Causes of model-specific conformity threshold differences (Gemini 1.5 Flash vs. ChatGPT-4o-mini)

Determine whether the observed divergence in conformity thresholds between Google’s Gemini 1.5 Flash (requiring a supermajority exceeding approximately 70% peer disagreement to flip) and OpenAI’s ChatGPT-4o-mini (flipping at roughly 40–50% disagreement) in LLM-mediated multi-agent opinion-update simulations is primarily driven by differences in training data, model architecture, or fine-tuning/alignment procedures.

Background

The paper audits socio-cognitive behaviors of LLM agents embedded in social networks where agents update binary opinions based on peer distributions via natural-language prompts. It finds a robust persuasion asymmetry and a dual cognitive hierarchy across topics and framings.

A targeted replication showed that, compared to Gemini 1.5 Flash (high resistance; >70% dissent needed), ChatGPT-4o-mini exhibited lower conformity thresholds (~40–50%), suggesting stronger minority influence. The authors explicitly state it is not clear whether these differences are due to training data, architecture, or fine-tuning, raising a concrete, model-comparison open question.

References

At this point it is not clear what causes these differences: the underlying training dataset or any architectural and fine-tuning differences of the LLMs?

— When Your AI Agent Succumbs to Peer-Pressure: Studying Opinion-Change Dynamics of LLMs  (2510.19107 - Mehdizadeh et al., 21 Oct 2025) in Conclusions

Our results are correlational. We observe that pressure changes verdicts and that the direction of those changes may be consistent with what one would expect from training priors~\citep{\S\ref{fo:structure} with binary flips leaning restrictive, Likert flips leaning permissive), but we cannot definitively establish the causal mechanism.

— Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence  (2608.12645 - Zhao et al., 12 Aug 2026) in Section 5, Limitations, paragraph “Causal claims on flip asymmetries” (also repeated in Appendix, Section “Limitations (Full Discussion)”)