MICA: A New RL Framework for Multi-Turn Emotional Support Dialogue
A novel framework called MICA (Multi-granularity Intertemporal Credit Assignment) has been introduced by researchers to enhance multi-turn emotional support dialogues using large language models, eliminating the need for a critic in reinforcement learning. MICA tackles the challenges of sparse rewards and ineffective credit assignment by utilizing a common potential function to derive both immediate and delayed credit based on the user's structured support state. It incorporates Incremental Distance Reward for assessing progress towards the target state and employs Monte Carlo returns to account for delayed impacts. These signals, once normalized, create a mixed advantage for consistent per-turn optimization without requiring matched-state comparisons. This framework is particularly aimed at long-horizon emotional support tasks that influence future user states. The paper can be found on arXiv with ID 2603.06194.
Key facts
- MICA stands for Multi-granularity Intertemporal Credit Assignment
- It is a critic-free RL framework for multi-turn emotional support dialogue
- Addresses sparse rewards and poor per-turn credit assignment in LLMs
- Uses Incremental Distance Reward to measure per-turn progress
- Monte Carlo returns capture delayed effects of actions
- Scope-specific normalization creates a mixed advantage signal
- No matched-state comparison is needed
- Paper available on arXiv: 2603.06194
Entities
Institutions
- arXiv