AFlow: Affective Flow Language Model Enhances Emotional Support Conversations
AFlow is a newly proposed framework for emotional support conversation (ESC) that enhances dialogue prefixes through detailed supervision by modeling a continuous emotional flow across multiple turns. In contrast to current alignment methods that depend on limited outcome-level indicators, AFlow assesses intermediate utility across explored trajectories and develops transitions that align with user preferences. By utilizing a subpath-level flow-balance objective, preference signals are transmitted to intermediate states, which boosts the coherence of strategies and the quality of empathetic responses. Experimental results demonstrate notable and consistent advancements compared to leading baselines.
Key facts
- AFlow is a framework for emotional support conversation (ESC).
- It models a continuous affective flow along multi-turn trajectories.
- It provides fine-grained supervision on dialogue prefixes.
- Existing alignment schemes rely on sparse outcome-level signals.
- AFlow estimates intermediate utility over searched trajectories.
- It learns preference-consistent strategy transitions.
- A subpath-level flow-balance objective propagates preference signals to intermediate states.
- Experiments show consistent and significant improvements over competitive baselines.
Entities
—