PRISM: Personalized Recommendation of In-Context Strategies with Grounded Explanations for Effective Dialogue

This paper introduces PRISM, a personalized dialogue framework that recommends context-specific persuasive primitives with evidence-grounded reasoning, then scores delivered utterances conditioned on that recommendation. The central thesis is that conditioning strategy recommendation and evaluation on both latent speaker–audience embeddings and retrieved, explicit evidence from dialogue and persona context enables more accurate, robust, and transparent persuasive-strategy guidance than context-only or non-personalized baselines.

The paper proposes three technical contributions: oracle-driven dataset construction with outcome-grounded primitive labels and counterfactual scenario augmentation; a hybrid architecture combining separate actor and audience encoders with retrieval-augmented generation evidence; and joint training that aligns strategy generation, recommendation-aware scoring, and conditional contrastive learning.

Evaluation on persuasion benchmarks shows that PRISM outperforms non-personalized baselines on both strategy recommendation accuracy and utterance effectiveness prediction, while the grounded explanation component provides interpretable evidence that makes the system suitable for high-stakes communication assistance contexts.