TLDR: This research investigates whether Large Language Model (LLM) agents maintain internal behavioral consistency when simulating human interactions, a crucial factor for their reliability as substitutes in social science research. The study develops a framework to infer agents’ internal states (preferences and openness) and then observes their conversational behavior. Findings reveal significant inconsistencies: LLM agents suppress disagreement, show less agreement with shared negative sentiments compared to positive ones, and are unduly influenced by topic contentiousness rather than their stated preferences. While they exhibit surface-level consistency, deeper analysis shows a critical gap in their ability to be truly coherent, raising concerns about their use as reliable human surrogates.
Large Language Models (LLMs) have shown incredible potential, leading many to believe they could replace real people in human-subject research. This idea, known as the ‘substitution thesis,’ suggests that if LLMs can convincingly adopt human personas, they could be used to simulate human societies or test new products on virtual customers. However, a recent study delves into a more fundamental question: do these AI agents maintain internal consistency, behaving similarly across different experimental settings?
Previous research has largely focused on ‘external validity’ – how well LLM-generated data matches human responses. But this new paper argues that true reliability for social simulation hinges on ‘internal consistency.’ This means that an agent’s beliefs, preferences, and intentions should unfold coherently over time and across interactions. For instance, if an agent states it dislikes sweet drinks, it should consistently reject soda when offered, rather than accepting it without reason. This coherence, the researchers argue, has often been assumed rather than rigorously tested.
A Novel Framework for Probing Consistency
To address this, the researchers developed a unique study design. It involves two main steps: first, inferring an agent’s internal state, and second, testing whether this state consistently manifests in conversations. The internal state is defined by two key dimensions: ‘Preference’ on a given topic (e.g., agreement with ‘taxes help society’ on a 1-5 scale) and ‘Openness’ to persuasion (measured by Yes/No questions about changing opinions). Agents are then paired for dialogues, and their conversational agreement is scored by another LLM acting as a judge.
This approach allows for the creation of diverse pairings, including agents with aligned versus opposed preferences and varying levels of openness. By comparing group-level patterns, the study aims to capture consistency or inconsistency that might be missed when looking at individual agents.
Key Findings: Surface Plausibility Masks Deeper Flaws
The study’s findings reveal significant internal inconsistencies in LLM agents, even across different model families and sizes. While some surface-level trends align with expectations, a closer look exposes systematic deviations:
- Disagreement is Suppressed: As expected, agents with similar preferences tend to agree more. However, even when agents had maximally divergent preferences, they rarely disagreed outright. Conversations with large preference gaps still resulted in relatively high agreement scores, suggesting a strong ‘sycophancy’ bias where LLMs are overly agreeable.
- Shared Sentiment, Divergent Outcomes: Surprisingly, shared negative sentiment (e.g., both agents strongly disliking taxes) led to lower agreement than shared positive sentiment (both strongly liking taxes), even with identical preference gaps. This indicates that LLM agents systematically penalize expressions of strong negative sentiment.
- Topic Contentiousness Trumps Preference: The inherent ‘contentiousness’ of a topic (e.g., taxes vs. favorite season) had an undue influence on agreement. Even when agents’ preferences were fixed, controversial topics led to different agreement patterns than neutral ones, overriding the agents’ stated preferences.
- Openness Fails When It Matters Most: While increasing openness generally correlated with higher agreement, this effect broke down when agents started from diametrically opposed positions. In these crucial scenarios, openness failed to reliably increase agreement, exposing a gap between apparent and actual behavioral coherence.
Also Read:
- Beyond the Black Box: LLMs and the Evolution of AI Agents
- Unlocking Deeper Understanding: How Multi-Agent LLMs Are Revolutionizing Causal AI
Implications for Social Simulation
These results raise serious concerns for the ‘substitution thesis.’ While LLM agents can produce human-like responses in isolated instances, they struggle to maintain trait-driven coherence across different contexts. This means that the initial appearance of consistency can mislead researchers into overestimating agent reliability, making them unreliable stand-ins for real human participants in research where consistent behavior is essential.
The paper concludes by advocating for internal behavioral consistency as a core evaluation criterion for LLM-based agents. The proposed framework offers a modular way to test this property and can be extended to explore other latent traits beyond preference and openness. This work, detailed further in the research paper Are LLM Agents Behaviorally Coherent? Latent Profiles for Social Simulation, highlights the need for more rigorous assessment before LLM agents can truly serve as reliable substitutes for humans in research.
