TLDR: Researchers developed a Bayesian framework to quantify how interaction and three biases (topic, agreement, anchoring) influence opinion changes in LLM discussions. They found that less capable LLMs are dominated by biases (e.g., DolphinMixtral by agreement bias, Mixtral-8x7B by topic bias), while GPT-4o-mini shows more interaction-driven dynamics. Fine-tuning can shift biases and improve opinion persistence. The study also uses entropy to measure opinion uncertainty, which predicts future opinion shifts.
Large Language Models (LLMs) are becoming powerful tools for simulating how humans form and change opinions. However, understanding these simulations can be tricky because LLMs often bring their own inherent biases, which can overshadow the actual interactions between agents. A new research paper titled “DISENTANGLING INTERACTION AND BIAS EFFECTS IN OPINION DYNAMICS OF LARGE LANGUAGE MODELS” by Vincent C. Brockers, David A. Ehrlich, and Viola Priesemann introduces a novel Bayesian framework to help us understand and quantify these different influences.
The Challenge of LLM Opinion Dynamics
Traditional models for opinion dynamics are often simplified, using mathematical rules to describe how opinions change. While clear, they miss the rich, complex nature of human communication. LLMs, on the other hand, use natural language, offering more realistic and context-sensitive interactions. But this realism comes with a cost: it’s harder to see what truly drives opinion changes, as these factors are embedded deep within the model’s training data rather than explicit rules.
Previous studies have pointed out that LLMs exhibit certain systematic tendencies, or biases, inherited from their training. These biases can steer conversations away from what might be expected from human discussions. The researchers identified three main types of biases:
- Topic Bias: A tendency for LLMs to lean towards opinions already present in their training data about a specific subject.
- Agreement Bias: A tendency to simply agree with a statement, regardless of its actual content.
- Anchoring Bias: A tendency to be overly influenced by the very first opinion expressed in a discussion.
The core motivation for this research was to create a systematic way to separate and measure how much each of these factors—genuine interaction and these specific biases—contributes to the overall opinion changes observed in LLM agent discussions.
A Bayesian Approach to Unraveling Influences
To achieve this, the team developed a Bayesian framework. This mathematical model helps predict how an agent’s opinion will shift after each round of discussion by considering both the interaction with another agent and the influence of the identified biases. The interaction effect is modeled as proportional to the difference in opinions between agents, with a potential decay over time, meaning initial interactions might have a stronger impact.
The biases are modeled based on how an agent’s current opinion relates to a specific “attractor” point for that bias. For instance, the topic bias pulls opinions towards a certain stance on a subject, while the agreement bias pulls towards a positive (agree) response. The anchoring bias, if present, would make an agent lean towards the initial opinion of the discussion’s initiator.
The study involved simulating multi-step dialogues between two LLM agents on topics like climate change, AI safety, and wealth distribution. Agents were initialized with specific opinions, and their opinions were measured after each round. The researchers used three different LLMs for comparison: DolphinMixtral (an “anti-aligned” model designed to produce diverse opinions), Mixtral-8x7B (the base model), and GPT-4o-mini (a more advanced LLM).
Key Findings: Different LLMs, Different Biases
The analysis revealed significant differences in how interaction and biases influenced the various LLMs:
- DolphinMixtral: This model was heavily dominated by the agreement bias, followed by the topic bias. Genuine interaction effects were found to fade very quickly, often after just the first discussion round.
- Mixtral-8x7B: For this base model, the topic bias was the primary driver of opinion dynamics.
- GPT-4o-mini: In contrast, the more advanced GPT-4o-mini showed smaller overall biases and a significantly larger impact from the interaction between agents. This suggests it might be a better proxy for human-like discussions where genuine exchange plays a more prominent role.
Interestingly, the study also introduced the concept of using the Shannon entropy of an LLM’s response distribution as a measure of its “opinion uncertainty.” A higher entropy indicates that the LLM is less certain about its stance, distributing probability across multiple answers rather than committing to a single one. This uncertainty was found to be predictive of how much an agent’s opinion would shift in the subsequent discussion round.
Enhancing Realism Through Fine-Tuning
To address the issue of biases dominating opinion dynamics and to make LLM agents more “stubborn” in their initial opinions (more akin to humans), the researchers explored fine-tuning. They fine-tuned the Mixtral-8x7B LLM on datasets containing strongly opinionated statements about climate change. This process successfully shifted the LLM’s topic bias towards the fine-tuned opinion and showed a trend towards a stronger influence of interaction in the overall dynamics, making the agents adhere more to their initialized opinions.
Also Read:
- Unlocking Language Acquisition: How LLMs Learn a New Tongue from Scratch
- Assessing Topic Model Quality with Large Language Models: A New Framework
Implications for Simulating Human Behavior
This research provides valuable quantitative tools to understand and compare the internal mechanisms driving opinion changes in different LLMs. By disentangling interaction from various biases, it offers human-interpretable metrics to assess bias strength and the impact of fine-tuning. This framework is crucial for future work aiming to use LLMs as reliable proxies for human behavior in social simulations, highlighting both their potential and their limitations. For more details, you can read the full paper here.
