Pith. sign in

REVIEW 3 major objections 3 minor 31 references

CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation

T0 review · 3 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read CARD generates credit card discussion threads whose aggregate statistics are close enough to real Reddit threads that statistical tests find no difference on 91.7% of metric-test pairs, outperforming prior simulators.

desk verdict A useful system paper whose headline realism numbers are inflated by optimizing the same metrics it evaluates on; the framework itself is sound and worth building on. read the letter →

arxiv 2608.09790 v2 pith:MRFPIMNO submitted 2026-08-10 cs.AI cs.MAcs.SI

Yaoning Yu , Kai-Min Chang , Ye Yu , Yi-Chia Wang , Haojing Luo , Haohan Wang This is my paper · ORCID
classification cs.AIcs.MAcs.SI
keywords creditcarddiscussionsimulationLLMagentsdistributionalrealismthreadgenerationself-loopcalibrationconsumerfinancecommunicationRedditthreads
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

CARD is a framework for simulating online credit card discussions, and the paper claims it produces whole threads whose collective statistics match real Reddit threads far better than existing social simulators. The central evidence is distributional: across lexical, semantic, behavioral, and structural metrics, the generated collection is close enough to the real collection that standard tests fail to detect a difference on 91.7% of metric-test pairs, while the strongest comparison baseline reaches 37.5%. The realism comes from planning each comment's communicative role before any text is written, then revising the entire collection toward real distributions while protecting each comment's planned financial function. If the claim holds, researchers gain a controlled way to study how consumers talk about fees, rewards, eligibility, and trade-offs without running large real-world interventions.

What carries the argument

The load-bearing mechanism is the self-looped population calibrator, a revision loop that works metric by metric over the whole generated collection. In each round it diagnoses which metric shows a significant generated-versus-real difference, designates every other metric as protected, selects the threads and comments with the largest contribution to the gap, samples best-of-N candidate rewrites conditioned on the comment's planned financial role, locally accepts rewrites that reduce the target gap while passing a role-preservation check, and retains a whole round only if protected metrics stay within tolerance. The planner's control vector—one tuple per comment encoding parent, depth, length, financial discussion function, claim stance, reasoning move, issue type, tone, and opening style—is what anchors each rewrite so that calibration improves statistics without turning a recommendation into an explanation or dropping a cost–benefit trade-off. Context dropout during writing, inspired by dropout in neural networks, gives later comments varied and partial views of the discussion, which reduces repeated semantic patterns.

What would settle it

Reproduce the evaluation after randomly permuting which real thread is paired with each seed post; if the non-rejection rate stays near 91.7% under this permutation, the claimed match is not actually driven by the matched-pair design. A content-level audit would then be needed, since the benchmark's metrics do not check factual accuracy.

Watch

Extended reading notes

Core claim

The paper's central claim is that thread-level realism can be engineered by separating what a comment does within a discussion from what it literally says. CARD first plans each thread as a sequence of abstract communicative acts—reply structure, financial discussion function, claim stance, reasoning move, issue type, tone, length, and opening style—then a writer turns each plan into text conditioned on the seed post and a partial view of the conversation. A self-looped calibration pass then selects the threads and comments contributing most to a distributional gap, generates candidate rewrites, and accepts only those that shrink the target gap while preserving the planned financial role. Across the benchmark's twelve metrics, CARD with the two strongest backbones reaches a non-rejection rate of 91.7% on both Mann–Whitney U and Kolmogorov–Smirnov tests, and it shows smaller Cliff's $\delta$ and Wasserstein distances than the baselines. A blind Turing test also identifies CARD threads as human-written or indistinguishable more often than the baselines.

Load-bearing premise

The evaluation assumes that the benchmark's metric set, together with one matched real thread per seed post, adequately captures what makes a credit card discussion realistic; if the matching is unrepresentative or the metrics miss content-level accuracy, calibration can fit statistics while missing substance.

Editorial extensions

If this is right

  • Credit card discussion threads generated this way can stand in for real threads in studies of how consumers discuss rewards, fees, eligibility, and trade-offs, at least on the distributional dimensions the benchmark measures.
  • The same planner-writer-calibrator structure should transfer to other financial or non-financial discussion domains once domain-specific discussion functions and metrics are defined, as the paper's Future Work states.
  • Because calibration preserves each comment's planned role, statistical matching can be improved without distorting the substance of financial advice, disagreement, or product comparison.
  • The result strengthens the case that realism in social simulation is a collection-level property, not just a property of individual comments.

Reading between the lines

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

  • Editorial inference: the reported 91.7% is a claim about distributional similarity, not about content accuracy; the benchmark's metrics do not check whether fee amounts, reward rules, or eligibility claims are factually correct, so two statistically matching threads could still disagree on substance.
  • Editorial inference: a direct stress test would vary the matching procedure that pairs each seed post with a real thread; if non-rejection rates swing with the choice of matched thread, much of the result depends on the calibration target rather than on the framework's planning and revision machinery.
  • Editorial inference: applied to high-stakes domains such as medical or legal advice, the same recipe would need content-level guards; distributional realism alone would not certify that generated advice is safe, since a thread can match human style while asserting false rules.
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

3 major / 3 minor

Summary. The paper introduces CARD, a framework for generating simulated Reddit discussion threads about credit cards. Given a seed post and its matched real thread, a planner specifies discussion structure and comment-level roles (function, stance, tone, reasoning move, issue type), a writer generates comments with an LLM, and a self-looped calibrator revises comments to reduce distributional mismatches with the real thread collection across lexical, semantic, behavioral, and structural metrics. The authors evaluate CARD on 150 real Reddit credit-card seed posts against OASIS and SynthPAI baselines, reporting Mann–Whitney U and Kolmogorov–Smirnov p-values, Cliff's delta and Wasserstein distances, and a blind Turing test. The headline result is that CARD achieves non-rejection rates up to 91.7% of metric–test pairs (p>0.05) and smaller distributional distances than baselines.

Significance. If the claims are well supported, CARD would be a useful step toward controllable, realistic simulation of consumer-finance discussions, and the planner–writer–calibrator architecture is a sensible way to inject domain structure into LLM generation. The paper is strong in transparency: the revision loop, protection policy, and rollback mechanism are described in detail, and the evaluation covers multiple backbone models and reports effect sizes in addition to p-values. However, the central evidence is weakened by a calibration–evaluation circularity and by unequal access to the matched real thread across methods, so the significance depends on whether the authors can provide a more independent test of realism.

major comments (3)
  1. [§3.3.3, §4.2, Table 2] The self-looped calibrator selects target metrics from the same MiroBench metric set used in the evaluation, revises comments until the Mann–Whitney U and Kolmogorov–Smirnov tests no longer reject (p>0.05) or a maximum of seven iterations is reached, and Table 2 reports non-rejection rates on exactly those metrics. This makes the headline non-rejection rates posterior to fitting the evaluation distribution, not independent evidence of realism. Under the null, p>0.05 is expected with probability 0.95 per test, so the reported 91.7% is close to the trivial chance level. The paper itself correctly notes (§4.2) that non-rejection does not establish equivalence. To substantiate the central claim, the authors should report results on held-out metrics that the calibration loop does not optimize, or otherwise demonstrate that the improvement transfers beyond the calibrated metric set.
  2. [§3.3.1, §4.2, Table 5] The planner derives comment count, reply depth, and branch structure 'from the matched real discussion,' and the evaluation measures average reply depth and structural virality. Consequently, part of the reported structural match is constructed by copying the target thread's structure, while the baselines (OASIS and SynthPAI) do not receive the matched real thread at all. This makes the comparison unequal and weakens the claim that CARD generates realistic discussion structures in a target-free setting. The authors should either provide the same matched-thread information to the baselines or evaluate CARD in a setting where no matched real thread is used for planning.
  3. [§6, Abstract] The Future Work section explicitly concedes that content-level objectives—topic coverage, claim alignment, and factual consistency—are absent from the current framework. The abstract's phrase 'matches the distributions of real credit card discussions' thus overstates the scope of the realism claim; the evidence supports structural, lexical, and stylistic matching, not content-level realism. The conclusions and title should be restricted accordingly, and the evaluation should avoid implying equivalence from p>0.05.
minor comments (3)
  1. [§3.3] There is a typo in the description of the matching procedure: 'CARD retrieves one matched real with the same about the same financial scenario' should read 'one matched real thread about the same financial scenario.'
  2. [§4.3, Figure 3] The human evaluation description reports '20 matched discussion pairs for each method' and '72 unique discussion pairs across all methods,' but with two comparison methods (OASIS and SynthPAI) one would expect 40 unique pairs; the counting should be clarified.
  3. [§3.3.1] The matching procedure for selecting the 'matched real thread' for each seed post is not described; because the planner and the evaluator both rely on this matched thread, the authors should specify how matching was performed and report matching quality statistics.

Circularity Check

3 steps flagged · score 6.0 of 10

Distributional claim is a post-fit description: calibration optimizes the same MiroBench metric set used for evaluation, and reply structure is copied from the matched real thread.

  1. fitted input called prediction [Section 3.3.3 calibration loop; Section 4.3 Table 2]
    "Each revision round starts by comparing the current generated thread collection with the matched real thread collection using the metric families in Section 4.2. If statistical tests detect a significant difference between the generated and real distributions for a metric, the controller selects that metric as the target metric φ_t. ... For each target metric, the revision loop continues until either no significant difference is detected or the maximum of seven revision iterations is reached."

    The calibration loop uses the very metric families that Section 4.2 and Table 2 then use as the evaluation battery. Because the loop revises comments until p > 0.05 (or 7 iterations), the reported 91.7% non-rejection rate is a stopping-rule outcome, not an independent test. Under the null, 95% of p-values exceed 0.05, so 91.7% is close to trivial; the paper itself notes that p > 0.05 does not establish equivalence. Baselines get no such calibration access to the matched real thread, so the headline comparison measures target optimization rather than generalizable realism.

  2. fitted input called prediction [Section 3.3.1 Discussion planner; Section 4.2 Table 5]
    "It controls the number of comments, reply depth, branch structure, and the balance between top-level and nested replies. These quantities are derived from the matched real discussion and adjusted by global generation parameters, such as comment count and discussion size."

    Table 5 evaluates 'Average reply depth' and 'Structural virality' as structural metrics. Because the planner derives reply count, depth, and branch structure from the matched real thread, the structural pass rates in Table 2 are partly constructed by copying the target, not by predicting it. Baselines are not given the matched real thread, so the structural comparison is unequal. This is a second, independent channel by which the headline match is built into the input.

1 more flagged steps
  1. self citation load bearing [Section 4.2 Evaluation Metrics; reference [29]]
    "We follow the MiroBench [29] credit-card thread-level evaluation protocol. ... [29] Yaoning Yu, Ye Yu, Haojing Luo, and Haohan Wang. 2026. MiroBench: Benchmarking Realism in Agentic Simulation of Real-world Discussions."

    MiroBench is authored by four of this paper's authors, so the evaluation protocol that defines realism for the central claim is imported from the authors' own prior work. It is also exactly the same metric set used as the calibration objective in Section 3.3.3, so the self-citation supplies the loop's target and the evaluation yardstick. Because the metric definitions are restated in Table 5, this is a secondary aggravating factor rather than the primary circularity.

full rationale

The central claim is not wholly circular: CARD generates novel comment text, is compared against baselines on effect sizes and Wasserstein distances, and has a human Turing test that does not use the calibration metrics. However, the headline distributional claim is substantially self-confirming. Section 3.2 defines M as the evaluation metrics, Section 3.3.3 calibrates until p > 0.05 on M, and Section 4.3/Table 2 reports non-rejection on M; p > 0.05 is the null expectation 95% of the time, so 91.7% is close to trivial. Structural metrics (average reply depth, structural virality) are additionally 'derived from the matched real discussion' in Section 3.3.1, so part of Table 2 is copying the target. The MiroBench evaluation protocol is a self-citation by the same authors. Finally, Section 6 concedes that content-level realism (topic coverage, claim alignment, factual consistency) is not modeled. The independent components (Turing test, baselines without calibration) prevent an 8+, but the distributional comparison is largely a fit to the evaluation distribution, giving a score of 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the chosen evaluation metric set, the matched real-thread targets, and the calibration stopping rule. No new physical or ontological entities are introduced; the planner, writer, and calibrator are software components rather than invented entities. The free parameters that most affect the headline results are the revision budget and the unreported acceptance thresholds.

free parameters (5)
  • Maximum revision iterations per metric = 7
    Hand-chosen cap in Section 3.3.3; if the budget is exhausted a metric is left as not passing, so the value affects the reported non-rejection rates.
  • Minimum gap reduction threshold epsilon_t = not reported
    Local acceptance criterion for candidate rewrites in Section 3.3.3; the numeric value is never given, so the optimization strength is unspecified.
  • Protected-metric tolerance = not reported
    Round-level acceptance requires protected metrics to stay 'within tolerance' (Section 3.3.3), but the tolerance is never defined.
  • Context dropout and variation rates = not reported
    Writer variation controls in Section 3.3.2 determine how much discussion history each generated comment sees; no rates are given.
  • Comment count and discussion size parameters = not reported
    Planner adjusts reply counts and structure from the matched real discussion using unspecified global generation parameters (Section 3.3.1).
assumptions (5)
  • domain assumption The matched real discussion retrieved for each seed post is a valid distributional target for realism.
    Section 3.3 retrieves a real thread for each seed post and uses it as non-verbatim guidance; the matching procedure and representativeness are not described.
  • domain assumption The MiroBench metric set in Section 4.2 sufficiently captures discussion realism.
    The evaluation follows MiroBench [29]; Future Work concedes content-level objectives such as topic coverage and factual consistency are absent, so the metric set is a partial characterization of realism.
  • domain assumption Failure to reject the null hypothesis (p>0.05) is a useful success criterion.
    Section 4.2 explicitly notes that p>0.05 does not establish equivalence, yet the calibration loop and Table 2 treat non-rejection as the target outcome.
  • domain assumption LLM-based classifiers for politeness, impoliteness, emotion, and story probability are accurate enough for distributional comparison.
    Section 4.2 defines these metrics without validating the underlying classifiers or reporting error rates.
  • domain assumption Reddit credit card discussions are a representative sample of consumer financial communication.
    The introduction frames online credit card discussions as the natural setting for consumer communication, but no sampling or representativeness analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation." pith.science (2026). https://pith.science/paper/MRFPIMNO

@misc{pith2026260809790,
  author       = {Pith},
  title        = {Pith review of: CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRFPIMNO}},
  note         = {Machine review of arXiv:2608.09790}
}
read the original abstract

Online credit card discussions provide a natural setting for studying how consumers communicate about financial products. Simulating these discussions requires more than just generating individual comments, the generated threads should also match how real users express themselves and interact with others. We introduce CARD, a framework for generating realistic credit card discussion threads. Given a credit card post and its matched real thread, CARD uses non-verbatim guidance on reply structure, comment function, stance, tone, and conversational variation. A planner organizes these controls, a writer generates the discussion, and a calibration loop updates comments' populations that contribute to differences between the generated and real thread distributions. We evaluate CARD on real Reddit credit card discussions using lexical, semantic, behavioral, and structural metrics. CARD matches the distributions of real credit card discussions better than simulation baselines across multiple LLMs and also demonstrates smaller effect sizes and distribution distances across metrics. These results show that structured planning and targeted revision can generate the realism of simulated credit card discussions.

Figures

Figures reproduced from arXiv: 2608.09790 by the authors.

Figure 1
Figure 1. Overview of CARD. Given a credit card seed post and its matched real discussion. The planner controls reply structure, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distributional agreement with real credit-card dis [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Results of the blind Turing test over 20 discussion [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 8 linked inside Pith

  1. [1]

    Sumit Agarwal, Sujit Chakravorti, and Anna Lunn. 2010. Why do banks reward their customers to use their credit cards? Technical Report. Working paper

    2010

  2. [2]

    Sumit Agarwal, Souphala Chomsisengphet, Neale Mahoney, and Johannes Stroebel. 2015. Regulating consumer financial products: Evidence from credit cards. The Quarterly Journal of Economics 130, 1 (2015), 111–164

    2015

  3. [3]

    Selim Amrouni, Aymeric Moulin, Jared Vann, Svitlana Vyetrenko, Tucker Balch, and Manuela Veloso. 2021. ABIDES-gym: gym environments for multi-agent discrete event simulation and application to financial markets. In Proceedings of the Second ACM International Conference on AI in Finance. 1–9

    2021

  4. [4]

    Yu Bi, Lingxiao Zhao, Jinyun Tong, Zhe Feng, and Carmine Ventre. 2025. Agent- based Modeling and Simulation of Ambiguity in Catastrophe Insurance Markets. In Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems. 316–324

    2025

  5. [5]

    Leonardo Bursztyn, Bruno Ferman, Stefano Fiorin, Martin Kanz, and Gautam Rao. 2018. Status goods: experimental evidence from platinum credit cards. The Quarterly Journal of Economics 133, 3 (2018), 1561–1595

    2018

  6. [6]

    David Byrd, Maria Hybinette, and Tucker Hybinette Balch. 2019. Abides: Towards high-fidelity market simulation for ai research. arXiv preprint arXiv:1904.12066 (2019)

    arXiv 2019

  7. [7]

    Kshama Dwarakanath, Tucker Balch, and Svitlana Vyetrenko. 2024. ABIDES- Economist: Agent-Based Simulator of Economic Systems with Learning Agents. arXiv preprint arXiv:2402.09563 (2024)

    arXiv 2024

  8. [8]

    Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Jinghua Piao, Huandong Wang, Depeng Jin, and Yong Li. 2023. S3: Social-network Simulation System with Large Language Model-Empowered Agents. arXiv preprint arXiv:2307.14984 (2023)

    arXiv 2023

Show all 31 references
  1. [9]

    Shen Gao, Yuntao Wen, Minghang Zhu, Jianing Wei, Yuhan Cheng, Qunzi Zhang, and Shuo Shang. 2024. Simulating financial market via large language model based agents. arXiv preprint arXiv:2406.19966 (2024)

    2024 arXiv

  2. [10]

    John Gathergood, Neale Mahoney, Neil Stewart, and Jörg Weber. 2019. How do individuals repay their debt? The balance-matching heuristic. American Economic Review 109, 3 (2019), 844–875

    2019

  3. [11]

    Benedict Guttman-Kenney, Jesse Leary, and Neil Stewart. 2023. Weighing anchor on credit card debt. arXiv preprint arXiv:2305.11375 (2023)

    2023 arXiv

  4. [12]

    Conor Brian Hamill, Raad Khraishi, Simona Gherghel, Jerrard Lawrence, Salvatore Mercuri, Ramin Okhrati, and Greig Alan Cowan. 2025. Agent-based modelling of credit card promotions. International Journal of Bank Marketing 43, 4 (2025), 849–870

    2025

  5. [13]

    Patrick Hendy, Robert Slonim, and Kadir Atalay. 2021. Unsticking credit card repayments from the minimum: Advice, anchors and financial incentives.Journal of Behavioral and Experimental Finance 30 (2021), 100505

    2021

  6. [14]

    Hal E Hershfield and Neal J Roese. 2015. Dual payoff scenario warnings on credit card statements elicit suboptimal payoff decisions. Journal of Consumer Psychology 25, 1 (2015), 15–27

    2015

  7. [15]

    Harrison Hong, Jeffrey D Kubik, and Jeremy C Stein. 2004. Social interaction and stock-market participation. The journal of finance 59, 1 (2004), 137–163

    2004

  8. [16]

    Harrison Hong, Jeffrey D Kubik, and Jeremy C Stein. 2005. Thy neighbor’s portfolio: Word-of-mouth effects in the holdings and trades of money managers. The Journal of Finance 60, 6 (2005), 2801–2824

    2005

  9. [17]

    Benjamin J Keys and Jialan Wang. 2019. Minimum payments and debt paydown in consumer credit cards. Journal of Financial Economics 131, 3 (2019), 528–548

    2019

  10. [18]

    Nian Li, Chen Gao, Mingyu Li, Yong Li, and Qingmin Liao. 2024. EconA- gent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 15523–15536

    2024

  11. [19]

    Zeping Li, Guancheng Wan, Keyang Chen, Yu Chen, Yiwen Zhao, Philip Torr, Guangnan Ye, Zhenfei Yin, and Hongfeng Chai. 2026. Behavioral Consistency Val- idation for LLM Agents: An Analysis of Trading-Style Switching through Stock- Market Simulation. InFindings of the Associatio...

    2026

  12. [20]

    Xinyi Mou, Xuanwen Ding, Qi He, Liang Wang, Jingcong Liang, Xinnong Zhang, Libo Sun, Jiayu Lin, Jie Zhou, Huang Xuanjing, et al. 2026. From individual to society: A survey on social simulation driven by large language model-based agents. Comput. Surveys 58, 11 (2026), 1–41

    2026

  13. [21]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1–22

    2023

  14. [22]

    Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al . 2025. AgentSociety: Large- Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society. arXiv preprint ar...

    2025 arXiv

  15. [23]

    Alejandro Ponce, Enrique Seira, and Guillermo Zamarripa. 2017. Borrowing on the wrong credit card? Evidence from Mexico. American Economic Review 107, 4 (2017), 1335–1361

    2017

  16. [24]

    Hong Ru and Antoinette Schoar. 2016. Do credit card companies screen for behavioral biases? Technical Report. National Bureau of Economic Research

    2016

  17. [25]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15, 1 (2014), 1929–1958

    2014

  18. [26]

    Victor Stango and Jonathan Zinman. 2016. Borrowing high versus borrowing higher: price dispersion and shopping behavior in the US credit card market.The Review of Financial Studies 29, 4 (2016), 979–1006

    2016

  19. [27]

    Sachin Thukral, Suyash Sangwan, Arnab Chatterjee, Lipika Dey, Aaditya Agrawal, Pramit Kumar Chandra, and Animesh Mukherjee. 2022. Understanding how social discussion platforms like Reddit are influencing financial behavior. In 2022 IEEE/WIC/ACM International Joint Conference o...

    2022

  20. [28]

    Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, and Ji...

    2024 arXiv

  21. [29]

    Yaoning Yu, Ye Yu, Haojing Luo, and Haohan Wang. 2026. MiroBench: Bench- marking Realism in Agentic Simulation of Real-world Discussions.arXiv preprint arXiv:2606.14715 (2026)

    2026

  22. [30]

    Hanna Yukhymenko, Robin Staab, Mark Vero, and Martin Vechev. 2024. A syn- thetic dataset for personal attribute inference. Advances in Neural Information Processing Systems 37 (2024), 120735–120779

    2024

  23. [31]

    Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. 2023. SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents. arXiv:2310.11667 [cs.CL] https://arx...

    2023 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.