User-Aware Active Knowledge Acquisition for Emotional Support Dialogue

Recent advances in large language models (LLMs) like GPT-3 and BERT have significantly improved dialogue systems' fluency and coherence. However, emotional support requires models to recognize and adapt to users' implicit emotional needs, which are often weakly signaled and evolve over multiple turns. Prior works, such as retrieval-augmented generation (RAG) and memory-enhanced dialogue, have attempted knowledge integration but lack mechanisms for active exploration of emotional knowledge. Multi-agent ToM models have introduced reasoning about user beliefs but are computationally complex and limited in handling implicit needs. This paper builds on these foundations, emphasizing active knowledge acquisition driven by user need uncertainty, to address the persistent challenge of understanding and supporting users with subtle emotional cues.

The core challenge in emotional support dialogue is accurately inferring users' latent, often implicit needs that change dynamically. Existing methods tend to be passive, relying on static knowledge bases or reactive responses, which lead to redundant signals and poor personalization. The bottleneck lies in efficiently exploring the emotional knowledge space while balancing the need for robustness against distribution shifts. Without active disambiguation, systems risk providing mismatched support, reducing user satisfaction and trust. Developing a framework that actively probes user needs, expands knowledge, and maintains stability under diverse conditions remains an open problem.

The paper introduces UKA, which innovatively combines a gradient-free active learning strategy with a ToM-based uncertainty estimator. Key innovations include: 1) a diagnostic uncertainty mechanism that quantifies response discriminability across user need hypotheses, 2) a user-aware summary anchor for knowledge retrieval conditioned on dialogue context and belief states, 3) a response scoring strategy that balances knowledge support and uncertainty to guide active exploration during training, and 4) a phase-separated approach that distinguishes between knowledge acquisition and support during inference. These innovations enable the system to actively seek knowledge gaps, disambiguate user needs efficiently, and provide more personalized, emotionally intelligent responses.

• �� Construct a hypothesis set of user needs using language models, updating it based on user feedback likelihood scores.
• �� Calculate ToM uncertainty by simulating user reactions to candidate responses across hypotheses, measuring response discriminability.
• �� Retrieve relevant emotional knowledge from an external database by constructing a summary anchor that encodes dialogue history and current beliefs.
• �� Generate multiple candidate responses conditioned on retrieved knowledge and dialogue context.
• �� During training, select responses that maximize the difference in simulated user reactions minus knowledge support, promoting exploration.
• �� During testing, prioritize responses that are well-supported by knowledge and effectively disambiguate user needs, ensuring robustness.
• �� Update the knowledge base with new entries derived from feedback, enabling continuous learning without gradient updates.

Experiments utilize datasets ESConv, ExTES, and Sentient Eval, covering diverse emotional support scenarios. Models tested include Qwen3-32B, Seed-36B, Qwen3-235B, and GPT-OSS-120B. Baselines include prompting, principle-memory, and multi-agent ToM frameworks. Metrics such as success rate, average turns, and emotion scores evaluate performance. Ablation studies remove components like ToM uncertainty to assess their impact. Hyperparameters include candidate response count N=5, retrieval top-K=3, and specific training steps. Evaluation involves both automatic metrics and human-like judgment via GPT-4 and DeepSeek-V3, ensuring comprehensive assessment.

Across all models and datasets, UKA consistently outperforms baselines, with success rate improvements averaging 4.5% and emotion scores up by 12 points. Notably, in challenging negative personas, UKA maintains high performance, demonstrating robustness. Ablation results confirm the importance of ToM uncertainty, with a performance drop of approximately 8% without it. The method's ability to actively explore knowledge gaps and disambiguate needs leads to more personalized, emotionally appropriate responses, validating its effectiveness in real-world scenarios.

UKAs framework can be directly applied to mental health chatbots, customer service agents, and social robots, enhancing their ability to understand and support users' hidden emotional states. Its active knowledge exploration mechanism allows for continuous improvement and adaptation in dynamic environments. Long-term, integrating multi-modal signals and online learning can further extend its impact, enabling AI systems to become empathetic companions capable of nuanced emotional understanding and support in diverse contexts.

The reliance on simulated user models limits real-world applicability, as actual user behaviors are more unpredictable. Knowledge bases are currently limited in scope, risking incomplete coverage of emotional scenarios. Response selection strategies may struggle under extreme distribution shifts or highly novel needs, requiring more adaptive mechanisms. Future work should focus on real user data, larger knowledge repositories, and multi-modal integration to overcome these limitations.