Multimodal counselor response generation (MCRG) aims to produce appropriate counselor replies from multimodal dialogue histories. Two major gaps limit progress: existing datasets rarely capture sustained, human‑recorded counseling sessions by qualified professionals, and current methods do not explicitly enforce consistency between counseling reasoning and the generated response, which may undermine system reliability.
To address these issues we introduce MOCC, a multimodal counseling conversation corpus containing over 200 hours of interactions with 154 credential‑verified counselors. Based on MOCC we propose MOCC‑R1, a two‑stage framework that reinforces reasoning‑response consistency. In the cold‑start supervised fine‑tuning stage the model learns to output a structured trajectory: understanding the client’s state, a response intent that links a counseling principle to a planned action, and the final response. Reinforcement learning (RL) then rewards grounded plan coherence and execution, encouraging the inferred state and plan to be supported by the dialogue context and the response to realize that plan. Experiments demonstrate that MOCC‑R1 outperforms baselines in both consistency and practicality.
Review