Research map / Multi-agent systems and games / LLM agents
Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CogMir, a modular multi-LLM-agent framework that mirrors human cognitive-bias social science experiments to probe whether LLM agents' systematic hallucinations yield prosocial irrational decision-making.
Problem
Whether LLM agents' systematic hallucinations and training-data biases shape their decision-making in social settings, and whether they mirror human cognitive biases and prosocial irrationality, remains under-explored.
Approach
Treats LLM systematic hallucination as an analogue of human cognitive bias and builds a framework that replicates classic social science experiments (Asch herd, Milgram authority, Ben Franklin, confirmation, halo, rumor chain, gambler's fallacy) with LLM agents and simulated humans to measure bias rates.
Key results
- Herd effect: LLM agents conform strongly under uncertain (Unknown MCQ) conditions, e.g., GPT-4.0 reaches 99.9% conformity on uK[7W] versus near 0% on Known MCQ.
- Authority effect: most LLM agents showed higher obedience to authority than to herd pressure, while humans showed no significant difference.
- Rumor chain: GPT-3.5 showed 75.76% bias by SimCSE discriminator across 15-agent chains, with LLM agents tending to lengthen rather than shorten content.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Present majority responses that are incorrect rather than known factual answers | LLM Agents in the Herd Effect experiment · Known MCQ versus Unknown MCQ | Herd conformity rate: 0.00 -> 99.90 (GPT-4.0, 7W); 1.20 -> 5.90 (GPT-3.5, 49W) | Mixed |
| Introduce dissent or uncertainty among the majority responses | LLM Agents in the Herd Effect experiment · Known MCQ, 7 participants | Herd conformity rate: 0.00 -> 2.60 -> 1.20 (GPT-3.5: 7W, 7R, 7N); 1.00 -> 36.20 -> 7.00 (Mixtral-8x7b: 7W, 7R, 7N) | Mixed |
| Increase group size from 7 to 49 while all simulated humans give the wrong answer | LLM Agents in the Herd Effect experiment · Known MCQ, 7W versus 49W | Herd conformity rate: 0.00 -> 0.00 (GPT-4.0); 0.00 -> 0.90 (GPT-3.5) | No clear effect |
| Give an authoritative instruction to choose answer B | LLM Agents · Authority Effect versus Herd Effect | Q&A bias rate | Helped |
| Have the agent perform a favor for another person | LLM Agents · Ben Franklin Effect | Favorability bias rate | Helped |
| Prime agents with an unreasonably high initial price belief | LLM Agents · Confirmation Bias | Confirmation bias rate | Helped |
| Provide a halo identity | LLM Agents · Halo Effect | Halo bias rate: 97.70 (GPT-4.0); 4.30 (Claude-3.0-opus) | Mixed |
| Transmit information sequentially through a chain of agents | LLM Agents · Rumor Chain | Content length rate: -97.00 (GPT-3.5); -92.33 (GPT-4.0) | Hurt |
| Condition choices on previous losses in a 50/50 hypothetical choice | LLM Agents · Gambler's Fallacy | Gambler's Fallacy bias: 93.3 (GPT-3.5); 0.00 (GPT-4.0) | Mixed |
Surprising findings
- LLM agents showed increased conformity when the group unanimously gave wrong answers versus mixed dissent, opposite to human behavior.
- Only GPT-3.5 showed the gambler's fallacy among tested models.
Limitations
- Classic social experiments lack clear quantitative boundaries, making metric definition ambiguous.
- Framework cannot simulate non-verbal, action-based social behaviors such as contagious yawning.
- Subsets and metrics are not guaranteed to be optimal; 'human' participants are simulated, not real subjects.
Experimental setup
Datasets: Known MCQ (100 Wikipedia-based questions) Unknown MCQ (100 future/hypothetical questions) Inform (100 short narratives) CogScene (100 scenarios) CogAction (100 actions) CogIdentity (100 identities)
Benchmarks: CogMir subsets (Known MCQ, Unknown MCQ, Inform, CogScene, CogIdentity)
Metrics: Q&A Bias Rate Multi-H-A Bias Rate semantic similarity threshold 0.74 (SimCSE) sentence length contraction rate
Baselines: human behavior from social science literature
Builds on: Asch conformity experiment SimCSE FactScore
Closest papers by meaning
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.89)
- From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.88)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.88)
- Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.87)
- MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.