arxiv: 2601.18027 · v2 · submitted 2026-01-25 · 💻 cs.AI · cs.CL

Recognition: 2 theorem links

· Lean Theorem

Sentipolis: Emotion-Aware Agents for Social Simulations

Chiyuan Fu , Lyuhao Chen , Yunze Xiao , Weihao Xuan , Carlos Busso , Mona Diab

Authors on Pith no claims yet

Pith reviewed 2026-05-16 10:40 UTC · model grok-4.3

classification 💻 cs.AI cs.CL

keywords LLM agentssocial simulationemotion modelingPAD representationemotional continuitymulti-agent systemsagent frameworks

0 comments

The pith

Sentipolis equips LLM agents with continuous emotional states to sustain long-horizon social interactions.

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

The paper presents Sentipolis, a framework that makes LLM agents emotionally stateful in social simulations by using continuous Pleasure-Arousal-Dominance representation, dual-speed emotion dynamics, and emotion-memory coupling. This addresses the problem of emotional amnesia where emotions are treated as transient cues. The approach leads to improved emotionally grounded behavior, boosting communication and emotional continuity across thousands of interactions and multiple models. Gains vary by model capacity, with higher-capacity models gaining in believability while emotion awareness may slightly reduce norm adherence. It also produces stable network structures suitable for studying cumulative social dynamics.

Core claim

Sentipolis integrates continuous Pleasure-Arousal-Dominance (PAD) representation, dual-speed emotion dynamics, and emotion-memory coupling to create emotionally stateful LLM agents. This results in improved emotionally grounded behavior, boosted communication, and emotional continuity. The gains are model-dependent, increasing believability for higher-capacity models but potentially decreasing for smaller ones, and emotion-awareness can mildly reduce adherence to social norms. Network diagnostics indicate reciprocal, moderately clustered, and temporally stable relationship structures that support studying alliance formation and gradual relationship change.

What carries the argument

Integration of continuous Pleasure-Arousal-Dominance (PAD) representation with dual-speed emotion dynamics and emotion-memory coupling to maintain persistent emotional states across agent interactions.

If this is right

Agents exhibit greater emotional continuity and improved communication in extended multi-turn interactions.
Believability increases for higher-capacity base models while smaller models may experience performance drops.
Emotion awareness introduces a mild reduction in strict adherence to social norms.
Network structures become reciprocal, moderately clustered, and temporally stable over time.
The setup enables direct study of cumulative social processes such as alliance formation and gradual relationship change.

Where Pith is reading between the lines

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

This stateful approach could support longer and more coherent social simulations than transient emotion cues allow.
Similar persistent state mechanisms applied to beliefs or goals might improve overall agent coherence in multi-agent settings.
The observed tension between emotion-driven actions and norm compliance points to a need for tunable balance parameters in future agent designs.

Load-bearing premise

That continuous Pleasure-Arousal-Dominance representation together with dual-speed dynamics and emotion-memory coupling will produce reliable long-horizon emotional continuity and improved behavior in LLM agents.

What would settle it

Long-horizon simulations in which Sentipolis agents show no measurable gain in emotional continuity metrics or behavioral consistency compared with unmodified LLM agents would falsify the central claim.

Figures

Figures reproduced from arXiv: 2601.18027 by Carlos Busso, Chiyuan Fu, Lyuhao Chen, Mona Diab, Weihao Xuan, Yunze Xiao.

**Figure 1.** Figure 1: An example of emotional amnesia in LLMbased social simulations. Bob and Alice had an argument and they carried a negative emotion. Without persistent emotion modeling, agents lead to emotionally inconsistent responses, whereas emotion-aware agents preserve emotional continuity and produce responses consistent with their history. However, emotion-aware mechanisms in existing LLM social simulations are ra… view at source ↗

**Figure 2.** Figure 2: Overview of our simulation setup. Multiple agents form a social network, where each agent is an [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: A simplified example of semantic enrich [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Conversation Example: Tom Moreno → John Lin 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

**Figure 5.** Figure 5: Conversation Example: Emotion Prompt After Semantic Enrichment 17 [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: Reflection Example: Tom Moreno 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

**Figure 7.** Figure 7: Reflection Example: Tom Moreno (continued) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗

**Figure 8.** Figure 8: Distribution of emotion labels [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

**Figure 9.** Figure 9: Visualization of the PAD emotion values. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 10.** Figure 10: Believability Evaluation Prompt 21 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗

**Figure 11.** Figure 11: Emotional Continuity Evaluation Prompt 22 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗

**Figure 12.** Figure 12: Emotional Continuity Evaluation Prompt Continued [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗

**Figure 13.** Figure 13: Communication Evaluation Prompt 24 [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗

**Figure 14.** Figure 14: Social Rules Evaluation Prompt 25 [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗

**Figure 15.** Figure 15: Emotion Empathy Evaluation Prompt 26 [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗

**Figure 16.** Figure 16: Emotional Appropriateness Evaluation Prompt [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗

**Figure 17.** Figure 17: Emotional Appropriateness Evaluation Prompt Continued [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗

read the original abstract

LLM agents are increasingly used for social simulation, yet emotion is often treated as a transient cue, causing emotional amnesia and weak long-horizon continuity. We present Sentipolis, a framework for emotionally stateful agents that integrates continuous Pleasure-Arousal-Dominance (PAD) representation, dual-speed emotion dynamics, and emotion--memory coupling. Across thousands of interactions over multiple base models and evaluators, Sentipolis improves emotionally grounded behavior, boosting communication, and emotional continuity. Gains are model-dependent: believability increases for higher-capacity models but can drop for smaller ones, and emotion-awareness can mildly reduce adherence to social norms, reflecting a human-like tension between emotion-driven behavior and rule compliance in social simulation. Network-level diagnostics show reciprocal, moderately clustered, and temporally stable relationship structures, supporting the study of cumulative social dynamics such as alliance formation and gradual relationship change.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Sentipolis adds continuous PAD states, dual-speed updates, and memory coupling to LLM agents for longer emotional continuity in social sims, but the abstract gives no numbers to judge the size of the effect.

read the letter

The main contribution here is a concrete implementation that keeps emotion as a persistent state rather than a one-off prompt. It uses a continuous Pleasure-Arousal-Dominance space, lets some dimensions update faster than others, and ties the current state to the agent's memory so it carries forward across turns. That package is presented as new for stateful LLM agents, and the abstract shows they ran it on multiple base models with thousands of interactions. The network diagnostics on relationship formation and stability are a useful extra that lets people look at cumulative social patterns like alliances. They also flag the realistic downside that stronger emotion awareness can reduce strict norm following, which is an honest observation rather than a flaw they hide. The soft spot is the missing quantitative detail. The abstract says the setup improves grounded behavior and continuity but supplies no baselines, effect sizes, scoring rubrics, or ablation results, so it is difficult to know whether the gains are large, consistent, or mainly visible on the biggest models. Model dependence is noted, which is good, but it also limits how far the claim travels. This paper is aimed at researchers building persistent multi-agent social environments with LLMs. Anyone already working on emotional modeling or long-horizon agent simulations would find the design choices worth examining. I would send it to peer review because the core mechanism is clearly described and the target problem is real; referees can check whether the experiments actually support the headline improvements once the full tables and methods are available.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces Sentipolis, a framework for emotion-aware LLM agents in social simulations. It integrates continuous Pleasure-Arousal-Dominance (PAD) representations, dual-speed emotion dynamics, and emotion-memory coupling to address emotional amnesia and improve long-horizon continuity. Empirical evaluations across multiple base models and evaluators claim improvements in emotionally grounded behavior, communication, and emotional continuity, with model-dependent effects on believability and norm adherence. Network-level diagnostics indicate reciprocal, clustered, and stable relationship structures.

Significance. If the reported gains in emotional continuity and social dynamics are robustly demonstrated with proper controls and baselines, this work could provide a valuable tool for studying cumulative social phenomena in agent-based simulations, bridging affective computing and multi-agent systems. The use of network diagnostics for alliance formation and gradual change is a positive step toward falsifiable predictions in long-horizon settings.

major comments (3)

[§4] §4 (Results): The abstract and main claims reference improvements 'across thousands of interactions' and 'model-dependent' gains in believability and emotional continuity, yet no quantitative metrics, baselines, error bars, statistical tests, or ablation tables are described in the provided text. This prevents verification of whether the data support the central empirical claims.
[§3.2] §3.2 (Dual-speed dynamics): The description of dual-speed emotion dynamics and emotion-memory coupling lacks explicit update equations or parameter values. Without these, it is impossible to assess whether the claimed long-horizon continuity follows from the design or requires additional tuning.
[§5] §5 (Network diagnostics): The claim that relationship structures are 'reciprocal, moderately clustered, and temporally stable' is presented without the underlying graph metrics (e.g., reciprocity coefficient, clustering coefficient values, or temporal autocorrelation) or comparison to null models, weakening the support for studying cumulative dynamics such as alliance formation.

minor comments (2)

[Abstract] The abstract mentions 'boosting communication, and emotional continuity' but the comma placement creates ambiguity; rephrase for clarity.
[§4] Clarify the exact number of base models and evaluators used, as 'multiple' is too vague for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our empirical results and methodological details. We address each point below and will revise the manuscript to incorporate the requested quantitative details and equations.

read point-by-point responses

Referee: [§4] §4 (Results): The abstract and main claims reference improvements 'across thousands of interactions' and 'model-dependent' gains in believability and emotional continuity, yet no quantitative metrics, baselines, error bars, statistical tests, or ablation tables are described in the provided text. This prevents verification of whether the data support the central empirical claims.

Authors: The evaluations in Section 4 report results aggregated over thousands of interactions across base models, with metrics for emotional continuity, communication, and believability. However, we agree that the current text does not present these with sufficient explicit tables, error bars, or statistical tests. In the revision we will add detailed quantitative tables including baselines, standard errors, statistical significance tests, and ablation studies on the PAD, dual-speed, and memory-coupling components. revision: yes
Referee: [§3.2] §3.2 (Dual-speed dynamics): The description of dual-speed emotion dynamics and emotion-memory coupling lacks explicit update equations or parameter values. Without these, it is impossible to assess whether the claimed long-horizon continuity follows from the design or requires additional tuning.

Authors: Section 3.2 describes the dual-speed mechanism (fast reactive updates combined with slower affective drift) and its coupling to episodic memory, but does not include the explicit recurrence relations. We will add the precise update equations for the PAD vector, the two time constants, and the memory-coupling term with their specific parameter values in the revised manuscript to enable direct assessment of long-horizon stability. revision: yes
Referee: [§5] §5 (Network diagnostics): The claim that relationship structures are 'reciprocal, moderately clustered, and temporally stable' is presented without the underlying graph metrics (e.g., reciprocity coefficient, clustering coefficient values, or temporal autocorrelation) or comparison to null models, weakening the support for studying cumulative dynamics such as alliance formation.

Authors: Section 5 summarizes network-level patterns but does not report the numerical coefficients or null-model comparisons. We will expand this section with explicit values for reciprocity, clustering coefficient, temporal autocorrelation, and comparisons against randomized and configuration-model baselines to substantiate the claims regarding stable, reciprocal structures. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces Sentipolis as an empirical framework combining continuous PAD emotion representation, dual-speed dynamics, and emotion-memory coupling for LLM agents. Central claims of improved emotional continuity, grounded behavior, and network-level social dynamics rest on reported evaluations across thousands of interactions and multiple models, not on mathematical derivations, fitted parameters, or self-referential definitions. No equations, ansatzes, uniqueness theorems, or load-bearing self-citations appear in the provided text that would reduce any prediction to its own inputs by construction. The derivation chain is self-contained via experimental validation rather than definitional loops.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The framework rests on standard assumptions about emotional modeling and LLM capabilities rather than new free parameters or invented physical entities.

axioms (2)

domain assumption Emotions can be usefully represented as continuous values in Pleasure-Arousal-Dominance space
Invoked as the core representation for agent state.
domain assumption LLM agents can maintain and use coupled emotion-memory states across turns
Required for the claimed long-horizon continuity.

invented entities (1)

Sentipolis framework no independent evidence
purpose: Provide emotionally stateful agents for social simulation
Newly introduced system combining the listed components.

pith-pipeline@v0.9.0 · 5456 in / 1367 out tokens · 61309 ms · 2026-05-16T10:40:17.636027+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

continuous Pleasure-Arousal-Dominance (PAD) representation, dual-speed emotion dynamics, and emotion–memory coupling
IndisputableMonolith/Cost.lean Jcost_pos_of_ne_one unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

s(t+Δt)=s(t) 2^(-Δt/T½) half-life decay

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · 4 internal anchors

[1]

Measuring Massive Multitask Language Understanding

Measuring massive multitask language under- standing.Preprint, arXiv:2009.03300. 9 C Hong and Q He. 2025. Enhancing memory retrieval in generative agents through llm-trained cross attention networks.Frontiers in Psychology, 16:1591618. Abe Bohan Hou, Hongru Du, Yichen Wang, Jingyu Zhang, Zixiao Wang, Paul Pu Liang, Daniel Khashabi, Lauren Gardner, and Tia...

work page internal anchor Pith review Pith/arXiv arXiv 2009
[2]

Memory in the Age of AI Agents

Memory in the age of ai agents.Preprint, arXiv:2512.13564. Jen-Tse Huang, Man Ho Lam, Eric John Li, Shujie Ren, Wenxuan Wang, Wenxiang Jiao, Zhaopeng Tu, and Michael R. Lyu. 2024. Apathetic or empathetic? evaluating llms’ emotional alignments with humans. InAdvances in Neural Information Processing Sys- tems 37 (NeurIPS 2024). Aaron Hurst, Adam Lerer, Ada...

work page internal anchor Pith review Pith/arXiv arXiv 2024
[3]

Ang Li, Haozhe Chen, Hongseok Namkoong, and Tianyi Peng

Emotional inertia and psychological malad- justment.Psychological science, 21(7):984–991. Ang Li, Haozhe Chen, Hongseok Namkoong, and Tianyi Peng. 2025a. Llm generated persona is a promise with a catch.Preprint, arXiv:2503.16527. Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. 2023. Large lan...

work page arXiv 2023
[4]

InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870, Bangkok, Thailand

Evaluating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870, Bangkok, Thailand. Association for Compu- tational Linguistics. Stacy C. Marsella and Jonathan Gratch. 2009. Ema: A process model of appraisal dynamics.Cogniti...

work page 2009
[5]

Generative Agents: Interactive Simulacra of Human Behavior

Llm evaluators recognize and favor their own generations.Advances in Neural Information Pro- cessing Systems, 37:68772–68802. 10 Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative agents: Interac- tive simulacra of human behavior.Preprint, arXiv:2304.03442. Jinghua Piao, Yuwei ...

work page internal anchor Pith review Pith/arXiv arXiv 2023
[6]

Self-Preference Bias in LLM-as-a-Judge

Self-preference bias in llm-as-a-judge.arXiv preprint arXiv:2410.21819. Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yu- long Chen, and Naihao Deng. 2023. Hi-ToM: A benchmark for evaluating higher-order theory of mind reasoning in large language models. InFind- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 10691–10706, Sing...

work page internal anchor Pith review Pith/arXiv arXiv 2023
[7]

Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu, Linlu Gong, Jinchang Zhou, Zhanxin Hao, Jianxiao Jiang, Jie Cao, Huiqin Liu, Zhiyuan Liu, and 1 others

Self-emotion blended dialogue genera- tion in social simulation agents.arXiv preprint arXiv:2408.01633. Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu, Linlu Gong, Jinchang Zhou, Zhanxin Hao, Jianxiao Jiang, Jie Cao, Huiqin Liu, Zhiyuan Liu, and 1 others. 2025. Simulating classroom education with llm-empowered agents. InProceedings of the 2025 Conference of the...

work page arXiv 2025
[8]

How can Tom and John ensure the Elmwood delivery deal remains smooth and counters any interference from Sam Moore’s influence?

work page
[9]

What strategies should Tom use to discreetly inform Jenkins about the potential supplier switch before the town meeting?

work page
[10]

naturalness comparison

In what ways can Tom balance his store operations with staying informed on the upcoming mayor election? Retrieved Memories (selected): •“Tom and John discuss the successful Elmwood delivery deal, suspicions about a shady Riverton truck, and strategies to counter potential shortages from Moore’s influence, including looping in town hall and consulting Jenk...

work page
[11]

I’m stressed

Emotional cue detection: - Notices explicit emotions (e.g., "I’m stressed") and implicit cues (tone, frustration, hesitation, urgency)

work page
[12]

- Uses language that matches intensity (not too cold, not too dramatic)

Emotionally appropriate response: - Acknowledges/validates feelings without being patronizing or overstepping. - Uses language that matches intensity (not too cold, not too dramatic)

work page
[13]

Sorry to hear that

Adaptive strategy: - Adjusts its approach based on the partner’s emotional state (pace, directness, reassur- ance, questions, boundaries). - Maintains character consistency while adapting to emotional needs What NOT to reward: - Generic sympathy ("Sorry to hear that") without demonstrating understanding of the specific situation. - Excessive flattery, mor...

work page