No One Wins in Nuclear War: A Social Simulation of Military Decision-making

Anthony Wen-Ming Zang; Glenn Matlin; Isaac Song; Mark Riedl

REVIEW 3 major objections 3 minor 22 references

No One Wins in Nuclear War: A Social Simulation of Military Decision-making

T0 review · 3 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A deterministic rules engine makes a nuclear-war card game an auditable social-simulation testbed.

desk verdict A solid, honestly scoped environment paper; the machinery is real and reproducible, but the abstract overstates two things: byte-identical replay across press rungs and full conformance to the published rulebook. read the letter →

arxiv 2608.01868 v1 pith:NHM465F2 submitted 2026-08-03 cs.CY cs.AIcs.CLcs.MA

Glenn Matlin , Isaac Song , Anthony Wen-Ming Zang , Mark Riedl This is my paper

classification cs.CYcs.AIcs.CLcs.MA

keywords socialsimulationdeterministicrulesenginereplayvalidationwargamingnucleardecision-makingcommand-and-controlagentdecisioncontractpressladder

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 tries to establish that high-stakes organizational decision-making can be studied in a simulation where consequences are not left to a model or adjudicator but computed by a deterministic rules engine. It builds that engine for the published card game Nuclear War, traces it against the published rules, and validates it by replaying seeded games byte-for-byte across structural changes. Every strategic choice is surfaced as an explicit agent decision through a decision-point contract, and the same mechanical core can be run with different communication rules and different internal command structures. A sympathetic reader would care because this makes claimed outcomes checkable: any game can be replayed, audited, and compared across conditions.

What carries the argument

The load-bearing mechanism is the replay-validated rules engine plus its decision-point contract: a deterministic state machine over the game's turn phases that reports the current choice, applies only engine-approved LegalActions, runs mandatory steps automatically, and records every state transition to a replay JSON. Replay validation is anchored by a parity invariant—purely structural changes must leave seeded outcomes byte-for-byte identical, and any divergence must be explicitly asserted—checked across a 240-game sweep. The press ladder and faction command-and-control archetypes are layers that leave this core unchanged.

What would settle it

Run a postal-equipment launch under the published rules with the specified six-sided die and compare the launch-failure distribution to the engine's two-d10 spinner on the same seeds; the probabilities diverge, which would show the replay fidelity claim is not conformance to the published rulebook. More broadly, an independent rules check of a recorded replay JSON that finds a transition outside the published rules would falsify the engine's audited-fidelity claim.

Watch

Extended reading notes

Core claim

The central claim is that WOPR supplies the missing verifiable layer in social simulation: a deterministic, replay-validated rules engine that alone provides legal actions, records every state transition, and exposes a fixed contract so any agent or faction can be plugged into the same world. On that engine, a full game of Nuclear War is implemented with mechanics traced to the published rules, and a 240-game sweep shows seeded outcomes are preserved when the engine is structurally changed. The engine surfaces eight strategic decision types; communication capacity is varied on a four-rung press ladder without touching the rules; and each faction can be run as a collective command-and-control

Load-bearing premise

The whole replay-and-audit guarantee rests on the engine actually implementing the published game's rules, but that conformance is self-audited and at least one tracked simplification—postal equipment launches using the base two-d10 spinner instead of the specified six-sided die—changes a published probability.

Editorial extensions

If this is right

Any claimed game outcome can be replayed and checked against the recorded legal state transitions, making agent behavior auditable after the fact.
The same unchanged engine can be run across the four press rungs, so differences in escalation or cooperation can be attributed to communication conditions rather than rule variation.
A faction can be internally a sole authority, council, distributed release, or automated policy, yet remain externally indistinguishable to the engine, enabling direct comparison of command structures on identical games.
Because the decision-point contract is not specific to Nuclear War, the testbed can be extended to other verifiable rule systems.
The 240-game replay sweep gives future changes a concrete regression gate: a structural refactor that silently changes outcomes will fail the parity check.

Reading between the lines

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

If the fidelity claims hold up, this style of engine could become a shared benchmark for comparing human and model crisis behavior under identical, verifiable rules—something the paper gestures at but does not claim.
The press ladder's honor-system commitments, with violation analysis left to post-hoc traces, could be turned into a direct measure of promise-keeping under escalation pressure; the paper records commitments but does not test this.
The four C2 archetypes suggest a concrete experiment the paper leaves implicit: holding agents and press rung fixed, vary sole-authority versus council structure and measure release decisions; the engine's replay artifact would make that comparison exact.
A natural extension beyond war games: any rule-governed high-stakes scenario—emergency response, financial crisis management—could reuse the same contract if its rules can be encoded deterministically.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: WOPR is a genuinely useful environment paper, not a behavioral study, and it is honest about that. The new pieces are the deterministic, replay-validated rules engine, the decision-point contract that exposes only legal actions to agents, the four-rung press ladder, and the C2 aggregation layer. The code, configs, replay data, and the 240-game parity sweep are public and well-engineered. The paper also correctly identifies a real gap in the social-simulation literature: most prior work lacks a verifiable engine with auditable transitions. Credit where due: the reproducibility statement is serious, the test gate is real, and Appendix D tracks simplifications instead of hiding them.

Now the soft spots, in proportion. First, the claim in Section 4 and Appendix E that press "leaves the replay JSON byte-identical" to a no-press game does not hold as written. If press messages change what agents choose, those choices are recorded in the replay JSON, so the content cannot be identical. What is presumably meant is that the JSON schema is unchanged and no additional fields are added; that is worth stating accurately. This looks like a fixable wording problem, but as it stands it is internally inconsistent.

Second, rule fidelity. The abstract says the game is "traced against its published rules," but Appendix D records that postal equipment launch resolution uses the base two-d10 fallout spinner instead of the published six-sided Radioactive Fallout die, changing launch-failure probability. The parity sweep proves refactor stability, not rulebook conformance. The conformance matrix and single-game rules trace are self-authored, not an external oracle. None of this is fatal—the paper explicitly disclaims one-to-one reproduction—but the load-bearing novelty claim is that rules are checkable against a published source, and that claim is currently stronger than the evidence. Soften the phrasing or fix the die.

Third, the paper is early-stage: no controlled comparison across press rungs, only council C2 has a worked example, and there are no empirical claims about agent behavior. The authors say this themselves, so I do not count it heavily.

Overall: the engineering is real, the limitations are mostly visible, and the two overstated claims are localized and fixable. This deserves a serious referee—a workshop or journal editor should send it out rather than desk reject. I would bring it to a reading group interested in LLM-based social simulation and wargaming infrastructure, and I would cite it if I were working on reproducible social-simulation benchmarks. My verdict: conditional accept, contingent on the byte-identical claim being clarified and the postal-equipment simplification either resolved or explicitly excluded from the fidelity claim.

Referee Report

3 major / 3 minor

Summary. WOPR proposes a deterministic, replay-validated rules engine for social simulation, instantiated with the published card game Nuclear War. The paper's contributions are (1) the engine and its decision-point contract, which exposes every strategic choice as an explicit agent decision; (2) a four-rung press ladder (no-press, press-light, multi-turn public, full press) intended to vary communication while leaving the rules unchanged; and (3) four command-and-control archetypes (sole-authority, council, distributed, automated) that aggregate multiple member agents into one legal action. The paper reports a 240-game replay-parity sweep and a rules-conformance matrix, and states that all code and replay artifacts are public. It explicitly frames the work as early-stage, with no controlled experimental comparison across press rungs or C2 archetypes.

Significance. If the central claims hold, WOPR addresses a real gap in generative-agent social simulation: most prior work relies on model- or adjudicator-driven mechanics, whereas WOPR provides an auditable, deterministic engine with a clean agent interface. The released code, test suite, and replay artifacts are valuable infrastructure, and the decision-point contract is a reusable design. However, the paper's load-bearing claims are currently stronger than the evidence: "replay-validated" refers to internal refactor stability, and "traced against its published rules" is qualified by a documented simplification that changes a published probability. The press-ladder section also contains an internal inconsistency about replay JSON identity. These issues do not destroy the value of the engine as a reproducible simulation environment, but they require correction before the claims can be accepted as stated.

major comments (3)

[Section 4 and Appendix E] The paper repeatedly states that press "never mutates game state" and "leaves the replay JSON identical across rungs" (Section 4) or "byte-identical to a no-press game" (Appendix E). This is internally inconsistent with the stated purpose of the press ladder: press messages are injected into agents' memory (Appendix E: "which the system injects into agents' memory for the next round") and are intended to influence their decisions. If an agent changes its choice because of a message, the recorded legal action changes and the replay JSON necessarily differs. The only defensible statement is that speech is not an engine state transition; the causal effect on agent choices is reflected in the actions recorded in the replay. This needs rewriting, because the claim is central to the communication-axis contribution.
[Appendix D] The conformance appendix records that postal equipment launch resolution (space platforms, cruise missiles, killer satellites, Space Shuttle attacks, and MX per-segment rolls) uses the base two-d10 fallout spinner instead of the published six-sided Radioactive Fallout die, and explicitly states that "the launch-failure probability differs." This directly conflicts with the abstract's claim that the engine is "traced against its published rules." The engine is a variant with a tracked simplification, not a full implementation of the published game for that mechanic. Either implement the six-sided die, or soften the abstract and contribution statements to "traced for implemented mechanics" with a clearly documented divergence. Rule fidelity is the stated novelty, so this is a load-bearing issue, not a cosmetic one.
[Appendix B and Section 2] The replay-validation method has two anchors: a parity invariant and a 240-game sweep. The parity invariant proves that structural changes to the engine preserve seeded outcomes; it does not compare engine behavior to the published rulebook. The rules-trace summary covers one game and is self-authored, not an external oracle. Thus "replay-validated" currently conflates internal determinism/reproducibility with external rule conformance. The paper should explicitly distinguish "replay-validated engine" (seeded re-execution and refactor stability) from "rule-fidelity-verified against a published source," and ideally include an independent audit or a larger, externally checkable trace set. Without this distinction, readers may over-credit the fidelity claim.

minor comments (3)

[Abstract / Section 1] The phrase "traced against its published rules" appears in the abstract and introduction, but Appendix D later says "We do not claim a one-to-one reproduction." Consider aligning the language up front to avoid the impression of full rule fidelity.
[Section 4] The description of press-light says "each living agent may make one public statement per round." It would be clearer to state whether a "round" is a full turn cycle or a single decision point; the examples later use "per pass" and "per round" interchangeably.
[Appendix F, Table 2] The council example uses threshold 0.34, which is just above the equal-weight share of 1/3. The prose says "no option clears the threshold" and the default is the first-cast vote. This is correct, but it would help to state explicitly that all three members have equal weight in the example.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: replay-parity checks are internal consistency tests, not predictions derived from fitted inputs; the Appendix D rule simplification is an acknowledged fidelity limit, not a circular step.

full rationale

The paper's central contribution is a deterministic rules engine and a decision-point contract, not a fitted model that makes predictions. The 240-game sweep validates that structural changes preserve seeded outcomes relative to the prior engine head; this is an internal self-consistency check, not a derivation of a result from its own inputs. The conformance matrix and rules trace are self-authored, and Appendix D explicitly records a tracked simplification (postal equipment launch resolution uses the base two-d10 fallout spinner where the published rules specify a six-sided Radioactive Fallout die, changing launch-failure probability). This weakens the strength of the 'traced against its published rules' claim as a correctness or fidelity matter, but it does not make the argument circular. No load-bearing self-citations appear; the reference list contains no overlapping authors with the present paper. No fitted parameter is renamed as a prediction, no uniqueness claim is imported from the authors' prior work, and no known result is relabeled as a new framework. The paper's own limitations—early-stage framing, no controlled comparison, single-game press modes, unenforced commitments, and only a council worked example—are disclosed transparently. Under the strict definition of circularity, no step reduces an output to an input by construction or by definition, so the appropriate finding is no significant circularity.

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

The central claim does not rest on fitted constants, but it does rest on self-referential validation artifacts: a self-authored conformance matrix, an internal parity sweep, and a special baseline-agent convention. The C2 configuration parameters are user-set knobs rather than fitted values. The invented entities are software abstractions, and only the decision-point contract and press ladder have shipped evidence; the C2 archetypes are mostly promissory.

free parameters (4)

council threshold = 0.5 default
Config parameter for the council archetype, chosen by hand rather than fitted to data; the worked trace uses 0.34.
sole-authority deference = 0 or 1
Config parameter controlling whether the executive defers to the first advisor; user-set, not fitted.
distributed quorum = not specified
Number of distinct release actions required for the distributed archetype to act; user-set, not fitted.
automated policy action id = no default
Pre-armed action required for the automated archetype; user-set, not fitted.

assumptions (4)

domain assumption The self-authored conformance matrix correctly traces implemented mechanics to the published rules of Nuclear War.
Appendix D presents the matrix as the evidence for rule fidelity, but it is not machine-checked against an external rules oracle and includes tracked simplifications.
domain assumption The 240-game sweep with 3/4 players, heuristic/random agents, and seeds 1-60 is sufficient to establish replay parity and determinism.
Appendix B describes the sweep as a sample that preserves prior-head outcomes; it is not exhaustive over the full decision space, and it excludes the LLM agents that are the paper's motivating use case.
ad hoc to paper The baseline heuristic picking options[0] reproduces the deterministic engine policy without consuming RNG draws.
Appendix A/B introduces this design convention to make the parity invariant checkable; it is specific to this engine.
domain assumption Press speech never mutates engine state, and replay JSON is unchanged by press.
Section 4 and Appendix E assert this invariant, but it only holds if agent action choices are fixed; otherwise different press-induced decisions change the recorded transition sequence.

invented entities (3)

Decision-point contract independent evidence
purpose: Defines the observe/choose/apply interface between the rules engine and any agent or composite faction.
The contract is exposed in code and replay artifacts, giving a falsifiable handle independent of the paper's prose.
Four-rung press ladder independent evidence
purpose: Adds a communication axis (silence, one public message, multi-turn public, private channels with commitments) on an unchanged rules engine.
Example configs and a press-trace sidecar are shipped, but no batch evaluation exists yet, so evidence is limited to machinery.
C2 archetypes (sole-authority, council, distributed, automated)
purpose: Composite faction command-and-control structures that collapse member votes into one outward legal action.
Only the council archetype has a worked trace; the other three are defined but not demonstrated with recorded examples.

how reviews work

0 comments

Cite this review

Pith. "Pith review of No One Wins in Nuclear War: A Social Simulation of Military Decision-making." pith.science (2026). https://pith.science/paper/NHM465F2

@misc{pith2026260801868,
  author       = {Pith},
  title        = {Pith review of: No One Wins in Nuclear War: A Social Simulation of Military Decision-making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHM465F2}},
  note         = {Machine review of arXiv:2608.01868}
}

read the original abstract

WOPR is a social-simulation environment for studying how organizations make high-stakes decisions, built on a deterministic, replay-validated rules engine and using wargames as the vehicle. We instantiate it first with the published card game Nuclear War, traced against its published rules. We start with military decision-making because of its safety implications and because it needs further study, but the design is not specific to it: the decision-point contract that exposes the engine to agents is reusable across verifiable rule systems. Existing social-simulation work emphasizes persona fidelity and synthetic opinion, but lacks a verifiable rules engine with replay-checkable mechanics and private-channel negotiation. WOPR supplies that engine, and its contract makes every strategic choice an explicit agent decision. The method is agnostic to social-simulation frameworks; we adopt Concordia as the default harness for driving the game. On the same engine, WOPR layers a four-rung press ladder from silence to private single-recipient channels with structured commitments, and instantiates each faction as a collective command-and-control system rather than a single agent. We make all code, example configurations, and replay data publicly available at https://github.com/eilab-gt/wopr.

Figures

Figures reproduced from arXiv: 2608.01868 by the authors.

**Figure 2.** The turn-phase cursor shown as a deterministic cycle. Mandatory steps (green) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Replay workbench table view for a cards-enqueued event. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Population chart and event timeline for the same replay. At turn 0 the top line is [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Decision-trace inspector showing the selected action context. The trace shown [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Conversation inspector for a full-press game. The public statements shown carry [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Forensic inspector exposing replay JSON for audit. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Batch analysis view for the bundled three-run sample. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 9.** Figure 9: Live workbench connected to the local rules API. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: The press ladder as a communication-capacity axis on an unchanged rules engine. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

**Figure 11.** Figure 11: Four command-and-control archetypes a faction may instantiate, distinguished [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 4 linked inside Pith

[1]

and Cai, Carrie J

Park, Joon Sung and O'Brien, Joseph C. and Cai, Carrie J. and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S. , title =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST) , publisher =. 2023 , doi =

2023
[2]

and Aharon, Avia and Ziv, Ron and Matyas, Jayd and Duéñez-Guzmán, Edgar A

Vezhnevets, Alexander Sasha and Agapiou, John P. and Aharon, Avia and Ziv, Ron and Matyas, Jayd and Duéñez-Guzmán, Edgar A. and Cunningham, William A. and Osindero, Simon and Karmon, Danny and Leibo, Joel Z. , title =. doi:10.48550/ARXIV.2312.03664 , url =

arXiv doi:10.48550/arxiv.2312.03664
[3]

The Fourteenth International Conference on Learning Representations (ICLR) , year =

Hu, Tiancheng and Baumann, Joachim and Lupo, Lorenzo and Collier, Nigel and Hovy, Dirk and Röttger, Paul , title =. The Fourteenth International Conference on Learning Representations (ICLR) , year =
[4]

doi:10.48550/ARXIV.2505.23713 , url =

Xu, Zixiang and Wang, Yanbo and Huang, Yue and Ye, Jiayi and Zhuang, Haomin and Song, Zirui and Gao, Lang and Wang, Chenxi and Chen, Zhaorun and Zhou, Yujun and Li, Sixian and Pan, Wang and Zhao, Yue and Zhao, Jieyu and Zhang, Xiangliang and Chen, Xiuying , title =. doi:10.48550/ARXIV.2505.23713 , url =

arXiv doi:10.48550/arxiv.2505.23713
[5]

Social Sim'25 Workshop at COLM 2025 , year =

Mangla, Shashank and Hokamp, Chris and Boylan, Jack and Ghalandari, Demian Gholipour and Jauhari, Yuuv and Cassidy, Lauren and Duffy, Oisin , title =. Social Sim'25 Workshop at COLM 2025 , year =

2025
[6]

Advances in Neural Information Processing Systems 37 (NeurIPS) , year =

Piatti, Giorgio and Jin, Zhijing and Kleiman-Weiner, Max and Schölkopf, Bernhard and Sachan, Mrinmaya and Mihalcea, Rada , title =. Advances in Neural Information Processing Systems 37 (NeurIPS) , year =
[7]

Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) , pages =

Lamparth, Max and Corso, Anthony and Ganz, Jacob and Mastro, Oriana Skylar and Schneider, Jacquelyn and Trinkunas, Harold , title =. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) , pages =. 2024 , doi =

2024
[8]

Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT) , pages =

Rivera, Juan-Pablo and Mukobi, Gabriel and Reuel, Anka and Lamparth, Max and Smith, Chandler and Schneider, Jacquelyn , title =. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT) , pages =. 2024 , doi =

2024

Show all 22 references

[9]

doi:10.48550/ARXIV.2311.17227 , url =

Hua, Wenyue and Fan, Lizhou and Li, Lingyao and Mei, Kai and Ji, Jianchao and Ge, Yingqiang and Hemphill, Libby and Zhang, Yongfeng , title =. doi:10.48550/ARXIV.2311.17227 , url =

doi arXiv
[10]

Science , volume =

Human-level play in the game of. Science , volume =. 2022 , doi =

2022
[11]

and Brennen, Andrea , title =

Hogan, Daniel P. and Brennen, Andrea , title =. doi:10.48550/ARXIV.2404.11446 , url =

doi arXiv
[12]

2025 , url =

Deshpande, Anushka and Jin, Zhijing , booktitle =. 2025 , url =

2025
[13]

, title =

Hoffman, David E. , title =. 2009 , isbn =

2009
[14]

and Tertrais, Bruno , title =

Lewis, Jeffrey G. and Tertrais, Bruno , title =. 2019 , url =

2019
[15]

and Fink, Anya L

Woolf, Amy F. and Fink, Anya L. , title =. 2023 , url =

2023
[16]

Understanding Cyber Conflict: 14 Analogies , editor =

Feaver, Peter and Geers, Kenneth , title =. Understanding Cyber Conflict: 14 Analogies , editor =. 2017 , isbn =

2017
[17]

2012 , isbn =

Khan, Feroz Hassan , title =. 2012 , isbn =

2012
[18]

2017 , isbn =

Salik, Naeem , title =. 2017 , isbn =

2017
[19]

, title =

Yarynich, Valery E. , title =. 2003 , isbn =

2003
[20]

2010 , howpublished =

2010
[21]

1956 , howpublished =

National Security Council Report,. 1956 , howpublished =

1956
[22]

2002 , howpublished =

Cotta-Ramusino, Paolo and Martellini, Maurizio , title =. 2002 , howpublished =

2002

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

and Cai, Carrie J

Park, Joon Sung and O'Brien, Joseph C. and Cai, Carrie J. and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S. , title =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST) , publisher =. 2023 , doi =

2023

[2] [2]

and Aharon, Avia and Ziv, Ron and Matyas, Jayd and Duéñez-Guzmán, Edgar A

Vezhnevets, Alexander Sasha and Agapiou, John P. and Aharon, Avia and Ziv, Ron and Matyas, Jayd and Duéñez-Guzmán, Edgar A. and Cunningham, William A. and Osindero, Simon and Karmon, Danny and Leibo, Joel Z. , title =. doi:10.48550/ARXIV.2312.03664 , url =

arXiv doi:10.48550/arxiv.2312.03664

[3] [3]

The Fourteenth International Conference on Learning Representations (ICLR) , year =

Hu, Tiancheng and Baumann, Joachim and Lupo, Lorenzo and Collier, Nigel and Hovy, Dirk and Röttger, Paul , title =. The Fourteenth International Conference on Learning Representations (ICLR) , year =

[4] [4]

doi:10.48550/ARXIV.2505.23713 , url =

Xu, Zixiang and Wang, Yanbo and Huang, Yue and Ye, Jiayi and Zhuang, Haomin and Song, Zirui and Gao, Lang and Wang, Chenxi and Chen, Zhaorun and Zhou, Yujun and Li, Sixian and Pan, Wang and Zhao, Yue and Zhao, Jieyu and Zhang, Xiangliang and Chen, Xiuying , title =. doi:10.48550/ARXIV.2505.23713 , url =

arXiv doi:10.48550/arxiv.2505.23713

[5] [5]

Social Sim'25 Workshop at COLM 2025 , year =

Mangla, Shashank and Hokamp, Chris and Boylan, Jack and Ghalandari, Demian Gholipour and Jauhari, Yuuv and Cassidy, Lauren and Duffy, Oisin , title =. Social Sim'25 Workshop at COLM 2025 , year =

2025

[6] [6]

Advances in Neural Information Processing Systems 37 (NeurIPS) , year =

Piatti, Giorgio and Jin, Zhijing and Kleiman-Weiner, Max and Schölkopf, Bernhard and Sachan, Mrinmaya and Mihalcea, Rada , title =. Advances in Neural Information Processing Systems 37 (NeurIPS) , year =

[7] [7]

Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) , pages =

Lamparth, Max and Corso, Anthony and Ganz, Jacob and Mastro, Oriana Skylar and Schneider, Jacquelyn and Trinkunas, Harold , title =. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) , pages =. 2024 , doi =

2024

[8] [8]

Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT) , pages =

Rivera, Juan-Pablo and Mukobi, Gabriel and Reuel, Anka and Lamparth, Max and Smith, Chandler and Schneider, Jacquelyn , title =. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT) , pages =. 2024 , doi =

2024