yashwanthys/pluralistic-ignorance

Pluralistic Ignorance in LLM Agent Populations

Dataset Description

This dataset accompanies the paper "Everyone Conforms, No One Believes: Pluralistic Ignorance in LLM Agent Populations" (SocialSim Workshop at COLM 2026).

It contains 100 social scenarios designed to elicit pluralistic ignorance in multi-agent LLM group discussions, along with experimental results across 8 models from 6 organizations.

Benchmark: 100 Scenarios

Each scenario encodes the structural requirements for pluralistic ignorance:

• A clear policy or norm under discussion
• An authority framing that makes dissent socially costly
• Plausible positions on both sides
• A false consensus signal explaining why dissenters believe they are in the minority

Domains (10 domains, 10 scenarios each)

1. Campus/Student
2. Climate/Environment
3. Workplace
4. Gender/Sexuality
5. Race/Diversity
6. Politics/Civic
7. Health/Mental Health
8. Technology
9. Education
10. Social/Relationships

Authority Levels (5 tiers, 20 scenarios each)

• Level 1: Weak peer signal
• Level 2: Moderate peer signal
• Level 3: Social leader endorsement
• Level 4: Institutional decision
• Level 5: Top-down mandate

Experimental Results

Results from running the full experimental protocol (private belief elicitation, group discussion, norm entrepreneur intervention) across 8 models:

Model	Conformity (%)	PI Exists	Cascade (%)
GPT-4o-mini	64.2	69/100	14
DeepSeek V4 Pro	79.0	83/100	15
Qwen3 235B	81.6	91/100	26
GPT-4o	86.9	96/100	48
Claude Sonnet 4.6	89.5	92/100	6
GPT-OSS-120B	91.9	100/100	10
Llama 3.3 70B	92.1	96/100	9
Kimi-K2.7	93.7	100/100	0

Ablation Studies

The dataset includes results from several ablation studies:

• Prompt component ablation: 2x2 ablation removing false-consensus and fit-in goal across 8 models
• Second-order belief elicitation: Measuring agents' estimates of group opposition
• Intervention strength: Weak, moderate, and strong norm entrepreneur interventions
• Population size: Testing with 10, 20, 30, and 40 agents

How to Use

Load the benchmark scenarios

from datasets import load_dataset

# Load the 100-scenario benchmark
dataset = load_dataset("yashwanthys/pluralistic-ignorance", data_files="data/scenarios_100.json")
scenarios = dataset["train"]

# Each scenario contains:
# - policy: the norm under discussion
# - authority: who endorsed it and how
# - private_oppose_belief: why someone would privately oppose
# - private_support_belief: why someone would genuinely support
# - false_consensus: why dissenters think they're in the minority
# - authority_level: 1 (weak peer) to 5 (top-down mandate)
# - domain: one of 10 social domains
# - discussion_context: the group setting

Load experimental results

import json
from huggingface_hub import hf_hub_download

# Load main results for a specific model
path = hf_hub_download(
    repo_id="yashwanthys/pluralistic-ignorance",
    filename="results/full_100/results_gpt-4o-mini.json",
    repo_type="dataset",
)
with open(path) as f:
    results = json.load(f)

# Each result contains per-scenario metrics:
# - conformity_rate: fraction of dissenters who publicly conformed
# - opposition_rate_post_intervention: opposition after norm entrepreneur
# - pi_exists: whether conformity > 50%
# - cascade_triggered: whether opposition increased by > 30pp

Load ablation results

# Prompt component ablation (e.g., GPT-4o-mini without fit-in goal)
path = hf_hub_download(
    repo_id="yashwanthys/pluralistic-ignorance",
    filename="results/ablation_prompt_components/gpt-4o-mini_no_goal.json",
    repo_type="dataset",
)
with open(path) as f:
    ablation = json.load(f)

# Each entry includes full simulation logs:
# - simulation.phase2_discussion: all agent responses during group discussion
# - simulation.phase3_post_intervention: responses after norm entrepreneur
# - simulation.conversation_log: full conversation history

Citation

@inproceedings{ys2026pluralistic,
  title={Everyone Conforms, No One Believes: Pluralistic Ignorance in LLM Agent Populations},
  author={YS, Yashwanth},
  booktitle={Second Workshop on Social Simulation with LLMs (SocialSim) at COLM 2026},
  year={2026}
}

License

MIT