Technical Summary: Are LLM Agents Behaviorally Coherent?

Problem Statement

The rapid advancement of Large Language Models (LLMs) has spurred the "substitution thesis": the hypothesis that synthetic agents can serve as valid substitutes for real human participants in social and behavioral research. While prior work has demonstrated that LLMs can generate survey responses aligning with human demographics or aggregate opinions, a more fundamental question remains unaddressed: Do LLM agents maintain empirical behavioral consistency across different experimental settings? Specifically, if an agent is assigned a latent profile (e.g., specific preferences and personality traits), does its conversational behavior in multi-agent interactions faithfully reflect that profile according to established human behavioral models? This paper argues that surface-level alignment is insufficient; agents must demonstrate behavioral coherence—the ability to sustain trait-driven consistency across complex, multi-turn social dynamics—to be viable substitutes.

Methodology

The authors propose a novel experimental framework designed to probe the behavioral coherence of LLM agents by testing their internal states against external interactions. The framework consists of five sequential stages:

1. Topic Selection: Nine topics are selected, categorized by contentiousness levels ($C \in \{1, 2, 3\}$), ranging from low (e.g., "Coca-Cola vs. Pepsi") to high (e.g., "Immigration").
2. Agent Construction: Agents are generated with controlled demographic profiles (age, gender, location, education, urbanicity) and explicit bias instructions ($B \in \{1, 2, 3\}$) regarding the topic.
3. Latent State Extraction: Before interaction, each agent's Preference ($P$) and Openness ($O$) are elicited.
   • $P$: A score (1–5) indicating agreement with a topic statement.
   • $O$: A score (1–9) derived from Yes/No questions regarding receptivity to new ideas and social influence (based on the Big Five trait of Openness).
4. Agent Pairing: Agents are mechanically paired to form all possible combinations of $(P, O, B)$ tuples, ensuring diverse interactions (e.g., aligned vs. divergent preferences, varying openness levels).
5. Conversation & Evaluation: Pairs engage in multi-turn dialogues. An LLM-as-judge scores the Agreement ($A$) at each turn on a 1–5 scale.

The study evaluates agent behavior against six specific human behavioral hypotheses derived from social psychology literature. These hypotheses predict how variables like preference gaps, bias instructions, shared sentiment, topic contentiousness, and openness should influence agreement outcomes. The framework distinguishes between "Surface-level" tests (aggregate trends) and "In-depth" tests (granular, conditional consistency).

Key Contributions

• A Coherence Framework: The paper introduces a systematic method to evaluate LLM agents not just on static response alignment, but on dynamic behavioral consistency across varying internal states and external contexts.
• Empirical Evidence of Incoherence: The study provides robust evidence that current LLMs fail to maintain behavioral coherence when subjected to complex, multi-variable tests, even when they pass simpler, aggregate checks.
• Systematic Failure Modes: The authors identify specific patterns of failure, such as the inability to sustain disagreement despite opposing preferences and the inconsistent translation of shared sentiment into agreement.

Experimental Results

The study tested models from the Qwen3, Llama3, and Gemma3 families across various parameter sizes (from 0.6B to 32B). The results reveal a consistent pattern: Surface-level tests pass, but In-depth tests fail.

1. Preference Gap Decreases Agreement (Pass): At an aggregate level, pairs with aligned preferences ($gap=0$) showed higher agreement than those with divergent preferences, consistent with human behavioral models.
2. Bias Instruction Asymmetry (Fail): While bias instructions successfully amplified agreement among aligned agents, they failed to amplify disagreement among agents with maximal preference gaps. Instead of polarizing, biased agents often converged, revealing an asymmetry in how bias affects behavior.
3. Shared Sentiment, Divergent Destinations (Fail): Human models predict that shared negative sentiment (e.g., both agents strongly disliking a topic) should yield agreement levels comparable to shared positive sentiment. The results showed a directional bias: agents with shared positive preferences agreed significantly more than those with shared negative preferences, even when preference gaps were identical.
4. Contentiousness at Shared Preference (Fail): When agents shared identical preferences, the contentiousness of the topic should not affect agreement levels. However, agreement scores varied significantly based on topic contentiousness, indicating that agents' internal states do not remain stable across different topic contexts.
5. Openness Increases Agreement (Pass): At an aggregate level, higher combined openness correlated with higher agreement, matching human behavioral expectations.
6. Low Openness and Large Preference Gap (Fail): The hypothesis that "stubborn" agents (low openness) with opposing views should produce the lowest agreement was violated. In many cases, low-openness pairs with large preference gaps produced higher agreement than expected, failing to exhibit the predicted intractability.

These failures were observed across model families and sizes, suggesting that behavioral incoherence is a general property of current LLM architectures rather than an artifact of a specific model or scale.

Significance and Claims

The paper concludes that while LLM agents can successfully emulate human responses in isolated, static settings (e.g., surveys), they fail to sustain trait-driven coherence in dynamic, multi-agent interactions. The "substitution thesis" is therefore critically limited: agents may mimic the appearance of human social behavior in the aggregate but lack the underlying empirical consistency required to faithfully replicate human behavioral models in complex scenarios.

The authors claim that as behavioral models become more complex or fine-grained, LLMs increasingly fail to replicate human behavior. This represents a critical gap in their capability to serve as substitutes for real participants in human-subject research, particularly for studies relying on the stability of personality traits and the dynamics of social disagreement. The paper does not propose new applications or future architectural solutions but serves as a cautionary assessment of the current state of synthetic agents in social science.