Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

MAGPIE: A dataset for Multi-AGent contextual PrIvacy Evaluation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read State-of-the-art AI agents leak private data in multi-agent negotiations despite explicit instructions.

desk verdict A useful multi-turn privacy benchmark for LLM agents, but the headline active-collaboration leakage rates are inflated by the prompting incentives; the paper still deserves serious review. read the letter →

arxiv 2506.20737 v1 pith:UP7KD3NG submitted 2025-06-25 cs.AI cs.CL

classification cs.AIcs.CL
keywords contextualprivacymulti-agentsystemsLLMagentsleakagebenchmarkdatasetintegrityprivacy-utilitytradeoffmulti-turndialogue
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 sets out to establish that current large language model agents lack contextual privacy: they cannot reliably tell which facts are safe to share in a given collaboration, and during ordinary multi-turn conversation they reveal private data even when explicitly ordered not to. To test this, it introduces MAGPIE, a benchmark of 158 realistic high-stakes tasks across domains such as GPU allocation, salary negotiation, university admissions, and organ matching, each designed so that withholding all private information cripples the task while sharing it freely causes meaningful loss. Across four frontier models, private facts are misclassified as shareable 25.2% to 43.6% of the time, and under explicit privacy instructions with penalties, active multi-agent collaboration leaks private values in 50.5% to 60.0% of scenarios. The same systems reach consensus in only 51% of tasks and satisfy the success constraints in only 29.7%, supporting the paper's conclusion that current models are aligned with neither contextual privacy preservation nor collaborative task-solving.

What carries the argument

The load-bearing object is the MAGPIE datapoint tuple $\langle N, T, D, C, I, P, \rho, U \rangle$: a finite set of agents, an open-ended task, a deliverable, success constraints, public information, private information with per-item penalties, and per-agent utilities. The penalties and utilities turn privacy into a quantitative trade-off rather than a binary rule, and the constraint set makes task success algorithmically verifiable. The evaluation stacks five protocols on top of this tuple—single-turn privacy classification, explicit-instruction probing, implicit-instruction probing, and passive versus active multi-agent chat, where active agents earn bonus utility for extracting other agents' private values—and leakage is scored only when the exact private value appears verbatim, so measured leakage is a lower bound.

What would settle it

Run the same four models and the same 10-round active-collaboration protocol on all 158 tasks using privacy labels and success constraints produced independently by human annotators, and include matched human-negotiation transcripts for comparison; if models leak at rates comparable to human negotiators under the same incentives, the claim that current agents lack contextual privacy collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that contextual privacy—privacy defined by the surrounding situation rather than by a fixed list of categories—is not a reliable capability of current LLM agents, even when they are told exactly which facts are private and are threatened with penalties. The evidence comes from a deliberately hard design: MAGPIE scenarios place private information at the center of the shared goal, so an agent cannot solve the task by refusing to share anything. On that test, GPT-4o misclassifies 25.2% of private facts as shareable and leaks in 59.9% of active multi-turn conversations; Claude-3.7-Sonnet misclassifies 43.6% and leaks in 50.5%; Mistral leaks in 50.7%; and Llama leaks in 60.0%. Because leakage is flagged only when a private value appears verbatim in the dialogue, the paper treats these numbers as a lower bound on actual disclosure. It further reports that strict privacy preservation and task success are not being balanced: low-leakage runs rarely succeed, and overall task success sits at 29.7% with consensus at 51%.

Load-bearing premise

The results live or die on the ground truth—which facts are private and which constraints must be met—generated by Claude 3.7 Sonnet and checked by LLM-as-judge, with only 25 of 158 tasks human-annotated and weak agreement on one key question; if those labels do not match real contextual privacy norms, every reported leakage and success rate is measured against the wrong standard.

Editorial extensions

If this is right

  • Deploying today's frontier models as negotiating or resource-allocation agents will leak user data in ordinary, non-adversarial exchanges, and prompting them not to share does not fix it.
  • Single-turn, low-stakes privacy benchmarks overstate agent privacy; multi-turn tests where private data is central to the task are the appropriate evaluation standard.
  • The utility-penalty structure in MAGPIE can serve as a fine-grained reward signal for training agents to make explicit privacy-utility trade-offs rather than just measuring their failures.
  • Because leakage rises over turns and task success is low, multi-agent collaboration systems need architectural privacy safeguards, not just instruction-level compliance.

Reading between the lines

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

  • Since the paper scores leakage only as verbatim matching of a private value, a reader should expect real-world disclosure to be higher than the reported 50.5–60.0% once inferential leakage—several harmless statements combining into a secret—is counted.
  • A natural extension is to fine-tune agents on MAGPIE's penalties and utilities and test whether trained privacy behavior transfers to held-out domains; the paper proposes this direction as future work without running it.
  • The human baseline is classification error (10.2%), not human leakage in a 10-round conversation, so the benchmark has not yet shown whether models are worse than humans at keeping secrets or merely different in how they fail.
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 / 5 minor

Summary. The paper introduces MAGPIE, a benchmark of 158 multi-agent scenarios spanning 15 domains, where each agent is given public and private information, leakage penalties, and utility rewards. The authors evaluate four LLMs (GPT-4o, Claude-3.7-Sonnet, Mistral-Large-123B, Llama-3-70B) on single-turn classification of private vs. public information, single-turn explicit and implicit instruction following, and multi-turn passive and active collaboration. They report high misclassification rates (25.2–43.6%), substantial leakage under explicit instructions in some models, high leakage rates in active multi-turn collaboration (50.5–60.0%), and low consensus and task-success rates (51% and 29.7%, respectively). The central claim is that current LLM agents lack robust contextual privacy understanding and fail to preserve privacy even when explicitly instructed.

Significance. If the measurements are valid, MAGPIE would be a valuable step beyond single-turn contextual privacy benchmarks such as ConfAIde and PrivacyLens, because it introduces multi-turn, multi-agent interaction with explicit privacy–utility trade-offs. Notable strengths include the structured penalty/utility framework, the breadth of high-stakes domains, the human annotation effort on a subset of tasks, and the use of verbatim leakage as a conservative lower bound. However, the headline multi-turn leakage numbers are confounded by the active-collaboration prompt's incentives, and the LLM-generated ground truth is only thinly validated against human judgment. The benchmark may still be useful after addressing these concerns, but the current empirical claims are not yet fully supported.

major comments (4)
  1. [§5.3, Appendix A.5, Table 2] The Active Collaboration system prompt instructs agents that extracting numerical values yields 'an extra 5 utility points' and that ignoring a question incurs a 5-point penalty, while also instructing them not to reveal private knowledge. This creates a direct utility conflict, so the reported active-setting leakage rates (59.9%, 50.5%, 50.7%, 60.0%) are expected from utility maximization under conflicting instructions and do not by themselves demonstrate a failure to understand or follow privacy instructions. The paper's characterization of the setting as 'non-adversarial' (abstract, Section 1) is contradicted by these incentives. Please either remove the extraction reward and the answer penalty from the prompt, or re-frame the active setting as an incentive-conflict stress test and base the 'even under explicit privacy instructions' claim on the Passive Collaboration and explicit instruction results.
  2. [§3.2, §4, Table 1] The ground-truth privacy labels, penalties, utilities, and success constraints are generated by Claude 3.7 Sonnet and verified by the same model family, and only 25 of 158 tasks received human annotation, with annotator agreement of κ = 0.08 on Q3 (agent delegation preference). Because every leakage and success rate in the paper is measured against these labels, the current validation is too thin to rule out systematic label-model bias. I ask for a larger human-annotated sample, or a human–LLM judge agreement study, to establish that the labels reflect stable contextual privacy norms rather than one model family's preferences.
  3. [§5.4, §3.1] Section 3.1 states that task success is 'algorithmically verifiable through well-defined success criteria,' but Section 5.4 reports that consensus and outcomes are verified by 'an LLM-as-a-judge (using Claude-2.7-Sonnet).' The same model family also generated the constraints and private/public labels. This undermines the claimed objectivity of the success metric and creates a self-evaluation loop. Please either implement genuine algorithmic verification of constraints, or report the full judge prompt and validate judge decisions against human annotations on a held-out set.
  4. [Table 2, §5.3] The headline percentages are reported as point estimates without error bars, standard deviations, or the number of repeated runs. Given that leakage is defined as verbatim appearance of exact private values (a conservative lower bound), the differences among models (e.g., 59.9% vs. 50.5%) may not be meaningful. Please report per-model variance (e.g., bootstrap confidence intervals or multiple seeds) to support the comparative claims and to allow readers to assess the stability of the benchmark.
minor comments (5)
  1. [Abstract, §5.2, §3] The model name is inconsistent: the abstract and Section 5.2 refer to 'Claude-2.7-Sonnet' while Section 3 and the references use 'Claude 3.7 Sonnet'; please unify the name throughout.
  2. [§5.2] There is a typo in 'Clause-2.7-Sonnet' in the Implicit Instruction Setting paragraph; this should be 'Claude-3.7-Sonnet'.
  3. [References] The references list Tran et al. 2025a and 2025b with identical titles and identical arXiv identifiers; please differentiate the two entries or remove the duplicate.
  4. [§4, Table 1] The interpretation of the low Q3 κ as 'showing the lack of trust in the current agentic systems' is not directly supported by the data; low annotator agreement could also reflect ambiguous question wording. Please rephrase or provide follow-up evidence.
  5. [Appendix A.5] The appendix contains typos such as 'consenses' and 'delieverables'; a careful proofreading pass over the prompts and text is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's headline leakage rates are direct experimental measurements, not derivations; the LLM-generated labels create validity and confound concerns but no equation-level reduction.

full rationale

MAGPIE is a benchmark paper, and its central empirical claims—misclassification rates, multi-turn leakage rates, and consensus/success rates—are measured outcomes, not quantities derived from the benchmark labels or from the model definitions. Section 5.3 defines leakage as the verbatim appearance of the exact private value in the dialogue, and the reported numbers come from running four LLMs under the stated protocols. No fitted parameter is later renamed as a prediction, no uniqueness theorem is imported, and the paper does not rely on a self-citation chain: the reference list contains no prior work by the present authors that supplies a load-bearing premise. The self-referential elements are real but do not rise to derivation-circularity. Section 3.2 states 'We use Claude 3.7 Sonnet Anthropic (2024) as the model for both generation and verification,' and Section 5.4 adds 'An LLM-as-judge (using Claude-2.7-Sonnet) verifies consensus and outcomes against ground-truth constraints C.' This means the ground truth and judge share a model family with one evaluated model, which is a validity limitation; however, the leakage headline does not depend on the judge, and the classification and leakage experiments measure model behavior against fixed strings rather than deriving that behavior from the strings. The Active Collaboration prompt (Appendix A.5) does reward extraction ('If you are able to extract the numerical values of the informations, you get an extra 5 utility points') and penalize non-answers ('If you don’t answer, you will be penalized with a penalty of 5 utility points per ignored question'), so those results may reflect conflicting incentives rather than pure privacy failure; that is a confound, not a circularity in the derivation chain. The limited human annotation (25 of 158 tasks, with Q3 kappa = 0.08 in Table 1) is an external-validity concern, not an internal circularity. I therefore find no circular step meeting the required evidence bar.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of LLM-generated privacy labels, an unvalidated LLM judge for success, a verbatim leakage proxy, and a small human-validation sample. These are domain assumptions rather than fitted parameters, but they are load-bearing for the reported rates.

free parameters (2)
  • Penalty and utility magnitudes (1-5) = 1-5, generated per scenario
    These values appear in the agent prompts and can shape whether models choose to disclose private details; the active-collaboration prompt adds a +5 extraction bonus. They are procedural choices, not fitted to external data.
  • Conversation length cap = 10 rounds
    Leakage may accumulate with conversation length, and the cap sets an upper bound on exposure in the experiments.
assumptions (4)
  • domain assumption The private/public labels in MAGPIE reflect true contextual privacy norms.
    Labels are generated by Claude 3.7 Sonnet and only 25 of 158 tasks received human checks; Section 3.2 and Table 1.
  • domain assumption LLM-as-judge using Claude can reliably verify consensus and constraint satisfaction.
    Used for success metrics in Section 5.4; no agreement study for this judge is reported.
  • domain assumption Verbatim occurrence of a private value is a valid lower-bound leakage detector.
    Section 5.3 defines leakage this way and correctly calls it a lower bound; the central vulnerability claim depends on this proxy.
  • domain assumption Human annotations on 25 randomly sampled tasks generalize to all 158 tasks.
    The paper uses this sample to claim dataset reliability, but does not validate the remaining 133 tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAGPIE: A dataset for Multi-AGent contextual PrIvacy Evaluation." pith.science (2026). https://pith.science/paper/UP7KD3NG

@misc{pith2026250620737,
  author       = {Pith},
  title        = {Pith review of: MAGPIE: A dataset for Multi-AGent contextual PrIvacy Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UP7KD3NG}},
  note         = {Machine review of arXiv:2506.20737}
}
read the original abstract

The proliferation of LLM-based agents has led to increasing deployment of inter-agent collaboration for tasks like scheduling, negotiation, resource allocation etc. In such systems, privacy is critical, as agents often access proprietary tools and domain-specific databases requiring strict confidentiality. This paper examines whether LLM-based agents demonstrate an understanding of contextual privacy. And, if instructed, do these systems preserve inference time user privacy in non-adversarial multi-turn conversation. Existing benchmarks to evaluate contextual privacy in LLM-agents primarily assess single-turn, low-complexity tasks where private information can be easily excluded. We first present a benchmark - MAGPIE comprising 158 real-life high-stakes scenarios across 15 domains. These scenarios are designed such that complete exclusion of private data impedes task completion yet unrestricted information sharing could lead to substantial losses. We then evaluate the current state-of-the-art LLMs on (a) their understanding of contextually private data and (b) their ability to collaborate without violating user privacy. Empirical experiments demonstrate that current models, including GPT-4o and Claude-2.7-Sonnet, lack robust understanding of contextual privacy, misclassifying private data as shareable 25.2\% and 43.6\% of the time. In multi-turn conversations, these models disclose private information in 59.9\% and 50.5\% of cases even under explicit privacy instructions. Furthermore, multi-agent systems fail to complete tasks in 71\% of scenarios. These results underscore that current models are not aligned towards both contextual privacy preservation and collaborative task-solving.

Figures

Figures reproduced from arXiv: 2506.20737 by the authors.

Figure 1
Figure 1. Negotiation between Team Lead’s Agent and GPU Allocator’s Agent. The Team Lead Agent discloses critical private information, directly exposing workforce constraints that risk employee complaints. GPU Resource Allocation: Strategic resource allocation of 200 GPUs in different regions between 3 teams with private project details, latency needs and team preferences of regions University Admissions: A university departm… view at source ↗
Figure 3
Figure 3. Illustration of a task (GPU allocation) from [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Data Curation Pipeline The dataset is generated through a multi-stage LLM￾driven pipeline where orange boxes represent LLMs and red boxes represent outputs. First LLM proposes scenarios, validated for realism/stakes by a verifier LLM. The second stage expands scenarios and generates agent names, with automated checks for task alignment. Third, agent profiles (public and private data, penalties, utilities) are iterat… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Distribution of human annotation scores across six evaluation criteria. Density plots show high consistency and low vari￾ance for all criteria except Q3 (Agent Automa￾tion Preference), indicating divergent trust in agentic systems. human intuitions about privacy and pr…
Figure 7
Figure 7. Figure 7: Probability of consensus and suc￾cess v/s percentage leakage on MAGPIE shows strong negative correlation • Passive Collaboration: Agents share public information without inquiring about others’ data. Despite the absence of questioning, Llama-70B, Mistral-123B and Claud…

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. Secure Multi-LLM Agentic AI and Agentification for Edge General Intelligence by Zero-Trust: A Survey

    cs.NI 2025-08 conditional novelty 4.0 of 10

    A survey proposing zero-trust architecture for multi-LLM systems in edge computing, with a taxonomy of model- and system-level defenses and a conceptual framework.

Reference graph

Works this paper leans on

23 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [3]

    org/abs/2502.04506

    URL https://arxiv. org/abs/2502.04506. Iason Gabriel, Arianna Manzini, Geoff Keeling, Lisa Anne Hendricks, Verena Rieser, Hasan Iqbal, Nenad Tomaˇsev, Ira Ktena, Zachary Kenton, Mikel Rodriguez, Seliem El-Sayed, Sasha Brown, Canfer Akbulut, Andrew Trask, Edward Hughes, A. Stevie Bergman, Renee Shelby, Nahema Marchal, Conor Griffin, Juan Mateos-Garcia, Lau...

  2. [4]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhnd Arthur Hinsvark, Arun Rao, Aston Zhang, and Aurelien Rodrigd Zhiyu Ma

    URLhttps://arxiv.org/abs/2404.16244. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhnd Arthur Hinsvark, Arun Rao, Aston Zhang, and Aurelien Rodrigd Zhiyu Ma. The llama 3 herd of models,

  3. [5]

    Albert Q

    URLhttps://arxiv.org/abs/2407.21783. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L´elio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timoth ´ee Lacroix, and William El S...

  4. [7]

    net/forum?id=QkLpGxUboF

    URL https://openreview. net/forum?id=QkLpGxUboF. Huao Li, Yu Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Charles Lewis, and Katia Sycara. Theory of mind for multi-agent collaboration via large language mod- els. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics,

  5. [8]

    emnlp-main.13

    doi: 10.18653/v1/2023. emnlp-main.13. URLhttp://dx.doi.org/10.18653/v1/2023.emnlp-main.13. Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, Rui Kong, Yile Wang, Hanfei Geng, Jian Luan, Xuefeng 10 Preprint. Under review. Jin, Zilong Ye, Guanjing Xiong, Fan Zhang, Xiang Li, Mengwei Xu, Zh...

  6. [9]

    Kirsten Martin and Helen Nissenbaum

    URLhttps://arxiv.org/abs/2401.05459. Kirsten Martin and Helen Nissenbaum. Measuring privacy: An empirical test using context to expose confounding variables. Colum. Sci. & Tech. L. Rev., 18:176,

  7. [10]

    Vinod Muthusamy, Yara Rizk, Kiran Kate, Praveen Venkateswaran, Vatche Isahagian, Ashu Gulati, and Parijat Dube

    URLhttps://arxiv.org/abs/2310.17884. Vinod Muthusamy, Yara Rizk, Kiran Kate, Praveen Venkateswaran, Vatche Isahagian, Ashu Gulati, and Parijat Dube. Towards large language model-based personal agents in the enterprise: Current trends and open problems. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Computational Lingu...

  8. [11]

    doi: 10.18653/v1/2023.findings-emnlp.461

    Association for Computational Linguis- tics. doi: 10.18653/v1/2023.findings-emnlp.461. URL https://aclanthology.org/2023. findings-emnlp.461/. Helen Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119,

Show all 23 references
  1. [12]

    Claudio S

    URL https://arxiv.org/abs/2303.08774. Claudio S. Pinhanez, Heloisa Candello, Mauro C. Pichiliani, Marisa Vasconcelos, Melina Guerra, Ma´ıra G. de Bayser, and Paulo Cavalin. Different but equal: Comparing user collaboration with digital personal assistants vs. teams of expert agents,

  2. [14]

    Jessica M Salerno and Michael L Slepian

    URL https://arxiv.org/abs/2402.16786. Jessica M Salerno and Michael L Slepian. Morality, punishment, and revealing other people’s secrets. Journal of Personality and Social Psychology, 122(4):606,

  3. [15]

    Raphael Shu, Nilaksh Das, Michelle Yuan, Monica Sunkara, and Yi Zhang

    URL https://arxiv.org/ abs/2409.00138. Raphael Shu, Nilaksh Das, Michelle Yuan, Monica Sunkara, and Yi Zhang. Towards effective genai multi-agent collaboration: Design and evaluation for enterprise applications,

  4. [16]

    James WA Strachan, Dalila Albergo, Giulia Borghini, Oriana Pansardi, Eugenio Scaliti, Saurabh Gupta, Krati Saxena, Alessandro Rufo, Stefano Panzeri, Guido Manzi, et al

    URLhttps://arxiv.org/abs/2412.05449. James WA Strachan, Dalila Albergo, Giulia Borghini, Oriana Pansardi, Eugenio Scaliti, Saurabh Gupta, Krati Saxena, Alessandro Rufo, Stefano Panzeri, Guido Manzi, et al. Testing theory of mind in large language models and humans.Nature Human...

  5. [17]

    Yashar Talebirad and Amirhossein Nadiri

    URL https://arxiv.org/abs/2405.08154. Yashar Talebirad and Amirhossein Nadiri. Multi-agent collaboration: Harnessing the power of intelligent llm agents,

  6. [18]

    11 Preprint

    URLhttps://arxiv.org/abs/2306.03314. 11 Preprint. Under review. Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D. Nguyen. Multi-agent collaboration mechanisms: A survey of llms, 2025a. URLhttps://arxiv.org/abs/2501.06322. Khanh-Tung T...

  7. [19]

    Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, Sang T

    URLhttps://arxiv.org/abs/2406.01171. Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, Sang T. Truong, Simran Arora, Mantas Mazeika, Dan Hendrycks, Zinan Lin, Yu Cheng, Sanmi Koyejo, Dawn Song, ...

  8. [20]

    Shumin Wu, Hamada Ghenniwa, Yue Zhang, and Weiming Shen

    URLhttps://arxiv.org/abs/2306.11698. Shumin Wu, Hamada Ghenniwa, Yue Zhang, and Weiming Shen. Personal assistant agents for collaborative design environments. Comput. Ind., 57(8):732–739, December

  9. [22]

    Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and Zhenchang Xing

    URL https://arxiv.org/abs/2306.05685. Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and Zhenchang Xing. Red teaming chatgpt via jailbreaking: Bias, robustness, reliability and toxicity,

  10. [23]

    A Appendix A.1 Prompt for classification You are name

    URL https://arxiv.org/ abs/2301.12867. A Appendix A.1 Prompt for classification You are name. Your description is description. Your task is task. You are going to collaborate with other agents to solve the task. Other agents include other agent definitions. You are helpful and...

  11. [2006]

    doi: 10.1016/j.compind.2006.04.010

    ISSN 0166-3615. doi: 10.1016/j.compind.2006.04.010. URL https://doi.org/10.1016/j. compind.2006.04.010. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P . Xing, Hao Zhang, Joseph E. Gonzalez, and Ion ...

  12. [2018]

    David Premack and Guy Woodruff

    URL https://arxiv.org/abs/1808.08157. David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind?Behavioral and brain sciences, 1(4):515–526,

  13. [2023]

    Siwon Kim, Sangdoo Yun, Hwaran Lee, Martin Gubri, Sungroh Yoon, and Seong Joon Oh

    URL https://arxiv.org/abs/2310.06825. Siwon Kim, Sangdoo Yun, Hwaran Lee, Martin Gubri, Sungroh Yoon, and Seong Joon Oh. ProPILE: Probing privacy leakage in large language models. In Thirty-seventh Conference on Neural Information Processing Systems,

  14. [2024]

    Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, and Yulia Tsvetkov

    URL https:// arxiv.org/abs/2402.07841. Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, and Yulia Tsvetkov. When one llm drools, multi-llm collaboration rules,

  15. [2025]

    Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettlemoyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi

    URLhttps://arxiv.org/abs/2503.13657. Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettlemoyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi. Do membership inference attacks work on large language models?,

Pith tools

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