Technical Summary: User-Aware Active Knowledge Acquisition for Emotional Support Dialogue

1. Problem Formulation

Emotional support dialogue systems face a critical challenge: user needs are often weakly signaled, implicit, and evolve dynamically across multi-turn interactions. While Large Language Models (LLMs) possess strong reasoning capabilities, existing methods struggle to efficiently acquire and generalize emotionally relevant knowledge (EQ knowledge) under these conditions.

Current approaches suffer from three primary limitations:

1. Need Mismatch: Generic supportive responses or standalone intent inference often fail to align with the user's actual, evolving needs, leading to "pushback" (e.g., a user seeking empathy receiving advice).
2. Passive Knowledge Acquisition: Strategy-memory and agent-style planning methods often follow passively acquired LLM behaviors, resulting in redundant feedback and limited coverage of the EQ knowledge space.
3. Static User Modeling: Storing EQ knowledge in action-agnostic forms ignores that user feedback is contingent on the assistant's chosen response. Similar surface emotions may require different user-aware strategies, yet existing systems fail to distinguish these nuances.

The core problem is defined as knowledge-driven multi-turn emotional support dialogue under uncertain user needs. The system must simultaneously provide helpful support and interact in a way that improves its ability to infer latent user needs and acquire reusable EQ knowledge, all without relying on gradient-based policy updates.

2. Methodology: User-Aware Active Knowledge Acquisition (UKA)

The authors propose UKA, a gradient-free, active dialogue learning framework that treats emotional support as both a support process and an active knowledge-acquisition process. UKA operates via a three-stage pipeline:

A. EQ Knowledge Base Representation

The system maintains an external memory $K$ of EQ knowledge entries in a state-action form $k = (a, v)$:

• Anchor ($a$): A belief-aware summary describing the emotionally salient dialogue state and inferred user needs.
• Value ($v$): An actionable response strategy or behavior distilled from interaction feedback.
  Unlike factual knowledge, EQ knowledge is evaluated by interaction effectiveness rather than objective truth.

B. Stage 1: User Belief Update (Theory-of-Mind)

UKA maintains an explicit hypothesis set $U_t = \{u^{(1)}_t, \dots, u^{(M)}_t\}$ representing plausible user needs.

• Hypothesis Scoring: For each hypothesis, the system computes a "teacher-forced" compatibility score by simulating how likely the observed user reply is under that specific need.
• Belief Distribution: These scores are normalized to form a belief distribution $p_t(u|H_t)$.
• Refresh Mechanism: The hypothesis set is dynamically refreshed (expanded or pruned) if the current set is incomplete, ambiguous, or if a new hypothesis yields a higher compatibility score.

C. Stage 2: Belief-Aware Knowledge Retrieval

Instead of retrieving knowledge based solely on dialogue history, UKA constructs a summary anchor $s_t$ that conditions on both the dialogue context and the current belief distribution over user needs. This anchor retrieves the top-$K$ relevant EQ knowledge entries from the external memory.

D. Stage 3: Uncertainty-Driven Active Response Selection

Response selection is formulated as an active learning problem, with distinct objectives for training and testing:

1. Candidate Generation: The model generates a set of candidate responses $C_t$ conditioned on history and retrieved knowledge.
2. Scoring Metrics:
   • Knowledge Support ($S_k$): Measures how well a candidate is grounded in the retrieved knowledge (cosine similarity).
   • ToM Uncertainty ($S_u$): Measures the "diagnostic value" of a response. It quantifies the disagreement in simulated user reactions across competing user-need hypotheses. A high $S_u$ indicates the response is likely to elicit distinct feedback, helping to disambiguate the user's latent need.
3. Selection Strategy:
   • Training (Exploration): Selects responses to maximize $S_u - S_k$. This prioritizes responses that expose gaps in current knowledge and elicit informative feedback to expand the EQ memory.
   • Testing (Exploitation): Selects responses to maximize $S_k + \gamma S_u$. This prioritizes knowledge-grounded responses for stable support while retaining mild diagnostic value to handle uncertainty.

3. Key Contributions

1. UKA Framework: Introduction of a gradient-free, active dialogue framework that explicitly separates exploratory EQ knowledge acquisition from test-time knowledge usage.
2. ToM Uncertainty Mechanism: A novel diagnostic uncertainty estimation method that tracks user-need hypotheses and guides response selection via hypothesis-conditioned reaction disagreement, allowing the system to actively elicit informative feedback.
3. Empirical Validation: Demonstration that UKA consistently outperforms strong baselines (including prompting, memory-based, and multi-agent ToM approaches) across multiple emotional support benchmarks (ESConv, ExTES, Sentient Eval) and diverse LLM backbones.

4. Experimental Results

The framework was evaluated on four LLM backbones (Qwen3-32B, Seed-1.6-36B, Qwen3-235B, GPT-OSS-120B) across three benchmarks.

• Performance: UKA consistently achieved the highest Success Rate (SR) on ESConv and ExTES and the highest average emotion scores on Sentient Eval, including challenging "Negative" persona settings where users are resistant or skeptical.
• Ablation Studies:
  • Removing Knowledge acquisition caused the largest performance drop, confirming that gains stem from acquiring/reusing EQ knowledge rather than prompting alone.
  • Removing Active Response selection or ToM Uncertainty significantly degraded performance, particularly on negative personas, highlighting the necessity of candidate-level exploration and explicit belief tracking.
• Robustness: UKA maintained superior performance when evaluated with different user simulators (Cross-Simulator Robustness), suggesting the acquired knowledge captures transferable strategies rather than overfitting to a specific simulator.
• Human Evaluation: Blind pairwise human preference evaluations showed UKA was preferred over baselines in the majority of cases (average win rate ~57-65%), with moderate inter-annotator agreement.
• Knowledge Dynamics: Analysis showed that performance improves monotonically as the knowledge base grows, with the most significant gains occurring in the early stages of knowledge accumulation.

5. Significance and Claims

The paper claims that uncertainty-aware interaction is a promising mechanism for enabling dialogue agents to acquire, refine, and utilize emotional intelligence knowledge in simulated settings without requiring parameter updates.

• Active vs. Passive: Unlike passive memory systems that merely store interactions, UKA actively selects responses to disambiguate user needs and expand its knowledge base, leading to broader and more diverse action repertoires.
• User Alignment: By modeling user needs as a belief distribution and using ToM uncertainty to guide exploration, the system achieves better alignment with implicit and evolving user preferences.
• Practicality: The framework is gradient-free, making it applicable to various LLM backbones without the computational cost of fine-tuning.

Limitations and Future Directions:
The authors acknowledge several limitations:

• Approximate Modeling: The use of a small set of natural-language hypotheses is an approximation of real, complex user motivations.
• Simulation Bias: Reliance on LLM-based user simulators and critics may introduce biases and fail to capture real-world psychological defense mechanisms or long-term emotional dynamics.
• Efficiency: The framework incurs extra inference costs for hypothesis tracking and ToM uncertainty estimation, and the external memory may accumulate noisy knowledge, necessitating future work on pruning and validation.

The paper concludes that while UKA demonstrates significant improvements in user alignment and dialogue quality, real clinical deployment would require further safety validation, knowledge filtering, and human-in-the-loop safeguards.