Pith. sign in

REVIEW 2 major objections 5 minor 33 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

LLM social simulations get a classical-ABM reality check

2026-07-08 19:11 UTC pith:PQRVRBU6

load-bearing objection AgoraSim: A Hybrid Agent-Based Modeling Framework the 2 major comments →

arxiv 2607.05999 v1 pith:PQRVRBU6 submitted 2026-07-07 cs.AI

AgoraSim: A Hybrid Agent-Based Modeling Framework

Chung-Chi Chen This is my paper
classification cs.AI
keywords agorasimmodelingsocialagent-basedagentsclassicalcomparedynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

AgoraSim is a hybrid agent-based modeling framework that lets LLM-powered agents, vision-language agents, custom-endpoint agents, random agents, and classical rule-based agents coexist in the same simulation. The central mechanism is a shared structured decision object: every agent type, whether it reasons through language or follows a mathematical rule, emits the same compact set of fields — stance, sentiment, selected action, confidence, public statement, private rationale, and survey answers. This shared schema means that free-form LLM behavior is projected into a countable, comparable state space alongside classical dynamics. The paper argues that this projection is the necessary bridge between NLP-driven social simulation and traditional agent-based modeling, because it allows the same scenario to run under identical conditions — same action space, network, seed, interaction protocol, and metrics — whether the agents are language models or threshold rules. The framework resolves a natural-language or multimodal artifact (a policy announcement, an advertisement, a video) into an editable simulation configuration, then runs ratio-controlled populations where the mix of agent types is itself an experimental variable. Classical ABM reference dynamics — threshold/Bass adoption, bounded confidence, SIR contagion, herding, DeGroot/voter learning, and discrete choice — serve not as weak baselines but as interpretive anchors. If an LLM-agent trajectory closely follows a simple threshold model, the observed pattern may not require rich language understanding to explain. If it diverges, the divergence becomes a hypothesis about whether the artifact's wording, the model provider's bias, or network amplification is driving the difference. The paper explicitly disclaims predictive validity: AgoraSim is positioned as a scenario exploration and comparison tool, not a population predictor.

Core claim

The paper's central construction is the shared structured decision object that all agent types emit, which makes LLM free-form text and classical rule outputs directly comparable under the same experimental conditions. This is not a claim about a new algorithm or a proven behavioral result; it is a framework design claim that the right way to evaluate LLM-based social simulations is to run them alongside matched classical ABM reference dynamics — same scenario, same action space, same seed, same metrics — and treat agreement or divergence as diagnostic information about whether language understanding is actually contributing to the observed trajectory.

What carries the argument

The shared structured decision object (stance, sentiment, action, confidence, public statement, private rationale, survey answers, debug fields) is the central mechanism. Around it sit three components: the scenario workbench that resolves text or media into editable ABM configurations, the hybrid ABM engine that assigns agents to ratio-controlled model slots, and the comparison controller that launches matched classical reference runs. The action space — compact labels like SUPPORT, OPPOSE, WAIT, ADAPT — is the bridge through which both LLM reasoning and classical rule updates are projected into a common state representation.

Load-bearing premise

The framework assumes that projecting LLM free-form responses into a compact shared action space — labels like SUPPORT, OPPOSE, WAIT — preserves enough behavioral information to make the comparison with classical ABM dynamics meaningful. If this mapping strips out the nuanced reasoning, conditional stances, and deliberation variance that make LLM agents interesting, then the comparison is between a lossy projection of LLM behavior and full classical behavior, which would bias

What would settle it

Run a scenario where LLM agents produce rich, conditional, multi-step reasoning that clearly differs from simple threshold or contagion dynamics in the full text. Then check whether the projected action labels collapse that difference — if the LLM trajectory matches a classical reference under the shared action space but diverges in the raw text, the projection is lossy and the comparison framework's core diagnostic is compromised.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the framework's comparison logic holds, researchers can determine whether observed social simulation patterns are genuinely produced by language understanding or are artifacts of simple diffusion dynamics that any classical model would reproduce.
  • The ratio-controlled model mix turns agent-type composition into an experimental variable: a population that is 30% LLM and 70% classical can be compared against 70% LLM and 30% classical, isolating the marginal effect of language reasoning on collective dynamics.
  • The audit trail — prompts, parsed decisions, heard neighbor statements, costs, rule diagnostics — makes it possible to trace any aggregate trajectory back to individual agent decisions, which could serve as a debugging tool for identifying when LLM agents collapse toward population-typical responses.
  • The framework's separation of agent behavior from interaction protocol means the same scenario can test whether results are robust to different exposure structures (independent survey vs. social diffusion vs. broadcast stream), which could reveal whether LLM simulation findings in the literature are artifacts of a particular network assumption.
  • The local deployment model with user-owned API keys and no key persistence makes the framework usable for sensitive scenario analysis where sending content to external providers is a concern, though the paper notes that live LLM calls still transmit prompts to providers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The shared action space is a lossy projection: mapping LLM free-form reasoning into labels like SUPPORT or OPPOSE may discard the conditional reasoning, nuanced stances, and multi-step deliberation that make LLM agents worth studying. If this projection strips out the variance that distinguishes LLM behavior from classical behavior, then the comparison framework could systematically make LLM and c
  • The framework could be extended to test the projection itself: run the same LLM agents with progressively richer action spaces (binary, ternary, continuous, free-form with post-hoc clustering) and measure at what granularity LLM trajectories begin to diverge from classical references. This would identify the resolution at which language understanding becomes behaviorally detectable.
  • The comparison logic implicitly assumes that classical reference dynamics span the space of plausible social dynamics. If there exist collective patterns that no classical model in the current set (threshold, Bass, bounded confidence, SIR, herding, DeGroot, discrete choice) can reproduce, then a divergence between LLM and all classical references would be uninterpretable — it could mean the LLM is

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper presents AgoraSim, a hybrid agent-based modeling framework that integrates LLM, vision-language, custom-endpoint, random, and classical rule-based agents under a shared structured decision schema. The system resolves natural-language or multimodal artifacts into editable ABM configurations, runs ratio-controlled mixed populations, and launches matched classical reference dynamics (threshold/Bass, bounded confidence, SIR contagion, herding, DeGroot/voter, discrete choice) under the same action space, interaction protocol, metrics, and seed. The framework is exposed through a local UI, Python SDK/CLI, and REST API. The paper is framed as a system/demo description and explicitly disclaims predictive validity. The congestion-pricing example in §4 illustrates the workflow narratively. Appendices document the configuration schema, interaction protocols, classical reference parameters, and SDK/CLI/REST usage.

Significance. The core design contribution—placing LLM agents and classical ABM reference dynamics in the same action space, interaction protocol, and metric schema so that trajectories are directly comparable—is a well-motivated response to a real methodological gap in LLM-based social simulation. The inclusion of six standard classical reference families (Table 1, Table 4) with editable parameters, the separation of interaction protocols from agent behavior (Table 3), and the full configuration schema (Table 2) are concrete design artifacts. The public code package and demo video, if functional and maintained, add reproducibility value. The paper is honest about limitations and does not overclaim predictive authority. However, the significance is substantially undercut by the complete absence of any empirical output from the system: no trajectory plots, no comparison metrics, no divergence/convergence analysis, and no evidence that the classical reference dynamics are correctly implemented or that the comparison metrics discriminate meaningfully between trajectory types.

major comments (2)
  1. §4 (Example Scenario): The congestion-pricing example is described entirely narratively. No actual results are presented—no trajectory plots, no comparison metrics (e.g., distance between hybrid and classical trajectories), no stance distributions over rounds, and no audit-record excerpts. The paper's central value proposition is 'direct comparison of scenario trajectories against matched classical ABM reference dynamics' (abstract, §1, §2.3), but this comparison is never shown. For a system/demo paper, at least one concrete worked example with populated results is essential to demonstrate that the framework produces meaningful comparative insights and that the classical reference implementations are correct. Without this, the reader cannot assess whether the system works as described or whether the comparison is informative.
  2. §2.1, action space design: The shared structured decision object (stance, sentiment, selected action, confidence, public statement, private rationale, sharing intent, survey answers, debug fields) is the load-bearing bridge between LLM free-form output and classical ABM state. The paper states that 'open-ended language behavior is projected into an auditable state space' but does not specify the projection mechanism: how are LLM responses parsed into discrete actions (SUPPORT, OPPOSE, WAIT, ADAPT)? Is this done via prompt instructions, a separate classifier, or structured output constraints? The fidelity of this projection is critical—if it is lossy or inconsistent, the comparison with classical dynamics is compromised. The paper should describe the mapping mechanism and, ideally, report inter-rater or test-retest reliability for the action extraction.
minor comments (5)
  1. Table 1 and Table 4 overlap significantly in content (both describe the six classical reference dynamics). Consider merging or cross-referencing more explicitly to avoid redundancy.
  2. §2.2: The list of interaction protocols is mentioned in the text and detailed in Appendix B (Table 3), but the main text does not forward-reference the appendix. A brief pointer would help readers.
  3. Appendix D: The SDK example uses run_experiment with a budget_usd parameter, but the configuration schema in Table 2 does not list budget as a core field. Clarify whether budget is part of the resolved configuration or only a run-time constraint.
  4. The paper references Wu et al. (2026) for the claim that LLM agents 'collapse toward high-authority or population-typical responses.' This is cited multiple times (§1, §5 Impact Statement). Ensure the citation is accessible and the claim is accurately represented.
  5. Figures 2 and 3 captions could note that the screenshots show the UI before execution (no results populated), to set reader expectations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for a careful and constructive reading. Both major comments identify genuine gaps in the current manuscript: the congestion-pricing example in §4 is entirely narrative with no empirical output, and the projection mechanism mapping LLM responses to discrete actions is underspecified. We agree with both points and will revise accordingly.

read point-by-point responses

  1. Referee: §4 (Example Scenario): The congestion-pricing example is described entirely narratively. No actual results are presented—no trajectory plots, no comparison metrics, no stance distributions over rounds, and no audit-record excerpts. The paper's central value proposition is 'direct comparison of scenario trajectories against matched classical ABM reference dynamics,' but this comparison is never shown. For a system/demo paper, at least one concrete worked example with populated results is essential.

    Authors: The referee is correct. The congestion-pricing example in §4 is described narratively and does not present any actual system output—no trajectory plots, no comparison metrics, no stance distributions, and no audit-record excerpts. This is a genuine gap: the paper's central value proposition is direct comparison of hybrid LLM-agent trajectories against matched classical ABM reference dynamics, and the current manuscript never demonstrates that comparison with concrete results. We will revise §4 to include at least one fully worked example with: (1) stance distribution plots over rounds for the hybrid run and the matched classical reference runs (threshold/Bass and discrete choice); (2) trajectory comparison metrics quantifying the distance between hybrid and classical trajectories; (3) at least one audit-record excerpt showing the per-agent decision object, public statement, and parsed action. We will also verify that the classical reference implementations produce expected dynamics on pure-classical runs (e.g., threshold model producing S-curve adoption, SIR producing bell-shaped contagion curve) and report these as implementation validation. This revision is essential and we commit to it. revision: yes

  2. Referee: §2.1, action space design: The shared structured decision object is the load-bearing bridge between LLM free-form output and classical ABM state. The paper states that 'open-ended language behavior is projected into an auditable state space' but does not specify the projection mechanism: how are LLM responses parsed into discrete actions (SUPPORT, OPPOSE, WAIT, ADAPT)? Is this done via prompt instructions, a separate classifier, or structured output constraints? The fidelity of this projection is critical—if it is lossy or inconsistent, the comparison with classical dynamics is compromised. The paper should describe the mapping mechanism and, ideally, report inter-rater or test-retest reliability for the action extraction.

    Authors: The referee correctly identifies that the projection mechanism is underspecified in the current manuscript. In the implementation, LLM agents are prompted to return a structured JSON object that includes the discrete action field (e.g., SUPPORT, OPPOSE, WAIT, ADAPT) alongside free-form fields (public statement, private rationale). The action space is enumerated in the prompt, and the model is instructed to select exactly one action from the allowed set. This is primarily a prompt-instructed structured-output approach rather than a separate post-hoc classifier. We agree that this mechanism and its fidelity need to be explicitly described in the manuscript. In the revision we will: (1) document the projection mechanism in §2.1, including the prompt structure and how the action space is communicated to LLM agents; (2) note the fallback behavior when an LLM response does not conform to the expected schema (retry, default, or flag as degraded); (3) report test-retest reliability for action extraction by running the same scenario with the same seed and temperature=0 across multiple trials and reporting the consistency rate of extracted actions. We agree that without this information the reader cannot assess whether the comparison with classical dynamics is meaningful. Regarding inter-rater reliability: because the action is self-reported by the LLM agent rather than extracted by an external annotator, classical inter-rater reliability does not directly apply, but we will report the analogous consistency metric (test-retest agreement under fixed seeds) and discuss the limitation that self-reported actions may not match what a human coder would assign from the free-text statement. revision: yes

Circularity Check

0 steps flagged

No circularity detected: framework description with externally-sourced classical ABM references and no fitted predictions or self-citation chains

full rationale

AgoraSim is a system/demo paper, not a derivation paper. It makes no first-principles predictions, fits no parameters to data and then 'predicts' related quantities, and invokes no self-cited uniqueness theorem or ansatz. The classical ABM reference dynamics (threshold/Bass, bounded confidence/Deffuant, SIR contagion, herding, DeGroot/voter, discrete choice) are all standard models from the external literature (Schelling 1971, Bass 1969, DeGroot 1974, Deffuant et al. 2000, Granovetter 1978, Clifford and Sudbury 1973), not results authored by the present paper's author. The shared structured decision object (stance, sentiment, action, confidence, etc.) is a design artifact, not a derived result that is defined in terms of its own inputs. The congestion-pricing example in §4 is explicitly narrative—no trajectories, metrics, or comparison results are shown—but this is a gap in empirical demonstration, not circularity. The paper's central claim (that the framework enables comparison between LLM-agent runs and classical reference dynamics under shared action spaces, protocols, metrics, and seeds) is a design claim about the system's architecture, not a derived result that reduces to its inputs by construction. The absence of demonstrated output is a correctness/evaluation risk, not a circularity risk.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

The axiom ledger is moderate because this is a system/demo paper, not a theoretical derivation. The free parameters are user-controlled configuration options rather than fitted constants. The key axioms are domain assumptions about the meaningfulness of the action-space projection and the classical-ABM comparison, neither of which is empirically validated. No invented physical or mathematical entities are introduced; the 'invented entities' are software architectural components.

free parameters (4)
  • model_mix ratios = user-specified
    The ratio of LLM to classical to random agents is a free parameter set by the user for each run. See §2.2.
  • classical ABM parameters (threshold β, Bass p/q, confidence bound, transmission probability, herding strength, neighbor = demonstration defaults
    Appendix C Table 4 lists editable parameters for each classical reference dynamic. The paper states 'The default values are demonstration defaults rather than calibrated domain-specific theories.'
  • network topology parameters = user-specified
    Network topology and neighbor sampling settings are part of the resolved configuration (Appendix A, Table 2) but no specific topology or parameter values are specified in the paper.
  • action space vocabulary = scenario-dependent
    The compact set of mutually interpretable actions (e.g., SUPPORT, OPPOSE, WAIT, ADAPT) is generated per scenario by the resolver. See §2.1.
axioms (3)
  • domain assumption LLM agents can be meaningfully projected into a compact discrete action space without losing the behavioral variance that makes them interesting.
    This is the core unstated assumption behind the shared decision object design. The paper does not validate it. See §2.1.
  • domain assumption Classical ABM reference dynamics (threshold, Bass, Deffuant, SIR, herding, DeGroot/voter, discrete choice) provide meaningful comparison baselines for LLM-agent trajectories.
    The paper asserts this in §1 and §2.3 but provides no empirical evidence that the comparison produces the interpretive value claimed.
  • domain assumption The resolver can accurately translate natural-language or multimodal artifacts into editable ABM configurations.
    The scenario workbench depends on this translation being useful. See §2.1. No evaluation of resolver quality is provided.
invented entities (2)
  • AgoraSim framework no independent evidence
    purpose: Hybrid agent-based modeling for scenario-oriented social reaction analysis
    The framework itself is the contribution. It is a software system, not a theoretical entity. Evidence of its utility would require empirical demonstration, which the paper does not provide.
  • Shared structured decision object no independent evidence
    purpose: Common output schema enabling all agent types to be compared on the same metrics
    The decision object schema (stance, sentiment, selected action, confidence, public statement, private rationale, sharing intent, survey answers, debug fields) is an invented architectural entity. Its adequacy is not validated.

pith-pipeline@v1.1.0-glm · 12584 in / 2974 out tokens · 290674 ms · 2026-07-08T19:11:32.422493+00:00 · methodology

0 comments
read the original abstract

LLM-agent simulations make natural-language social scenarios easy to instantiate, but their outputs can be overread as predictions and are often difficult to compare with explicit social dynamics. We present AgoraSim, a hybrid agent-based modeling framework for scenario-oriented social reaction analysis. AgoraSim resolves textual or multimodal artifacts into editable ABM configurations, runs ratio-controlled populations that mix LLM, vision-language, custom-endpoint, random, and classical agents, and compares the same scenario against matched classical reference dynamics. All agents emit a shared structured decision object, enabling common action spaces, interaction protocols, metrics, and audit records. Exposed through a local UI, Python SDK/CLI, and REST API, AgoraSim helps users inspect scenario trajectories, compare modeling assumptions, and identify cases that warrant empirical validation.

Figures

Figures reproduced from arXiv: 2607.05999 by Chung-Chi Chen.

Figure 1
Figure 1. Figure 1: AgoraSim architecture. A natural-language or multimodal artifact is resolved into an editable simulation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Creating a scenario experiment. Users enter a natural-language or multimodal artifact, edit the generated [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison and inspection. Users select classical ABM reference dynamics with editable parameters, then [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 3 internal anchors

  1. [1]

    Management Science , volume =

    A New Product Growth for Model Consumer Durables , author =. Management Science , volume =

    work page

  2. [2]

    American Journal of Sociology , volume =

    Threshold Models of Collective Behavior , author =. American Journal of Sociology , volume =

    work page

  3. [3]

    Advances in Complex Systems , volume =

    Mixing Beliefs among Interacting Agents , author =. Advances in Complex Systems , volume =

    work page

  4. [4]

    Wilensky, Uri , year =

    work page

  5. [5]

    Proceedings of the 14th Python in Science Conference , year =

    Mesa: An Agent-Based Modeling Framework , author =. Proceedings of the 14th Python in Science Conference , year =

    work page

  6. [6]

    Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , year =

    Generative Agents: Interactive Simulacra of Human Behavior , author =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , year =

    work page

  7. [7]

    Political Analysis , volume =

    Out of One, Many: Using Language Models to Simulate Human Samples , author =. Political Analysis , volume =

    work page

  8. [8]

    Proceedings of Machine Learning Research , volume =

    Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies , author =. Proceedings of Machine Learning Research , volume =

    work page

  9. [9]

    2023 , institution=

    Large language models as simulated economic agents: What can we learn from homo silicus? , author=. 2023 , institution=

    work page 2023

  10. [10]

    , title =

    Schelling, Thomas C. , title =. The Journal of Mathematical Sociology , volume =. 1971 , doi =

    work page 1971

  11. [11]

    Journal of Conflict Resolution , volume =

    Axelrod, Robert , title =. Journal of Conflict Resolution , volume =. 1997 , doi =

    work page 1997

  12. [12]

    and Axtell, Robert , title =

    Epstein, Joshua M. and Axtell, Robert , title =

    work page

  13. [13]

    Proceedings of the National Academy of Sciences , volume =

    Bonabeau, Eric , title =. Proceedings of the National Academy of Sciences , volume =. 2002 , doi =

    work page 2002

  14. [14]

    , title =

    DeGroot, Morris H. , title =. Journal of the American Statistical Association , volume =. 1974 , doi =

    work page 1974

  15. [15]

    , title =

    Clifford, Peter and Sudbury, A. , title =. Biometrika , volume =. 1973 , doi =

    work page 1973

  16. [16]

    Social Simulacra: Creating Populated Prototypes for Social Computing Systems

    Park, Joon Sung and Popowski, Lindsay and Cai, Carrie J. and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S. , title =. 2022 , eprint =. doi:10.48550/arXiv.2208.04024 , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2208.04024 2022

  17. [17]

    Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia

    Vezhnevets, Alexander Sasha and Agapiou, John P. and Aharon, Avia and Ziv, Ron and Matyas, Jayd and Du. Generative Agent-Based Modeling with Actions Grounded in Physical, Social, or Digital Space Using Concordia , year =. 2312.03664 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Humanoid Agents: Platform for Simulating Human-like Generative Agents

    Wang, Zhilin and Chiu, Yu Ying and Chiu, Yu Cheung. Humanoid Agents: Platform for Simulating Human-like Generative Agents. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 2023. doi:10.18653/v1/2023.emnlp-demo.15

    work page doi:10.18653/v1/2023.emnlp-demo.15 2023

  19. [19]

    2023 , eprint =

    Gao, Chen and Lan, Xiaochong and Lu, Zhihong and Mao, Jinzhu and Piao, Jinghua and Wang, Huandong and Jin, Depeng and Li, Yong , title =. 2023 , eprint =

    work page 2023

  20. [20]

    2024 , eprint =

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

    work page 2024

  21. [21]

    2025 , eprint =

    Piao, Jinghua and Yan, Yuwei and Zhang, Jun and Li, Nian and Yan, Junbo and Lan, Xiaochong and Lu, Zhihong and Zheng, Zhiheng and Wang, Jing Yi and Zhou, Di and Gao, Chen and Xu, Fengli and Zhang, Fang and Rong, Ke and Su, Jun and Li, Yong , title =. 2025 , eprint =

    work page 2025

  22. [22]

    2026 , eprint =

    Huang, Renhong and Tang, Ning and Xu, Jiarong and Cao, Yuxuan and Tu, Qingqian and Guo, Sheng and Zheng, Bo and Liu, Huiyuan and Yang, Yang , title =. 2026 , eprint =

    work page 2026

  23. [23]

    2026 , booktitle=

    Position: LLM-Based Social Simulations Require a Boundary , author=. 2026 , booktitle=

    work page 2026

  24. [24]

    2026 , eprint =

    Zhang, Yongmao and Qiao, Kai and Wang, Zhengyan and Liang, Ningning and Ma, Dekui and Sun, Wenyao and Chen, Jian and Yan, Bin , title =. 2026 , eprint =

    work page 2026

  25. [25]

    2026 , eprint =

    Reji, Deepak John , title =. 2026 , eprint =

    work page 2026

  26. [26]

    LLM Agents Predict Social Media Reactions but Do Not Outperform Text Classifiers: Benchmarking Simulation Accuracy Using 120K+ Personas of 1511 Humans

    LLM Agents Predict Social Media Reactions but Do Not Outperform Text Classifiers: Benchmarking Simulation Accuracy Using 120K+ Personas of 1511 Humans , author=. arXiv preprint arXiv:2604.19787 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    work page 1972

  28. [28]

    Publications Manual , year = "1983", publisher =

    work page 1983

  29. [29]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  30. [30]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

    work page

  31. [31]

    Dan Gusfield , title =. 1997

    work page 1997

  32. [32]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    work page 2015

  33. [33]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

    work page