Skip to main content
Back

SI-AVPO Teaches AI to Persuade by Reading Users' Hidden Concerns

A new training framework teaches small LLMs to read between the lines and proactively steer conversations, beating GPT-4.1 and Claude on outbound sales benchmarks

Paper, Full Access
  • Core finding: Giving LLMs access to users' hidden concerns during training -- not at inference -- unlocks persuasive behavior that RL alone cannot produce.
  • New framework: SI-AVPO combines asymmetric self-distillation (AOPD) and turn-level state rewards (STPR) to train proactive dialogue agents from a Cognitive User Simulator.
  • Results: On Qwen3-8B, SI-AVPO beats the best RL baseline by +12.2% Call Success Rate and +12.2% Acceptance Rate on outbound-call benchmarks.
  • Beats frontier models: A small, deployable Qwen3-8B trained with SI-AVPO matches or exceeds GPT-4.1 and Claude on proactive dialogue metrics.
  • RL ceiling confirmed: GRPO, DAPO, and PPO all plateau on proactivity because they can only reinforce behaviors the passive base model already samples.
  • No code released yet: The paper is on arXiv; no weights or simulator code have been published alongside it.

Getting an LLM to answer questions is easy. Getting one to persuade someone -- to anticipate objections they haven't voiced yet and steer a conversation toward a goal -- is a fundamentally different problem. A new paper, "Unlocking Proactivity in Task-Oriented Dialogue", tackles exactly this gap, and its findings challenge some common assumptions about how reinforcement learning shapes LLM behavior.

The wall that RL keeps hitting

Among task-oriented dialogue scenarios, proactive settings such as outbound sales are especially challenging compared with conventional reactive ones: the agent must initiate or sustain the interaction, push the dialogue toward a predefined goal, and persuade users to take a target action. This is the kind of thing a skilled human sales rep does naturally -- reading the room, surfacing unspoken hesitations, and addressing them before they become objections.

Post-trained LLMs are inherently conservative, and reward-shaping RL (e.g., GRPO) struggles since it only re-weights what an already passive policy samples. GRPO (Group Relative Policy Optimization) is one of the most popular RL algorithms for post-training LLMs -- it works by generating multiple candidate responses and reinforcing the better ones. The problem: if the model never spontaneously produces a proactive, persuasive response in the first place, GRPO has nothing to reinforce.

Vanilla models (e.g., Qwen3-4B/8B) anchor at low operating points, and test-time scaling with Pass@16 leaves them clustered nearby, suggesting that more sampling alone cannot elicit the required persuasive behaviors. Reward-shaping RL consequently exhibits a similar ceiling, as it can only re-weight behaviors already sampled by the policy.

Scatter plot showing proactivity score vs acceptance rate, comparing reactive plateau against latent-concern regime

The missing variable: what the user is secretly thinking

The authors show that conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. In other words, the key ingredient isn't a smarter reward function or more rollouts -- it's giving the model access to what the user is privately worried about during training.