and both, using identical data splits and hyperparameters; if the joint model's accuracy is no better than the better single component, or the gain is additive rather than interactive, the paper's central co-design claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the adaptation gap in zero-shot anomaly detection is best closed by architectural co-design rather than by improving either representation or fusion alone. On the representation side, Conv-LoRA injects local inductive biases into CLIP's visual features, which the paper identifies as the missing ingredient for fine-grained dense prediction. On the fusion side, DFG replaces a one-way or static fusion with a bidirectional one: visual context adaptively modulates the text prompts that guide detection. The paper reports extensive experiments on diverse industrial and medical benchmarks showing superior accuracy and robustness, and states that this synergistic co-

Load-bearing premise

The load-bearing premise is that the benchmark gains were produced under a genuine zero-shot protocol, with the adapter and fusion gate trained without exposure to the target anomaly classes and without benchmark-specific hyperparameter tuning.

Editorial extensions

If this is right

If the claim holds, zero-shot anomaly detection on unseen industrial and medical categories becomes practical with CLIP-scale foundation models and only a small number of trainable parameters.
The co-design principle becomes a template for other dense perception tasks, such as segmentation or defect localization, where CLIP-like models currently underperform.
Bidirectional visual-to-text modulation could replace the standard static prompt-template fusion in downstream vision-language applications.
The parameter efficiency of Conv-LoRA means the method can be applied in compute-limited settings where full fine-tuning of a vision-language model is infeasible.

Reading between the lines

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

Caveat: the full text supplied after the abstract is a different manuscript (on LLM expert annotators), so this summary rests on the abstract alone; the claims should be re-checked against the actual ACD-CLIP full text.
The paper's 'synergy is critical' claim implies a falsifiable interaction effect: removing either Conv-LoRA or DFG should hurt performance more than removing a comparable single component, and an additive ablation pattern would undercut the co-design thesis.
The reported 'zero-shot' label is only meaningful if the adapter and fusion gate were not tuned on the target benchmark classes; a version evaluated with truly held-out classes and fixed hyperparameters would sharpen the claim.

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

Colleague,

You need to know two things. First, the actual manuscript we got is not the paper: the full text attached is a COLM paper on LLM expert annotators, arXiv:2508.07827, not the ACD-CLIP paper described in the header. So any review is abstract-only. Second, the abstract itself is a competent methods pitch: it names a concrete problem (CLIP lacks local inductive biases and inflexible fusion for dense anomaly detection), proposes two specific components (Conv-LoRA adapter and Dynamic Fusion Gateway), and promises code.

What's new: the combination of a convolutional low-rank adapter and a visual-conditioned fusion gate for zero-shot anomaly detection is plausibly new. Each ingredient is familiar from the VLM adaptation literature, and the abstract doesn't cite related work, so I can't judge the novelty precisely. But the pitch is coherent.

Soft spots: There are no numbers, no baselines, no ablations, no protocol details. The phrase 'superior accuracy and robustness' is a claim, not evidence. The load-bearing risk is the zero-shot protocol. If the Conv-LoRA adapter or the fusion gate is trained on normal images from the target benchmarks, or if the hyperparameters (rank, gate scaling) are tuned on the target validation sets, then this is supervised adaptation, not zero-shot. The abstract doesn't state the training source or hyperparameter selection. Also, 'co-design is critical' requires ablations of each component alone and combined; none are visible. These are not flaws in the abstract itself—an abstract can't carry all that—but they are exactly what a referee needs to check.

I don't see internal inconsistency in the abstract. The claims are empirical, not circular. But without the actual paper, I can't vote on soundness. The mismatch in the full text is a pipeline problem, I assume, but it means the record under review is just one page.

Who this is for: someone working in zero-shot anomaly detection with CLIP-style backbones might find the idea worth a glance, and the code link helps. I wouldn't cite it until the real paper is available and I can check the protocol.

Recommendation: if a coherent full paper exists and matches the abstract, it deserves a serious referee—the question is empirical and the method is well-specified enough to test. Desk-rejecting a concrete proposal in an active subfield because the abstract lacks detail would be too harsh; desk-rejecting because the submitted text is a different paper is fair if the authors can't fix it. My take: ask for the correct manuscript first, then send it to review.

Referee Report

3 major / 4 minor

Summary. The submission claims to introduce ACD-CLIP, an architectural co-design for zero-shot anomaly detection (ZSAD). Based on the abstract, the method combines a parameter-efficient Convolutional Low-Rank Adaptation (Conv-LoRA) adapter, intended to inject local inductive biases into CLIP's representation, with a Dynamic Fusion Gateway (DFG) that modulates text prompts from visual context, and the authors assert that this co-design yields superior accuracy and robustness on industrial and medical benchmarks. The full text supplied with the submission, however, is an unrelated paper, arXiv:2508.07827, 'Evaluating Large Language Models as Expert Annotators,' which studies LLM annotation in finance, law, and biomedicine. The record therefore contains no methods, experiments, ablations, tables, or related-work discussion for ACD-CLIP; the scientific content of the claimed paper is entirely absent.

Significance. If substantiated, the ACD-CLIP proposal could be relevant to zero-shot anomaly detection, where CLIP's limited local inductive biases and rigid fusion mechanisms are recognized obstacles. The abstract makes concrete, potentially testable commitments: an open-source code release, a parameter-efficient adapter, and a dynamic fusion mechanism. However, because the submitted record contains none of the supporting material, the contribution cannot currently be assessed. The significance is thus conditional and, on this record, unverifiable.

major comments (3)

[Full text (arXiv:2508.07827)] The full text is a completely different paper: 'Evaluating Large Language Models as Expert Annotators.' Its methods, datasets, results, and references concern LLM annotation performance on finance, law, and biomedicine tasks, not zero-shot anomaly detection, CLIP, Conv-LoRA, or the Dynamic Fusion Gateway. None of the abstract's substantive claims about ACD-CLIP are supported by any methods, equations, experimental protocol, tables, or ablations in this record. This is a load-bearing absence, not a presentation issue.
[Abstract] The abstract asserts 'superior accuracy and robustness' on 'diverse industrial and medical benchmarks' and states that experiments validate that 'synergistic co-design is critical,' but it reports no numerical results, no benchmark names, no baseline list, and no ablations. The causal claim about co-design being critical requires at least component ablations (e.g., Conv-LoRA-only, DFG-only, and combined); none are present in the record. Consequently, both the accuracy claim and the attribution claim are unverifiable.
[Abstract — zero-shot protocol] The zero-shot claim is load-bearing and currently unspecified. The abstract does not state what data were used to train or configure the Conv-LoRA adapter and the DFG, whether target-benchmark anomaly classes or target-domain normal images were seen during training or validation, or how hyperparameters (e.g., adapter rank, gate settings) were selected. If target-domain data or target validation sets informed tuning, the result would not be a genuine zero-shot evaluation. The manuscript must specify the training protocol and hyperparameter selection; this is not a minor missing detail.

minor comments (4)

[Abstract] The phrase 'superior accuracy and robustness' should be replaced with quantitative comparisons against specific baselines, including confidence intervals or significance tests where appropriate.
[Abstract] The 'diverse industrial and medical benchmarks' should be named explicitly, along with the evaluation metrics (e.g., AUROC, AP) and the anomaly-class definition per dataset.
[Code availability] The source-code URL is given, but the manuscript does not specify the code's license, dependencies, or whether a reproducibility statement was verified. If the code is publicly available, please include a clear reproducibility appendix.
[References] The full-text reference list is unrelated to the ACD-CLIP abstract. Any corrected submission must supply appropriate citations for LoRA/Conv-LoRA, zero-shot anomaly detection benchmarks, and prior CLIP adaptation methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the ACD-CLIP abstract is an empirical architecture proposal with no derivation-by-construction, no fitted-parameter-turned-prediction, and no load-bearing self-citation. The full-text mismatch is an evidence-availability problem, not circularity.

full rationale

The supplied material consists only of the ACD-CLIP abstract; the full text attached is arXiv:2508.07827, an unrelated COLM paper on LLM expert annotators. Under the circularity rubric, I can only flag a step as circular when the paper itself exhibits a specific reduction: an equation defined in terms of the target, a fitted parameter renamed as a prediction, or a central premise justified solely by a self-citation. No such step exists in the abstract. The abstract proposes two components (Conv-LoRA and Dynamic Fusion Gateway) and claims empirical superiority and robustness on industrial and medical benchmarks. This is an empirical claim, not a derivation from first principles. There are no equations, no fitted values, no uniqueness theorems, and no self-citations in the abstract. The statement that 'synergistic co-design is critical' is a causal claim that would normally be supported by ablations; ablation-based support on the same benchmarks is standard practice and, while internal rather than external evidence, it is not circular in the sense defined here. The zero-shot protocol and hyperparameter choices are underspecified in the abstract, which is a transparency and verifiability concern. In particular, if the adapter or gate were tuned on target-benchmark validation sets, the 'zero-shot' label would be misleading, but that would be a factual/protocol failure, not a circular derivation. The reviewer rule asks that limitations and missing support be flagged explicitly; I flag the full-text mismatch and the absence of methods/ablations in the record as a serious evidence-supply limitation. However, circularity scoring is reserved for arguments that reduce to their own inputs by construction, and no such reduction is quotable from the ACD-CLIP abstract. Therefore the honest finding is no significant circularity, score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 2 invented entities

The building blocks (LoRA, convolutions, text-prompt modulation) come from prior literature, so they are not re-invented here. What the paper adds are two new modules and a co-design claim; since the body is unavailable, none of their hyperparameters or evaluation constraints can be audited. No physical entities are postulated.

free parameters (2)

Conv-LoRA rank and scaling hyperparameters
Abstract reports a parameter-efficient adapter but gives no rank, alpha, or layer placement; reported performance plausibly depends on these choices, and such choices are usually tuned on validation data.
DFG gating hyperparameters
Modulation strength, temperature, or thresholds of the dynamic fusion gateway are undisclosed in the abstract; benchmark results may depend on them and on per-domain settings.

assumptions (3)

domain assumption CLIP's adaptation gap for dense prediction is caused by missing local inductive bias and inflexible fusion.
This diagnostic premise motivates both modules in the abstract; if the true bottleneck is elsewhere, the specific design is not justified, though it could still work empirically.
domain assumption The evaluation is genuinely zero-shot: adapters and fusion gate do not see target benchmark anomaly classes during training.
The abstract calls the setting 'Zero-Shot Anomaly Detection'; maintaining that protocol end to end is load-bearing for the headline claim and is unverifiable from the abstract.
domain assumption Reported benchmark gains transfer across industrial and medical domains without per-domain retuning.
The abstract claims accuracy and robustness on 'diverse industrial and medical benchmarks'; whether the same hyperparameters and settings were used across domains is not stated.

invented entities (2)

Conv-LoRA adapter
purpose: Inject convolutional low-rank parameters into CLIP layers to add local inductive bias for dense prediction.
The only evidence for this module's effectiveness is the paper's own benchmark numbers, which are not present in this record; no external falsifiable handle is provided.
Dynamic Fusion Gateway (DFG)
purpose: Use visual context to adaptively modulate text prompts for bidirectional image-text fusion.
Same situation as Conv-LoRA: a new architectural component whose evidence is internal to the paper; listed here so the ledger shows what the paper introduces.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection." pith.science (2026). https://pith.science/paper/HUKZT7TP

@misc{pith2026250807819,
  author       = {Pith},
  title        = {Pith review of: ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HUKZT7TP}},
  note         = {Machine review of arXiv:2508.07819}
}

read the original abstract

Pre-trained Vision-Language Models (VLMs) struggle with Zero-Shot Anomaly Detection (ZSAD) due to a critical adaptation gap: they lack the local inductive biases required for dense prediction and employ inflexible feature fusion paradigms. We address these limitations through an Architectural Co-Design framework that jointly refines feature representation and cross-modal fusion. Our method proposes a parameter-efficient Convolutional Low-Rank Adaptation (Conv-LoRA) adapter to inject local inductive biases for fine-grained representation, and introduces a Dynamic Fusion Gateway (DFG) that leverages visual context to adaptively modulate text prompts, enabling a powerful bidirectional fusion. Extensive experiments on diverse industrial and medical benchmarks demonstrate superior accuracy and robustness, validating that this synergistic co-design is critical for robustly adapting foundation models to dense perception tasks. The source code is available at https://github.com/cockmake/ACD-CLIP.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Deep Learning-Based Desikan-Killiany Parcellation of the Brain Using Diffusion MRI
eess.IV 2025-08 unverdicted novelty 5.0 of 10

A two-stage deep network parcellates the brain into Desikan-Killiany regions directly from diffusion MRI maps, reportedly beating registration-based methods on two public datasets.

Reference graph

Works this paper leans on

26 extracted references · 9 canonical work pages · cited by 1 Pith paper

[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,

arXiv
[4]

Must read: A systematic survey of computational persuasion

Nimet Beyza Bozdag, Shuhaib Mehri, Xiaocheng Yang, Hyeonjeong Ha, Zirui Cheng, Esin Durmus, Jiaxuan You, Heng Ji, Gokhan Tur, and Dilek Hakkani-T ¨ur. Must read: A systematic survey of computational persuasion. arXiv preprint arXiv:2505.07775,

arXiv
[5]

Can GPT models be Financial Analysts? An Evaluation of ChatGPT and GPT-4 on mock CFA Exams

Ethan Callanan, Amarachi Mbakwe, Antony Papadimitriou, Yulong Pei, Mathieu Sibue, Xiaodan Zhu, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. Can gpt models be finan- cial analysts? an evaluation of chatgpt and gpt-4 on mock cfa exams. arXiv preprint arXiv:2310.08678,

work page Pith review arXiv
[6]

Reconcile: Round-table conference improves reasoning via consensus among diverse llms

Justin Chih-Yao Chen, Swarnadeep Saha, and Mohit Bansal. Reconcile: Round-table conference improves reasoning via consensus among diverse llms. arXiv preprint arXiv:2309.13007,

arXiv
[7]

Evalu- ating large language models trained on code

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evalu- ating large language models trained on code. arXiv preprint arXiv:2107.03374,

arXiv
[9]

Improving factuality and reasoning in language models through multiagent debate

Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325,

arXiv
[10]

Measuring the persuasiveness of language models,

10 Published as a conference paper at COLM 2025 Esin Durmus, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. Measuring the persuasiveness of language models,

work page 2025
[12]

Large language model based multi-agents: A survey of progress and challenges

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680,

arXiv

Show all 26 references

[13]

Annollm: Making large language models to be better crowdsourced annotators

Xingwei He, Zhenghao Lin, Yeyun Gong, Alex Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, Weizhu Chen, et al. Annollm: Making large language models to be better crowdsourced annotators. arXiv preprint arXiv:2303.16854,

arXiv
[15]

Coda-19: Using a non-expert crowd to annotate research aspects on 10,000+ abstracts in the covid-19 open research dataset

Ting-Hao Kenneth Huang, Chieh-Yang Huang, Chien-Kuang Cornelia Ding, Yen-Chia Hsu, and C Lee Giles. Coda-19: Using a non-expert crowd to annotate research aspects on 10,000+ abstracts in the covid-19 open research dataset. In ACL 2020 Workshop on Natural Language Processing fo...

2020
[16]

Pubmedqa: A dataset for biomedical research question answering

Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natur...

2019
[18]

Self-refine: Iterative refinement with self-feedback

11 Published as a conference paper at COLM 2025 Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Informa...

2025
[19]

URL https://openai.com/index/openai-o3-mini/. Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean- baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding ...

arXiv
[20]

Gpqa: A graduate-level google- proof q&a benchmark

David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google- proof q&a benchmark. arXiv preprint arXiv:2311.12022,

arXiv
[21]

Large language models encode clinical knowledge

Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge. Nature, 620(7972):172–180, 2023a. Karan Singhal, Tao Tu, Juraj Gottweis, ...

arXiv
[22]

Are expert-level language models expert-level annotators? arXiv preprint arXiv:2410.03254, 2024a

Yu-Min Tseng, Wei-Lin Chen, Chung-Chi Chen, and Hsin-Hsi Chen. Are expert-level language models expert-level annotators? arXiv preprint arXiv:2410.03254, 2024a. Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Yu-Ching Hsu, Jia-Yin Foo, Chao-Wei Huang, and Yun-Nung Chen. Two tales...

2024 arXiv
[23]

Llmaaa: Making large language models as active annotators

12 Published as a conference paper at COLM 2025 Ruoyu Zhang, Yanzeng Li, Yongliang Ma, Ming Zhou, and Lei Zou. Llmaaa: Making large language models as active annotators. arXiv preprint arXiv:2310.19596,

2025 arXiv
[24]

Can chatgpt reproduce human-generated labels? a study of social computing tasks

Yiming Zhu, Peixian Zhang, Ehsan-Ul Haq, Pan Hui, and Gareth Tyson. Can chatgpt reproduce human-generated labels? a study of social computing tasks. arXiv preprint arXiv:2304.10145,

arXiv
[25]

In this task, annotators are tasked to label each segment as background, purpose, method, finding/contribution, or other sections

codes each segment aspect of English abstracts in the COVID-19 Open Research Dataset (Wang et al., 2020). In this task, annotators are tasked to label each segment as background, purpose, method, finding/contribution, or other sections. To ensure the quality of the labels, we ...

2020
[26]

15 Published as a conference paper at COLM 2025 Figure 6: The annotation guideline of FOMC dataset

13 Published as a conference paper at COLM 2025 �� ...

2025
[2020]

Cuad: An expert-annotated nlp dataset for legal contract review

Dan Hendrycks, Collin Burns, Anya Chen, and Spencer Ball. Cuad: An expert-annotated nlp dataset for legal contract review. arXiv preprint arXiv:2103.06268,

arXiv
[2021]

Gpts are multilingual annotators for sequence generation tasks

Juhwan Choi, Eunju Lee, Kyohoon Jin, and YoungBin Kim. Gpts are multilingual annotators for sequence generation tasks. arXiv preprint arXiv:2402.05512,

arXiv
[2022]

Encouraging divergent thinking in large language models through multi-agent debate

Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Zhaopeng Tu, and Shuming Shi. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118,

arXiv
[2023]

Open-source large language models out- perform crowd workers and approach chatgpt in text-annotation tasks

Meysam Alizadeh, Ma ¨el Kubli, Zeynab Samei, Shirin Dehghani, Juan Diego Bermeo, Maria Korobeynikova, and Fabrizio Gilardi. Open-source large language models out- perform crowd workers and approach chatgpt in text-annotation tasks. arXiv preprint arXiv:2307.02179,

arXiv
[2024]

Neel Guha, Julian Nyarko, Daniel Ho, Christopher R´e, Adam Chilton, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, et al

URL https:// blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/ #ceo-message. Neel Guha, Julian Nyarko, Daniel Ho, Christopher R´e, Adam Chilton, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, et al. Legalbench: A ...

2024
[2025]

Parikshit Bansal and Amit Sharma

URL https://www.anthropic.com/ news/claude-3-7-sonnet . Parikshit Bansal and Amit Sharma. Large language models as annotators: Enhancing generalization of nlp models at minimal cost. arXiv preprint arXiv:2306.15766,

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[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,

arXiv

[2] [4]

Must read: A systematic survey of computational persuasion

Nimet Beyza Bozdag, Shuhaib Mehri, Xiaocheng Yang, Hyeonjeong Ha, Zirui Cheng, Esin Durmus, Jiaxuan You, Heng Ji, Gokhan Tur, and Dilek Hakkani-T ¨ur. Must read: A systematic survey of computational persuasion. arXiv preprint arXiv:2505.07775,

arXiv

[3] [5]

Can GPT models be Financial Analysts? An Evaluation of ChatGPT and GPT-4 on mock CFA Exams

Ethan Callanan, Amarachi Mbakwe, Antony Papadimitriou, Yulong Pei, Mathieu Sibue, Xiaodan Zhu, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. Can gpt models be finan- cial analysts? an evaluation of chatgpt and gpt-4 on mock cfa exams. arXiv preprint arXiv:2310.08678,

work page Pith review arXiv

[4] [6]

Reconcile: Round-table conference improves reasoning via consensus among diverse llms

Justin Chih-Yao Chen, Swarnadeep Saha, and Mohit Bansal. Reconcile: Round-table conference improves reasoning via consensus among diverse llms. arXiv preprint arXiv:2309.13007,

arXiv

[5] [7]

Evalu- ating large language models trained on code

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evalu- ating large language models trained on code. arXiv preprint arXiv:2107.03374,

arXiv

[6] [9]

Improving factuality and reasoning in language models through multiagent debate

Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325,

arXiv

[7] [10]

Measuring the persuasiveness of language models,

10 Published as a conference paper at COLM 2025 Esin Durmus, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. Measuring the persuasiveness of language models,

work page 2025

[8] [12]

Large language model based multi-agents: A survey of progress and challenges

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680,

arXiv