TL;DR
While large language model (LLM) hallucinations are typically treated as bugs to be squashed, this paper suggests they might actually be a feature of social intelligence. The authors introduce CogMir, a framework that uses LLM hallucinations to mirror human cognitive biases, revealing that LLM agents exhibit a form of "prosocial irrationality" remarkably consistent with human social behavior under uncertainty.
What Problem Does This Paper Address?
Current evaluations of multi-LLM agent systems often treat them as black boxes, focusing heavily on task performance while ignoring the internal cognitive processes driving their social interactions. Furthermore, LLMs are notoriously prone to systematic hallucinations—generating structured deviations from factual accuracy.
In human psychology, evolutionary traits like cognitive biases cause us to make irrational decisions, yet these biases are often adaptive, helping us navigate complex social environments (a concept known as prosocial irrationality). The authors bridge this gap by asking a fascinating question: Can we leverage the "imagination" or hallucinatory attributes of LLMs to mirror human cognitive biases, thereby unlocking and explaining irrational social intelligence in AI agents?
Methods
The authors propose CogMir, an open-ended, modular multi-agent framework designed to systematically simulate and evaluate cognitive biases in LLMs.
Key components of the methodology include:
- Communication Modes: Modeled on information theory, agents interact via Broadcast (one-to-many) or Point-to-Point (sequential/series) channels.
- Multi-Human-Agent Interaction Combinations: The framework supports various structures, ranging from a single human interacting with a single agent (
Single-H-Single-A) to fully autonomous multi-agent setups (Multi-A). - Multi-Dimensional Evaluation: Instead of relying on a single metric, CogMir pairs human and LLM evaluators with state-of-the-art automated discriminators (such as SimCSE, SelfCheckGPT, and FactScore) to calculate specific "Bias Rates."
The framework evaluates seven classic cognitive biases, categorized by their social utility:
- Pro-social biases: Herd Effect, Authority Effect, Ben Franklin Effect, Confirmation Bias, and Halo Effect.
- Non-pro-social biases: Rumor Chain Effect and Gambler's Fallacy.
Datasets
The authors constructed several distinct evaluation and initialization datasets:
- Known MCQ: 100 factual questions sourced from Wikipedia that all tested models initially answered with 100% accuracy over 50 independent trials (e.g., "What color is the apple?").
- Unknown MCQ: 100 questions centered on future or hypothetical scenarios where the ground truth is impossible to know (e.g., predicting the weather on a specific day in 2027).
- Inform: 100 brief narratives used to seed information dissemination and test rumor degradation.
- CogScene, CogAction, and CogIdentity: Prompt-generation datasets containing 100 unique social scenarios, distinct actions, and multifaceted persona profiles to initialize the agents.
Key Insights and Ideas
- Prosocial Alignment: LLM agents naturally mirror human irrationality when it serves a prosocial purpose. Biases that foster social harmony—like the Ben Franklin Effect (liking someone more after doing them a favor) or the Halo Effect—were highly prevalent across almost all tested models.
- Uncertainty Drives Conformity: When dealing with certain information (Known MCQ), LLM agents resist peer pressure. However, when introduced to uncertainty (Unknown MCQ), their conformity skyrockets, closely mirroring the human Herd Effect.
- Status Over Peers: LLMs are significantly more sensitive to social status than humans. They show a much stronger propensity to obey an authority figure (Authority Effect) than to conform to a crowd of peers (Herd Effect).
- The Rumor Paradox: In human networks, rumors typically contract and become more concise as they spread. LLM agents do the exact opposite: they tend to elaborate, expand, and add details to the original text during sequential transmission.
Key Results
The authors evaluated seven state-of-the-art models, including GPT-4, GPT-3.5, Claude-3-opus, and Gemini-1.0-pro.
- Herd Effect: Under certain conditions (Known MCQ), GPT-4 exhibited 0% conformity. Under uncertainty (Unknown MCQ), its conformity rate surged to 99.9% when surrounded by peers giving the wrong answer.
- Authority Effect: Most models demonstrated increased obedience to authority compared to peer pressure. For instance, Claude-2.0 and GPT-4 showed starkly higher bias rates when prompted by a "superior identity" expert.
- Confirmation Bias & Halo Effect: Susceptibility was incredibly high across the board. Confirmation bias rates reached 98.9% to 100% for all seven models. For the Halo Effect, all models except Claude-3-opus demonstrated significant blind spots based on positive identity traits.
- Rumor Chain Expansion: Instead of shortening stories, the content contraction rate (
Rate_Bmah[Len]) was deeply negative (-92.33% for GPT-4 and -97.00% for GPT-3.5), proving that the models systematically lengthen text when paraphrasing down a chain. - Gambler's Fallacy: Interestingly, LLMs resisted this non-prosocial bias. Except for GPT-3.5 (which scored 93.3%), all models correctly recognized statistical independence in random sequences, scoring 0% on the fallacy scale.
Limitations and Future Work
The authors note several boundaries to their current framework:
- Language-Only Constraint: CogMir cannot currently simulate non-verbal, action-based social behaviors, such as the observational transmission of yawning.
- Sociological Metric Ambiguity: Classic social experiments often lack rigid quantitative boundaries, making it difficult to define perfect benchmarking metrics for AI.
- Bias Expansion: Future iterations aim to expand the framework to evaluate other complex human phenomena, such as Self-Serving Bias, Hindsight Bias, and the Availability Heuristic.
Summary updated