TLDR: This research paper introduces a theory-driven framework, grounded in Bayesian Persuasion, to systematically evaluate and train large language models (LLMs) in strategic persuasion. By repurposing human-human persuasion datasets, the study creates environments where LLMs act as both persuaders (Senders) and decision-makers (Receivers). Findings show that frontier LLMs achieve high persuasion gains and exhibit sophisticated strategies. Furthermore, reinforcement learning significantly enhances the persuasive capabilities of even smaller LLMs, demonstrating that models can learn effective information design principles. The work highlights the importance of adaptive strategies and intermediate prior beliefs for successful persuasion, offering a principled approach to understanding LLM influence.
Large language models, or LLMs, have shown a remarkable ability to persuade, sometimes even matching human capabilities. This presents exciting possibilities, but also raises important questions about how these powerful tools might be used. A new research paper from the University of Illinois Urbana-Champaign dives deep into this topic, offering a structured way to understand and improve how LLMs persuade.
The paper, titled “TOWARDS STRATEGIC PERSUASION WITH LANGUAGE MODELS,” by Zirui Cheng and Jiaxuan You, addresses a key challenge: it’s hard to consistently measure how good LLMs are at persuasion because its effectiveness can vary so much depending on the situation. To tackle this, the researchers adopted a theory-driven approach, using a concept called Bayesian Persuasion.
Understanding Persuasion Through a Strategic Lens
At its core, Bayesian Persuasion views persuasion as a strategic interaction. Imagine one party, the “Sender” (in this case, an LLM), wants to influence the decisions of another party, the “Receiver” (also an LLM in their experiments, validated by human studies). The Sender doesn’t force decisions or offer bribes; instead, it strategically reveals information to guide the Receiver’s beliefs towards an outcome favorable to the Sender. This framework suggests that carefully chosen, partial information disclosure can often be more effective than either full transparency or complete secrecy.
The researchers developed specific ways to measure an LLM’s persuasive power. They looked at “persuasion gains,” which is how much the Sender’s desired outcome improves compared to if no persuasion happened. They also analyzed “persuasion signals,” examining how much relevant information the LLM chooses to reveal over time, indicating how adaptive and strategic its communication is.
Building a Benchmark and Training LLMs
To test their ideas, the team created a benchmark focused on “opinion change” tasks. They repurposed existing datasets from human-human persuasion scenarios, such as debates and discussions on controversial topics. In these environments, LLMs played both the role of the Sender and the Receiver. A human study involving 45 participants confirmed that the LLMs’ belief-updating processes were generally reasonable, making the setup plausible for evaluation.
Beyond just evaluating, the paper also explored how to make LLMs better persuaders. They used reinforcement learning (RL), a method where models learn by trial and error, receiving rewards for successful persuasion. By training a Sender LLM against a Receiver LLM, they found that even smaller LLMs could significantly boost their persuasive abilities. This improvement wasn’t just specific to the training partner; the enhanced persuasion skills transferred to different Receiver models, suggesting the LLMs learned general strategic principles rather than just exploiting a specific model’s quirks.
Also Read:
- Multiplayer Nash Preference Optimization: A Robust Framework for LLM Alignment
- TruthRL: A Framework for More Reliable Language Models
Key Findings from Experiments
The experiments revealed several interesting insights:
- Model Size Matters: Larger, more advanced models like DeepSeek-R1, Claude 3.7 Sonnet, and GPT-4o consistently achieved higher persuasion gains than smaller models.
- Dynamic Interaction is Powerful: The biggest improvements in persuasion were seen in dynamic settings, where the LLMs could interact over multiple rounds. This highlights the importance of adaptive, evolving strategies in persuasion.
- Prior Beliefs Influence Success: Consistent with Bayesian Persuasion theory, persuasion was most effective when the Receiver’s initial beliefs were moderate, rather than strongly held. It’s easier to shift an opinion that isn’t already at an extreme.
- Sophisticated Signaling: Larger models demonstrated more diverse and adaptive signaling strategies over time, aligning with theoretical predictions for effective persuasion.
- Common Strategies: LLMs primarily relied on strategies like providing “evidence,” establishing “credibility,” and emphasizing “impact” in their persuasive messages.
- Receiver Susceptibility: While DeepSeek-R1 showed strong persuasive gains across various Receiver models, some models, like Mistral-7B, were found to be more susceptible to persuasion.
This research provides a foundational framework for understanding and enhancing the persuasive capabilities of LLMs. By bridging game theory with language model development, it offers a scientific lens to examine the societal impacts of LLM persuasion, encouraging the development of ethical safeguards and responsible deployment practices. You can read the full paper here: TOWARDS STRATEGIC PERSUASION WITH LANGUAGE MODELS.
