Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

CoTGuard: Using Chain-of-Thought Triggering for Copyright Protection in Multi-Agent LLM Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read CoTGuard claims that by planting a task-specific trigger phrase in each agent's prompt, it can detect when multi-agent LLM systems reproduce copyrighted content in their intermediate reasoning traces, catching leaks that final-output…

desk verdict A trigger-pattern detector dressed as copyright protection; the evaluation never touches copyrighted content, so the central claim is untested. read the letter →

arxiv 2505.19405 v1 pith:HRO5ZL3W submitted 2025-05-26 cs.CL cs.CR

classification cs.CLcs.CR
keywords chain-of-thoughtmulti-agentLLMsystemscopyrightprotectiontrigger-basedwatermarkingreasoningtracedetectionleakagepromptinjectionagents
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

This paper sets out to show that copyright leakage in multi-agent LLM systems can be caught during the reasoning process rather than only in final outputs. The proposed method, CoTGuard, plants a task-specific trigger phrase into each agent's prompt and then inspects the chain-of-thought traces—the step-by-step reasoning records agents exchange—looking for residues of that trigger. On math, logic, and planning benchmarks, the authors report that this approach detects leaked content at rates up to 95.7 percent while keeping task accuracy close to the unprotected baseline. The upshot, if the method holds, is a practical and interpretable audit tool for collaborative agent workflows.

What carries the argument

The central mechanism is the trigger pattern $\tau$: a verbose, task-specific instruction (for example, "explain like a patient teacher") that is appended to an agent's prompt. The pattern is designed to blend into the reasoning style so that it reappears in the chain-of-thought steps and persists when those steps are shared with other agents. Detection rests on a three-stage pipeline—parsing the candidate trace into spans, scoring each span against known trigger patterns with similarity metrics such as embedding-based cosine similarity, and aggregating the scores into a single leakage score $\delta$—which converts the prompt-level bias into a quantifiable watermark.

What would settle it

Run the paper's detector on triggered traces whose intermediate reasoning steps have been paraphrased into a neutral, style-free register—no trigger metaphor, no key phrase, no distinctive phrasing—while keeping the final answers identical. If the leakage score stays above the detection threshold, the watermark is robust to paraphrase; if it falls to the level of untriggered traces, the method's central claim that the signature persists through reasoning fails.

Watch

Extended reading notes

Core claim

The central claim is that injecting a deterministic trigger pattern into the prompt of a multi-agent reasoning system creates a persistent stylistic and semantic signature in the intermediate chain-of-thought traces, and that this signature survives when traces are passed from one agent to the next. The paper formalizes this with a trigger key $k$ and task type $t$ mapped by a deterministic function $T(k,t)$ to a verbose trigger pattern $\tau$, which is appended to the task instruction. A detection function $D$ then parses a candidate trace, computes similarity scores against a repository of known trigger patterns using lexical, structural, or embedding-based metrics, and aggregates them into a leakage score $\delta$; if $\delta$ exceeds a threshold, the trace is flagged as exhibiting unauthorized reuse of trigger-guided reasoning. The paper evaluates this scheme across three model families and seven benchmarks spanning mathematics, logic, and planning, reporting high leak-detection rates with minimal accuracy degradation relative to an output-perturbation baseline.

Load-bearing premise

The load-bearing premise is that every agent in the chain will faithfully carry the trigger pattern into its own reasoning steps, so the watermark survives as the trace is shared, summarized, and restated; the paper's own adaptive-attack results show that when an attacker rewrites the prompt to break the chain-of-thought pattern, the detection rate drops from over 85 percent to roughly 68–78 percent.

Editorial extensions

If this is right

  • Copyright monitors for multi-agent systems can shift from checking final outputs to auditing intermediate reasoning traces, catching partial or paraphrased leakage that appears only during collaboration.
  • Operators who control prompt construction can mark their own agent workflows, then later check external traces against their trigger repository to detect unauthorized reuse of their reasoning patterns.
  • The task-specific trigger design means the same watermarking scheme transfers across mathematics, logic, and planning tasks without modifying the underlying model.
  • Detection requires only externally observed reasoning logs and a repository of known triggers, so it works in black-box deployments where no model internals are accessible.
  • The paper's adaptive-attack results delimit the protection: it is robust to output-level post-processing but degrades when an attacker deliberately rewrites the prompt to break the chain-of-thought pattern.

Reading between the lines

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

  • A consequence the authors do not draw: if trigger keys are reused across tasks or deployments, an adversary who learns one key could strip or replicate the watermark across an entire family of tasks; key diversification per deployment would be a natural hardening step.
  • Quantifying how much the trigger attenuates when agents summarize or abstract intermediate steps (rather than passing full traces) is a direct extension that would bound the method's real-world coverage.
  • The same triggering mechanism could serve as a provenance tool: because the watermark identifies the originating prompt style, it can attribute a generated reasoning chain back to the operator who injected the trigger, not just flag copyright violation.
  • Extending to multilingual or multimodal agents would require a trigger signal that survives translation or modality conversion, which the current embedding-based comparison does not yet demonstrate.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes CoTGuard, a framework that embeds task-specific trigger patterns into Chain-of-Thought (CoT) prompts of multi-agent LLM systems and then detects potential copyright leakage by scoring externally observed reasoning traces for similarity to a repository of known trigger patterns. The authors evaluate the method on math, logic, and planning benchmarks using GPT-3.5, GPT-4o, and Claude, reporting high Leakage Detection Rates (LDR) with minimal degradation of task accuracy, together with ablations and an adaptive-attack analysis.

Significance. The paper targets a timely and underexplored problem: monitoring intermediate reasoning in multi-agent LLM systems for content leakage. It has genuine strengths: a concrete pipeline with prompt injection and detection algorithms, experiments across three model families, multiple datasets, a code repository, and both ablation and adaptive-attack evaluations. However, the evaluation never uses copyrighted content, and the detection metric is computed as similarity to the very trigger patterns the system injected. The reported LDR therefore measures trigger-pattern presence, not unauthorized reproduction of protected material, so the central claim in the abstract and introduction is not supported by the evidence. The method may be a style-based watermark for CoT traces, but the paper does not establish that it detects copyright violations.

major comments (4)
  1. [§5.1, Tables 1–2; Algorithm 4] The evaluation contains no copyrighted material. All datasets (GSM8K, MATH, Omni-MATH, PrOntoQA, ContextHub, FOLIO, TravelPlanner) are public reasoning benchmarks. LDR is defined as the percentage of triggers successfully detecting leakage, and Algorithm 4 computes it by measuring embedding similarity between each reasoning step and the known trigger pattern set K. A high LDR thus indicates that the trace resembles the injected trigger phrase, not that it reproduces protected content. The headline numbers (e.g., LDR 95.7% on Omni-MATH) do not test the paper's claimed ability to detect copyright violations.
  2. [§5.1, Table 2 (Vanilla baseline)] The Vanilla baseline is described as the standard setting without any copyright protection or signal injection, yet it reports LDR values of roughly 50–64% across all models and tasks. With no trigger injected, a sensible detector should not label clean traces as leaking at rates near chance. This anomaly indicates that the LDR metric has no reliable ground-truth notion of leakage and undermines the interpretation of the CoTGuard LDR numbers throughout the paper.
  3. [§3.3, Theorem 1] Theorem 1 asserts the existence of a leakage detection function D and states that a leakage score above a threshold implies reuse of trigger-induced patterns. This is a restatement of the detection procedure rather than a theorem with a formal proof, and it identifies leakage with the presence of the authors' own injected triggers. Because the triggers are added by the system itself, detecting them is circular and does not establish detection of unauthorized reproduction of copyrighted content.
  4. [§5.5, Table 4] The adaptive attack 'Rewriting Prompt (Anti-CoT)' reduces LDR from 85–96% to 68–79% across tasks. If the traces still contained the same copyrighted material after the prompt rewrite, a copyright-leakage detector would be expected to continue flagging them. The observed drop shows that LDR tracks the stylistic trigger pattern rather than the presence of protected content, which is consistent with the circularity concern raised above.
minor comments (6)
  1. [§1] The sentence 'Building on the generalization and zero-shot capabilities of LLMs, researchers have developed LLM-based agent systems [23] that simulate autonomous agents capable of planning cite xie2024travelplanner, collaboration [27], and task execution [33]' contains the literal text 'cite xie2024travelplanner', which appears to be an unresolved citation command.
  2. [References, [56]] Reference [56] attributes the MATH benchmark to 'A. Zelikman et al.'; the MATH dataset was introduced by Hendrycks et al. (2021), not Zelikman et al.
  3. [References, [5]] Reference [5] is 'Extracting training data from diffusion models,' but the related-work text uses it to support a statement about LLMs memorizing training data; the cited work concerns diffusion models, not language models.
  4. [§5.2] The sentence 'For example, GPT-3.5 with CoTGuard achieves 90.1%' is incomplete because it does not specify the dataset or metric; Table 1 shows that 90.1% is the Ours column value for GSM8K, but the text should state this explicitly.
  5. [Figure 1] Figure 1 contains rendering artifacts such as '♂robotAgent' and 'chalkboard-◎eacher', which obscure the pipeline description and should be fixed in any revision.
  6. [Appendix A.5, Algorithm 4] Algorithm 4 states 'Normalize δ' without giving a formula or describing how normalization is performed; the paper also does not specify how the detection threshold θ is selected beyond saying it 'balances' sensitivity and false positives.

Circularity Check

2 steps flagged · score 8.0 of 10

CoTGuard's LDR measures the injected trigger's presence, not copyright leakage; the claimed copyright detection is self-definitional.

  1. self definitional [Section 4.3 'Trigger Detection Algorithm' and Algorithm 1; Appendix A.5, Algorithm 4]
    "The main goal of the detection phase is to determine whether a reasoning trace has been influenced by our trigger-based watermarking system. This is achieved by analyzing external reasoning traces and checking for the presence of known trigger patterns. ... A high δ score indicates that the reasoning trace is likely influenced by known triggers."

    The paper's advertised output is detection of 'unauthorized content reproduction,' but the detector's input is the repository K of trigger patterns that the authors themselves injected. Algorithm 1/4 returns a score δ by aggregating similarity between each reasoning step and K, and Appendix A.5 states that a high δ means the trace was 'influenced by known triggers.' Thus 'leakage' is defined as the presence of the injected marker; no copyrighted content is ever compared. The LDR numbers in Table 2 therefore measure the detector's ability to recognize its own trigger, making the central claim true by construction rather than by independent evidence.

  2. fitted input called prediction [Section 5.1 'Evaluation Metrics' and Table 2; Section 5.3]
    "Leakage Detection Rate (LDR): The percentage of triggers successfully detecting leakage. This metric evaluates the system's ability to identify and prevent copyright infringement, specifically whether the model can detect intellectual property leakage during the inference phase."

    The metric equates successful trigger detection with detected copyright leakage. Because Section 4.1 constructs every protected prompt as Instruction(p) + τ and Appendix B.1 states that detection is tested on triggered traces against K, a high LDR is the expected recall of a similarity search over traces that contain the searched-for patterns. The evaluation contains no copyrighted works—only GSM8K, MATH, Omni-MATH, PrOntoQA, ContextHub, FOLIO, and TravelPlanner—so the reported LDR cannot validate the copyright-protection claim independently of the trigger-injection construction. The adaptive-attack results in Section 5.5 confirm this reading: breaking the trigger pattern drops LDR even though any stolen content would still be stolen.

full rationale

The central derivation chain reduces to self-definition. The paper defines copyright leakage operationally as the presence of trigger patterns that the authors injected into agent prompts, and then reports the detector's recall on those same patterns as 'Leakage Detection Rate.' No experiment introduces copyrighted text, and the detector never compares reasoning traces against any copyrighted corpus. Consequently, the high LDR values are a direct consequence of the experimental design: the 'signal' being detected is the very signal the method places in the prompt. This is not a minor self-citation issue or a mere external-benchmark gap; it is a construct-level circularity in which the claimed phenomenon (copyright leakage) is replaced by the method's own watermark, and the watermark's presence is then presented as evidence of the phenomenon.

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

No new entities (particles, forces, conserved quantities, etc.) are introduced. The 'trigger pattern' is a prompt addition, not a postulated entity.

free parameters (3)
  • detection threshold θ = not reported
    Threshold that balances detection sensitivity and false positives; stated in Section 3.3 as tunable using clean and synthetic traces.
  • trigger key k and trigger pattern τ = e.g., 'patient teacher', 'reason like a cautious analyst'
    Manually generated per task; the paper says 'we manually generate a pool of trigger keys' (Appendix B.1). These are hand-chosen, not derived from data.
  • embedding model for similarity = Sentence-BERT
    Choice of sentence embeddings for cosine similarity; affects all detection scores (Appendix B.1), a design parameter not compared to alternatives.
assumptions (3)
  • domain assumption LLM agents will faithfully reproduce the injected trigger pattern in their chain-of-thought reasoning, even when intermediate steps are passed between agents.
    Stated in Section 4.2: 'the trigger pattern propagates through the reasoning chain' and 'maintaining its presence throughout the collaborative process.' If agents ignore or drop the trigger, detection fails.
  • domain assumption The detector has access to the full intermediate reasoning traces and to a repository of known trigger keys and task types.
    Stated in Section 3.3: 'We assume access to a repository of known trigger keys and their associated task types.' This is an access assumption that may not hold in real deployments.
  • ad hoc to paper Embedding-based cosine similarity between a trace and a trigger pattern is a valid indicator of leakage.
    The paper uses Sentence-BERT cosine similarity (Appendix B.1) without justifying that this metric separates trigger-influenced from clean traces. The Vanilla baseline's high LDR suggests the separation is poor.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoTGuard: Using Chain-of-Thought Triggering for Copyright Protection in Multi-Agent LLM Systems." pith.science (2026). https://pith.science/paper/HRO5ZL3W

@misc{pith2026250519405,
  author       = {Pith},
  title        = {Pith review of: CoTGuard: Using Chain-of-Thought Triggering for Copyright Protection in Multi-Agent LLM Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRO5ZL3W}},
  note         = {Machine review of arXiv:2505.19405}
}
read the original abstract

As large language models (LLMs) evolve into autonomous agents capable of collaborative reasoning and task execution, multi-agent LLM systems have emerged as a powerful paradigm for solving complex problems. However, these systems pose new challenges for copyright protection, particularly when sensitive or copyrighted content is inadvertently recalled through inter-agent communication and reasoning. Existing protection techniques primarily focus on detecting content in final outputs, overlooking the richer, more revealing reasoning processes within the agents themselves. In this paper, we introduce CoTGuard, a novel framework for copyright protection that leverages trigger-based detection within Chain-of-Thought (CoT) reasoning. Specifically, we can activate specific CoT segments and monitor intermediate reasoning steps for unauthorized content reproduction by embedding specific trigger queries into agent prompts. This approach enables fine-grained, interpretable detection of copyright violations in collaborative agent scenarios. We evaluate CoTGuard on various benchmarks in extensive experiments and show that it effectively uncovers content leakage with minimal interference to task performance. Our findings suggest that reasoning-level monitoring offers a promising direction for safeguarding intellectual property in LLM-based agent systems.

Figures

Figures reproduced from arXiv: 2505.19405 by the authors.

Figure 1
Figure 1. General pipeline of the CoTGuard. The process starts by enhancing the original prompt [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

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

  1. Towards Provable (In)Secure Model Weight Release Schemes

    cs.CR 2025-06 conditional novelty 6.0 of 10

    Defines game-based security for weight release schemes and breaks TaylorMLP with a near-complete, low-cost parameter extraction attack.

Reference graph

Works this paper leans on

63 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

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

  2. [2]

    Claude: A family of language models

    Anthropic. Claude: A family of language models. 2025

  3. [3]

    Bender, Timnit Gebru, Alexis McMillan-Major, and Margaret Shmitchell

    Emily M. Bender, Timnit Gebru, Alexis McMillan-Major, and Margaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021

  4. [4]

    Quantifying memorization across neural language models

    Nicholas Carlini, Kyle Lee, Florian Tramer, Eric Wallace, Matthew Jagielski, Abhinav Jagan- natha, Dawn Song, and Ulfar Erlingsson. Quantifying memorization across neural language models. In IEEE Symposium on Security and Privacy, 2022

  5. [5]

    Extracting training data from diffusion models

    Nicholas Carlini, Askhat Triastcyn, Matthew Jagielski, Florian Tramer, Eric Wallace, Abhinav Jagannatha, Dawn Song, and Ulfar Erlingsson. Extracting training data from diffusion models. arXiv preprint arXiv:2305.15269, 2023

  6. [6]

    How to protect copyright data in optimization of large language models? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17871–17879, 2024

    Timothy Chu, Zhao Song, and Chiwun Yang. How to protect copyright data in optimization of large language models? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17871–17879, 2024

  7. [7]

    Cobbe et al

    K. Cobbe et al. Gsm8k: A large-scale dataset for math word problems. In Proceedings of the 2021 International Conference on Machine Learning (ICML), 2021

  8. [8]

    Multi-agent systems: A survey

    Ali Dorri, Salil S Kanhere, and Raja Jurdak. Multi-agent systems: A survey. Ieee Access, 6:28573–28593, 2018

Show all 63 references
  1. [9]

    Improving multi-agent collaboration with chain-of-thought reasoning

    Yujia Du, Ximing Liu, Yujun Bai, Yitao Liang, and Xiang Ren. Improving multi-agent collaboration with chain-of-thought reasoning. arXiv preprint arXiv:2305.14325, 2023

  2. [10]

    Omni-math: A universal olympiad level mathematic benchmark for large language models

    Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, Zhengyang Tang, Benyou Wang, Daoguang Zan, Shanghaoran Quan, Ge Zhang, Lei Sha, Yichang Zhang, Xuancheng Ren, Tianyu Liu, and Baobao Chang. Omni-math: A universal ...

  3. [11]

    Auto-gpt: An autonomous gpt-4 experiment, 2023

    Significant Gravitas. Auto-gpt: An autonomous gpt-4 experiment, 2023

  4. [12]

    Multi-agent deep reinforcement learning: a survey

    Sven Gronauer and Klaus Diepold. Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55(2):895–943, 2022

  5. [13]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

  6. [14]

    Domain watermark: Effective and harmless dataset copyright protection is closed at hand

    Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36:54421–54450, 2023

  7. [15]

    Coda: Copyright detection in artificial intelligence-generated content via natural tracing

    Ruiqi Guo, Xudong Wang, Haotian Xu, Hongxia Jin, Yuhong Li, and Huayi Xu. Coda: Copyright detection in artificial intelligence-generated content via natural tracing. arXiv preprint arXiv:2305.18829, 2023

  8. [16]

    Stealthy wa- termarking of text generation via multi-token encoding

    Simeng He, Wayne Zhao, Zhiyuan Lin, Zhou Yu, and William Yang Wang. Stealthy wa- termarking of text generation via multi-token encoding. arXiv preprint arXiv:2306.04636, 2023. 10

  9. [17]

    Multi-modal and multi-agent systems meet rationality: A survey

    Bowen Jiang, Yangxinyu Xie, Xiaomeng Wang, Weijie J Su, Camillo Jose Taylor, and Tanwi Mallick. Multi-modal and multi-agent systems meet rationality: A survey. In ICML 2024 Workshop on LLMs and Cognition, 2024

  10. [18]

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative language models

    Zexuan Jiang, Deming Ye, Yilun Xu, Jindong Wang, Peng Liu, and Minlie Zhang. Selfcheckgpt: Zero-resource black-box hallucination detection for generative language models. arXiv preprint arXiv:2301.05228, 2024

  11. [19]

    A watermark for large language models

    Julian Kirchenbauer, Jonas Geiping, Henrik Bauermeister, Micah Goldblum, and Tom Goldstein. A watermark for large language models. arXiv preprint arXiv:2301.10226, 2023

  12. [20]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022

  13. [21]

    Double-i watermark: Protecting model copyright for llm fine-tuning

    Shen Li, Liuyi Yao, Jinyang Gao, Lan Zhang, and Yaliang Li. Double-i watermark: Protecting model copyright for llm fine-tuning. arXiv preprint arXiv:2402.14883, 2024

  14. [22]

    Camel: Communicative agents for "mind" exploration

    Tiansi Li, Yuxuan Zhang, Yuxuan Liu, Yujia Zhang, Yujie Liu, Wayne Xin Zhao, and Ji-Rong Wen. Camel: Communicative agents for "mind" exploration. arXiv preprint arXiv:2303.17760, 2023

  15. [23]

    A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth, 1(1):9, 2024

  16. [24]

    Coordination of networked nonlinear multi-agents using a high-order fully actuated predictive control strategy

    Guo-Ping Liu. Coordination of networked nonlinear multi-agents using a high-order fully actuated predictive control strategy. IEEE/CAA Journal of Automatica Sinica, 9(4):615–623, 2022

  17. [25]

    Liu et al

    L. Liu et al. Prontoqa: A dataset for logic-based question answering. In Proceedings of the 2021 Conference on Artificial Intelligence (AAAI), 2021

  18. [26]

    Shield: Evaluation and defense strategies for copyright compliance in llm text generation

    Xiaoze Liu, Ting Sun, Tianyang Xu, Feijie Wu, Cunxiang Wang, Xiaoqian Wang, and Jing Gao. Shield: Evaluation and defense strategies for copyright compliance in llm text generation. arXiv preprint arXiv:2406.12975, 2024

  19. [27]

    Dynamic llm-agent net- work: An llm-agent collaboration framework with agent team optimization

    Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. Dynamic llm-agent net- work: An llm-agent collaboration framework with agent team optimization. arXiv preprint arXiv:2310.02170, 2023

  20. [28]

    A dynamic llm-powered agent network for task-oriented agent collaboration

    Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. A dynamic llm-powered agent network for task-oriented agent collaboration. In First Conference on Language Modeling, 2024

  21. [29]

    Babyagi, 2023

    Yohei Nakajima. Babyagi, 2023

  22. [30]

    Codegen: An open large language model for code with multi-turn program synthesis

    Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint arXiv:2203.13474, 2022

  23. [31]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. 2023

  24. [32]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Pro- ceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1–15, 2023

  25. [33]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, Michael S Bernstein, et al. Generative agents: Interactive simulacra of human behavior. Org (2023, April 7) https://arxiv. org/abs/2304.03442 v2, 2023

  26. [34]

    Reasoning with large language models, a survey

    Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 11

  27. [35]

    Decentralized and centralized planning for multi-robot additive manufacturing

    Laxmi Poudel, Saivipulteja Elagandula, Wenchao Zhou, and Zhenghui Sha. Decentralized and centralized planning for multi-robot additive manufacturing. Journal of Mechanical Design, 145(1):012003, 2023

  28. [36]

    ChatDev: Communicative agents for software development

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. ChatDev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Associa...

  29. [37]

    Sentence-bert: Sentence embeddings using siamese bert- networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 3982–3992, Hong Kong, China, 2019. Association for Computational Lin- guistics

  30. [38]

    Copyright protection in generative ai: A technical perspective

    Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Pei Huang, Lingjuan Lyu, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024

  31. [39]

    Trust but verify: A simple method for detecting hallucinations in large language models

    Shuo Shen, Wenhao Ruan, Chen Liu, Mo Yu, Yansong Gao, Kai-Wei Chang, and Xiang Ren. Trust but verify: A simple method for detecting hallucinations in large language models. arXiv preprint arXiv:2303.16549, 2023

  32. [40]

    Privacy risks of general-purpose language models

    Congzheng Song and Vitaly Shmatikov. Privacy risks of general-purpose language models. In Proceedings of the 2020 IEEE Symposium on Security and Privacy, 2020

  33. [41]

    To cot or not to cot? chain- of-thought helps mainly on math and symbolic reasoning

    Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett. To cot or not to cot? chain- of-thought helps mainly on math and symbolic reasoning. arXiv preprint arXiv:2409.12183, 2024

  34. [42]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

  35. [43]

    Survey of containment control in multi-agent systems: concepts, communication, dynamics, and controller design

    Mourya Thummalapeta and Yen-Chen Liu. Survey of containment control in multi-agent systems: concepts, communication, dynamics, and controller design. International Journal of Systems Science, 54(14):2809–2835, 2023

  36. [44]

    Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models

    Baolin Wang, Xiaoxue Liu, Qixuan Zeng, Xinyu Li, and Minlie Huang. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. arXiv preprint arXiv:2305.04091, 2023

  37. [45]

    Espew: Robust copyright protec- tion for llm-based eaas via embedding-specific watermark

    Zongqi Wang, Baoyuan Wu, Jingyuan Deng, and Yujiu Yang. Espew: Robust copyright protec- tion for llm-based eaas via embedding-specific watermark. arXiv preprint arXiv:2410.17552, 2024

  38. [46]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

  39. [47]

    Badchain: Backdoor chain-of-thought prompting for large language models

    Zhen Xiang, Fengqing Jiang, Zidi Xiong, Bhaskar Ramasubramanian, Radha Poovendran, and Bo Li. Badchain: Backdoor chain-of-thought prompting for large language models. arXiv preprint arXiv:2401.12242, 2024

  40. [48]

    Travelplanner: A benchmark for real-world planning with language agents

    Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. Travelplanner: A benchmark for real-world planning with language agents. arXiv preprint arXiv:2402.01622, 2024

  41. [49]

    Adversarial behavior in multi-agent systems: Challenges and approaches

    Hao Xu, Shuo Li, and Tianyu Wang. Adversarial behavior in multi-agent systems: Challenges and approaches. IEEE Transactions on Autonomous Systems, 2024. 12

  42. [50]

    Can large vision-language models detect images copyright infringement from genai? arXiv preprint arXiv:2502.16618, 2025

    Qipan Xu, Zhenting Wang, Xiaoxiao He, Ligong Han, and Ruixiang Tang. Can large vision-language models detect images copyright infringement from genai? arXiv preprint arXiv:2502.16618, 2025

  43. [51]

    Auto-gpt for online decision making: Benchmarks and additional opinions

    Hui Yang, Sifu Yue, and Yunzhong He. Auto-gpt for online decision making: Benchmarks and additional opinions. arXiv preprint arXiv:2306.02224, 2023

  44. [52]

    Enhancing model defense against jailbreaks with proactive safety reasoning

    Xianglin Yang, Gelei Deng, Jieming Shi, Tianwei Zhang, and Jin Song Dong. Enhancing model defense against jailbreaks with proactive safety reasoning. arXiv preprint arXiv:2501.19180, 2025

  45. [53]

    Tree of thoughts: Deliberate problem solving with large language models

    Shinn Yao, Jeffrey Zhao, Dian Yu, Izhang Zhao, Karthik Reynoso, Luyu Hou, Eric Cheng, Kevin Park, Shunyu Gao, Thomas Yu, et al. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023

  46. [54]

    Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making

    Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yuechen Jiang, Yupeng Cao, Zhi Chen, Jordan Suchow, Zhenyu Cui, Rong Liu, et al. Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making. Advances in Neural Inf...

  47. [55]

    Wordcraft: story writing with large language models

    Ann Yuan, Andy Coenen, Emily Reif, and Daphne Ippolito. Wordcraft: story writing with large language models. In Proceedings of the 27th International Conference on Intelligent User Interfaces, pages 841–852, 2022

  48. [56]

    Zelikman et al

    A. Zelikman et al. Math: A benchmark for mathematical reasoning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021

  49. [57]

    Prompting large language model for machine translation: A case study

    Biao Zhang, Barry Haddow, and Alexandra Birch. Prompting large language model for machine translation: A case study. In International Conference on Machine Learning, pages 41092– 41110. PMLR, 2023

  50. [58]

    Zhang et al

    H. Zhang et al. Contexthub: A dataset for context-aware reasoning. In Proceedings of the 2021 Conference on Machine Learning (ICML), 2021

  51. [59]

    Vasilakos

    Gejian Zhao, Hanzhou Wu, Xinpeng Zhang, and Athanasios V . Vasilakos. Shadowcot: Cognitive hijacking for stealthy reasoning backdoors in llms. arXiv preprint arXiv:2504.05605, 2025

  52. [60]

    Zhao et al

    X. Zhao et al. Folio: A formal logic dataset for evaluating reasoning models. In Proceedings of the 2022 Conference on Artificial Intelligence (AAAI), 2022

  53. [61]

    Denny Zhou, Nathanael Schärli, Lu Hou, Jason Wei, Xuezhi Wang, Tushar Khot, Ashish Sabharwal, Kelvin Guu, and Ed H. Chi. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2023

  54. [62]

    Unlearnable examples: Protecting data against unauthorized learning

    Di Zou, Zhizheng Liu, Xingjun Liu, Lingjuan Xie, Shouling Xu, and Dawn Song. Unlearnable examples: Protecting data against unauthorized learning. arXiv preprint arXiv:2302.01785, 2023. 13 Appendix A Detailed Algorithm and Method A.1 Notation Table We provide a summary of notat...

  55. [63]

    Explain like a teacher help- ing students understand fair sharing

    This follows the strategy in Section 4.1 of the main paper. Algorithm 2 Trigger-CoT Prompt Injection 1: Input: Task prompt p, trigger key k, task type t 2: Output: Modified prompt ˜p with embedded trigger 3: Generate trigger pattern τ from k and t: τ ← T (k, t) 4: Append trigg...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.