TLDR: Researchers used “linear probes” to analyze how Large Language Models (LLMs) persuade in multi-turn conversations. They found these probes can efficiently identify when persuasion occurs, detect personality traits of the person being persuaded, and recognize the persuasion strategies used. The study revealed differences in how persuasion unfolds in human-generated versus LLM-generated dialogues and identified correlations between personality traits like extroversion and the effectiveness of certain persuasion strategies.
Large Language Models (LLMs) are increasingly demonstrating an ability to influence human beliefs and opinions, sometimes with an efficacy comparable to human communicators. This phenomenon, known as LLM-based persuasion, presents both opportunities, such as in education and therapy, and concerns, like political targeting and the spread of misinformation.
Despite this growing influence, there’s a limited understanding of how persuasion dynamics unfold in conversations involving LLMs. Traditional methods for analyzing such complex, high-level behaviors, especially in multi-turn dialogues, can be computationally expensive and inefficient. This is where a recent research paper, “How Do LLMs Persuade? Linear Probes Can Uncover Persuasion Dynamics in Multi-Turn Conversations”, introduces a novel approach.
Unlocking Persuasion with Linear Probes
The researchers, Brandon Jaipersaud, David Krueger, and Ekdeep Singh Lubana, leveraged a technique called linear probing. Linear probes are lightweight analytical tools used to study the internal representations of models. They have previously been applied to understand various LLM capabilities, such as modeling user sentiment or political perspectives. Inspired by insights from cognitive science, the team applied these probes to analyze persuasion dynamics in natural, multi-turn conversations.
The study focused on three distinct aspects of persuasion: persuasion success (whether the persuadee was convinced), persuadee personality (using the Big-5 framework), and persuasion strategy (categorized into logical, emotional, or credibility appeals based on Aristotle’s rhetorical triangle). By training separate linear probes for each of these dimensions, the researchers aimed to gain fine-grained, turn-level insights into how persuasion occurs.
The Experimental Approach
To train their probes, the researchers generated synthetic, multi-turn conversational data using GPT-4o, simulating interactions between a persuader and a persuadee across various contexts. For evaluation, they used two distinct datasets: DailyPersuasion (DP), which contains dialogues between GPT-powered agents, and PersuasionforGood (PfG), consisting of human-human interactions where one participant attempts to persuade another to donate to charity.
They compared the performance of their linear probes against zero-shot prompting of the underlying Llama-3.2-3b model (on which the probes were trained) and a GPT-4.1-Nano baseline. A significant advantage of linear probes highlighted in the study is their computational efficiency. When applied to precomputed model activations, probes run much faster than prompting-based methods, especially for analyzing data at finer granularities like token or turn levels.
Key Findings and Insights
The study yielded several compelling insights into persuasion dynamics:
-
Individual Conversation Analysis: Probes were able to identify critical moments in individual conversations. For instance, in an unpersuasive dialogue, a sharp drop in persuasion probability directly corresponded to explicit rejection by the persuadee. Conversely, in persuasive conversations, the persuadee’s agreement was reflected in smoother persuasion probability trajectories.
-
Personality and Persuasion: The probes revealed correlations between personality traits and persuasion outcomes. Conversations where the persuadee remained unpersuaded often showed consistently low agreeableness and high neuroticism scores throughout. This suggests that certain personality traits might make individuals more or less susceptible to persuasion.
-
Rhetorical Strategy Detection: The probes successfully detected the rhetorical strategies employed by the persuader. For example, high probabilities of ‘credibility’ appeals directly aligned with instances where the persuader used credibility-based arguments.
-
Dataset-Level Generalizations: The study also uncovered population-wide persuasion behaviors. In the human-human PfG dataset, persuasion signals (moments where persuasion typically occurs) concentrated around the middle turns of the conversation. In contrast, for the LLM-generated DP dataset, persuasion signals were predominantly found in the final one or two turns, indicating a systematic difference in how persuasion unfolds across natural versus synthetic data.
-
Strategy-Personality Correlations: By correlating probe outputs, the researchers found that extroversion was moderately correlated with various persuasion strategies across both datasets. Specifically, extroversion showed a positive correlation with emotional appeals and a negative correlation with credibility or logical appeals. This suggests that extroverted individuals might be more receptive to emotional persuasion and less so to logical or credibility-based arguments.
These findings demonstrate that linear probes offer an efficient and effective alternative for large-scale dataset analysis, providing insights that would be computationally prohibitive with prompting-based methods alone. The successful application of probes in this work suggests a promising avenue for future research into other complex LLM behaviors like deception and manipulation, especially in multi-turn settings.
Also Read:
- Shaping AI Dialogues: A New Framework for Controlling Conversation Quality in LLMs
- Conversational Manipulation: A New Threat to AI Alignment
Limitations and Future Directions
While the study provides valuable insights, the authors acknowledge limitations. They focused on a minimal set of persuasion strategies and the Big-5 personality framework; exploring more granular strategies or alternative personality models could yield further insights. Additionally, the experiments were conducted on Llama-3.2-3b, and testing larger models or different architectures would strengthen generalization claims. The researchers also highlight potential risks, noting that the insights into strategy-personality interactions could be misused to develop targeted persuasion techniques.
