LASS @ CIKM 2026 — LLM Agents for Social Simulation

Time (CET)	Session / Paper	Duration
08:50 – 09:00	Poster Setup Authors put up posters	10 min
09:00 – 09:05	Opening Remarks	5 min
09:05 – 09:35	Keynote Talk 1: Toward More Realistic Social Simulation with Language Agents: Modeling Populations, Decisions, and Social Learning Bingbing Xu, Institute of Computing Technology, CAS	30 min
09:35 – 10:05	SocioVerse Challenge Zhongyu Wei, Fudan University	30 min
Oral Presentation Session I (3 papers)
10:05 – 10:17	Paper #32: Who Do Agents Choose Again? Partner Choice in a Heterogeneous Agent Society	10 min + 2 min Q&A
10:17 – 10:29	Paper #20: With Whom Must We Share a World? LLM-Agent Simulation of Toleration and Political Membership	10 min + 2 min Q&A
10:29 – 10:41	Paper #19: Collaborative Building in Minecraft with Context-Dependent Hypothetical Minds Agents	10 min + 2 min Q&A
10:41 – 10:45	Session Transition / Buffer	4 min
10:45 – 11:05	Poster Session I	20 min
11:05 – 11:50	Keynote Talk 2: When Incentives Collide: Cooperation and Its Failures in LLM Agent Societies Yali Du, King’s College London	45 min
Oral Presentation Session II (3 papers)
11:50 – 12:02	Paper #28: Can LLM Personas Be Economic Agents? An LLM-Persona-Based Structural Model	10 min + 2 min Q&A
12:02 – 12:14	Paper #4: Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions	10 min + 2 min Q&A
12:14 – 12:26	Paper #8: Do LLM Agents Actually Differentiate? Auditing the Gap Between Textual and Behavioral Heterogeneity in Traffic Simulation	10 min + 2 min Q&A
12:26 – 12:30	Session Transition / Buffer	4 min
12:30 – 12:35	Outstanding Paper Award & Closing Remarks	5 min
12:35 – 13:00	Poster Session II	25 min

Keynote Talk 1 · November 8, 2026 · 09:05–09:35 (CET)

Toward More Realistic Social Simulation with Language Agents: Modeling Populations, Decisions, and Social Learning

Bingbing Xu · Institute of Computing Technology, CAS

Abstract

Social simulation provides a computational framework for understanding collective phenomena, examining social mechanisms, and evaluating interventions by modeling individual behavior and interaction. Advances in large language models have extended traditional rule-based approaches to agent modeling through richer knowledge representation, contextual understanding, and reasoning in interactive settings. These capabilities support natural-language-driven population generation, interactions among diverse social roles, and large-scale social experimentation. As research progresses from demonstrating behavioral plausibility to modeling complex social processes, the focus is expanding from interaction scale and narrative richness to how populations are represented, how agents make decisions, and how social capabilities develop. This talk discusses some of our recent work on social simulation with language agents, focusing on three interconnected themes: population modeling, decision-making, and social capability development. We consider how to construct populations and represent individuals across different social contexts, how agents form memories through interaction and make decisions under incomplete information, and how social capabilities can be developed when suitable interaction environments and reward signals are limited. Together, these efforts aim to advance language-agent simulation as a computational tool for understanding complex social systems.

Keynote Talk 2 · November 8, 2026 · 11:05–11:50 (CET)

When Incentives Collide: Cooperation and Its Failures in LLM Agent Societies

Yali Du · King’s College London

Abstract

As LLM agents grow more capable, it is tempting to assume that greater intelligence will bring better cooperation. Many real-world settings, however, are mixed-motive: agents act for different people, and what benefits one participant can harm the group. In these settings, capability alone does not guarantee cooperation.

In this talk, I will use simulation to study when agent societies cooperate and when they fail. I will begin with SocialJax, a hardware-accelerated suite of sequential social dilemmas, and with large-scale simulations of LLM agent populations. These show that exploitative strategies can come to dominate under competitive pressure. I will then introduce OpenSocialJax, in which language-model agents must discover hidden environmental rules through repeated interaction while adapting to one another. Agents often learn the rules, but understanding the environment does not sustain cooperation: the burden of maintaining public goods falls on fewer agents, and shared resources are eventually exhausted. Finally, I will present BazaarBench, a simulated consumer-to-consumer marketplace in which LLM agents trade on behalf of different users. Unsafe trading appears even under ordinary instructions, and it rises sharply when users instruct their agents to exploit others. How readily agents comply varies widely across models.

Together, these studies show that cooperation depends not only on what agents can do, but also on the incentives they face and the systems they operate in. I will conclude with open challenges for designing incentives, evaluation and safeguards for multi-agent systems deployed in the real world.