Persuasion Propagation in LLM Agents

TL;DR

This work defines persuasion propagation as belief states persisting across tasks in multi-step LLM agents and introduces a behavior-centered evaluation framework that separates on-the-fly persuasion from prefilled belief conditioning. It demonstrates, across coding and web research tasks, that transient on-the-fly persuasion yields weak aggregate changes, while persisting belief states introduced before task execution produce more pronounced behavioral shifts in exploration patterns. Notably, belief-prefilled agents reduce information-seeking activity, with quantified effects such as a $26.9\%$ decrease in searches and a $16.9\%$ decrease in unique sources compared to neutral-prefilled baselines, highlighting the behavioral impact of belief integration. The study argues for process-level auditing and behavior-focused metrics to assess robustness and safety in agentic systems, and establishes a scalable methodology to examine how beliefs propagate into action rather than merely into stated stances.

Abstract

Modern AI agents increasingly combine conversational interaction with autonomous task execution, such as coding and web research, raising a natural question: what happens when an agent engaged in long-horizon tasks is subjected to user persuasion? We study how belief-level intervention can influence downstream task behavior, a phenomenon we name \emph{persuasion propagation}. We introduce a behavior-centered evaluation framework that distinguishes between persuasion applied during or prior to task execution. Across web research and coding tasks, we find that on-the-fly persuasion induces weak and inconsistent behavioral effects. In contrast, when the belief state is explicitly specified at task time, belief-prefilled agents conduct on average 26.9\% fewer searches and visit 16.9\% fewer unique sources than neutral-prefilled agents. These results suggest that persuasion, even in prior interaction, can affect the agent's behavior, motivating behavior-level evaluation in agentic systems.

Figures (10)

• Figure 1: Overview of On-the-fly Persuasion Propagation. An agent’s stance is first probed, then exposed to a task-irrelevant persuasive statement. The agent subsequently performs coding or web research tasks. Conditioned on whether the agent adopts the injected stance, its execution dynamics can differ (e.g., more revisions, more frequent strategy changes or searches concentrated in fewer domains).
• Figure 2: Pipeline to Study Persuasion Propagation.
• Figure 3: Prompt template for probing model stance. This probe is used to assess expressed belief before and after persuasion exposure. Stance probing evaluates belief adoption and persistence, but does not measure downstream behavior, motivating the need for behavior-level analysis in the main experiments. Here, the pair['topic'], pair['A'], and pair['B'] correspond to the Topic, Claim A, and Claim B in \ref{['tab:claim-pairs']}, respectively.
• Figure 4: Prompt template for generating persuasive claims. Claims are conditioned on topic, stance shift, and persuasion tactic, and constrained to a single concise sentence.
• Figure 5: Belief prefill prompt templates. Agents are conditioned into one of three initialization regimes: belief (B), disbelief (NB), or neutral exposure (C0). These templates are prepended to task prompts to study how persistent belief states, rather than transient persuasion, occur during an active task.