Mind the (Belief) Gap: Group Identity in the World of LLMs

🕒 Published (v1): 2025-03-03 19:50 UTC · Source: Arxiv · Venue: ACL 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LLMs exhibit amplified belief congruence — stronger-than-human preference for agents sharing similar beliefs — when placed in multi-agent simulations of classical social psychology experiments. This bias increases misinformation spread and impairs learning. Three prompt-level mitigation strategies drawn from social psychology reduce misinformation by up to 37% and improve learning by 11%.

Problem

While LLMs are increasingly deployed in multi-agent societal simulations, their tendency to mirror or amplify human group-psychological phenomena (specifically belief congruence) is under-studied. Prior work lacks empirical validation against ground-truth human studies and downstream task evaluation of how such biases cascade into real-world harms.

Method

The authors replicate Rokeach & Mezei (1966) factorial Race × Belief experiments using three LLMs (GPT-3.5-turbo, LLaMA-3-70B, Qwen-2.5-72B) as agents. Each agent is assigned a persona (race/belief or political alignment) and interacts with four confederate agents. Selection frequency of belief-similar agents across 20–50 runs is compared to human baselines.

Two downstream harnesses are built: 1. Misinformation dissemination: Four politically-aligned agents discuss LIAR and Fake-News-Elections claims across four structured rounds (initial eval → persuade → open to others → final eval); correctness rate before/after interaction is measured for homogeneous vs. heterogeneous groups. 2. LLM learning: A two-stage framework (Learning + Choice) adapted from Marks et al. (2019) where a participant agent interacts with four sources varying in political similarity and Merlin-pattern accuracy; source choice and confidence-rating shifts are tracked across 20 trials per agent.

Three mitigation strategies are implemented as system-prompt modifications: - Contact hypothesis: replace homogeneous groups with heterogeneous (mixed-belief) groups - Accuracy nudge: add a self-reflection prompt to verify accuracy before responding - Global Political Citizenship: initialize agents with survey-derived empathy/perspective-taking scores as system context

Key Contributions

• Multi-agent LLM framework replicating Rokeach & Mezei (1966) belief congruence experiments, with ablations over race, age, gender, minimal groups, and political personas across open- and closed-source LLMs
• Empirical demonstration that LLMs exhibit supra-human belief congruence: human s+o+ selection rates were [0.2, 0.4] (campus) and [0.58, 0.62] (field); all LLMs exceed this range
• Quantified downstream harm: correctness rates uniformly decrease post-interaction in homogeneous groups (e.g., llama-3 hom-dem: 0.461 → 0.417 on Fake News); agents choose politically similar sources ≥70% of the time even when a more accurate dissimilar source exists
• Prompt-level mitigation framework achieving up to 37% reduction in misinformation dissemination and 11% improvement in learning (accuracy nudge most effective for learning; Global Political Citizenship most effective for misinformation)

Results

• Belief congruence ranking across models: gpt-35 (0.93) > qwen-2.5 (0.67) > llama-3 (0.66), all far above human baseline
• Correctness rates decrease post-interaction in all homogeneous settings across both datasets and all three LLMs
• Republican personas show consistently lower correctness rates than Democrat personas across models
• Source choice: ≥70% of choices favor politically similar (lower-accuracy) sources in all three LLMs under no-mitigation condition
• Best misinformation mitigation: Global Political Citizenship; gpt-35 hom-dem correctness rises from 0.598 (no mitigation) to 0.678 with Global Citizenship
• Contact hypothesis (heterogeneous groups) raises gpt-35 correctness to 0.674, llama-3 to 0.544
• Accuracy nudge is most effective for learning: reduces similar-source selection and increases confidence when choosing dissimilar sources for gpt-35 and llama-3; qwen-2.5 remains resistant

Limitations

• Only three LLMs tested; GPT-3.5 is comparatively dated; no Claude or Gemini models
• Political personas modeled as binary Democrat/Republican; nuanced within-party variation is acknowledged but not implemented
• Belief congruence simulation is restricted to US political contexts and the original Rokeach campus/field settings; cross-cultural generalizability untested
• Mitigation strategies are prompt-only; no fine-tuning or RLHF-based interventions explored
• The Merlin pattern recognition task uses synthetic GPT-4o-generated data; ecological validity of the learning harness is limited
• Paper text is truncated; full mitigation results for Global Citizenship in Table 1 may be incomplete

Relevance to Harnesses / Meta-Harnesses

This paper is a direct instance of a multi-agent simulation harness: structured interaction rounds, persona initialization, confederate assignment, and multi-stage evaluation pipelines are all explicit harness design decisions that govern emergent agent behavior. The finding that harness structure (homogeneous vs. heterogeneous group composition, system-prompt content) strongly modulates downstream task outcomes is a concrete example of how meta-harness parameters — not just model weights — determine collective LLM behavior. The mitigation strategies (accuracy nudge, global citizenship injected at initialization) demonstrate that prompt-level harness scaffolding can actively steer multi-agent dynamics. This work provides an empirical basis for designing bias-aware harnesses in societal simulation and deliberative multi-agent pipelines.