To address error propagation across multi-stage reasoning in empathetic spoken dialogue, researchers from CUHK and Microsoft introduce EchoChat, a unified framework that organizes perception, theory-of-mind inference, strategy planning and speech response generation, constrains inter-stage dependencies with cascaded gated rewards, and applies step-level credit assignment for targeted updates. Accepted at NeurIPS 2026, the work reframes empathy as a decomposable cognitive process rather than a direct input-to-response mapping.
Theory of Mind
心智理论
The ability to infer others' beliefs, intentions, and mental states, often used to describe AI modeling of interaction partners.
Curated Intelligence
A team from the University of Tokyo introduced the MIDA benchmark for multimodal interactive deception assessment, accepted at EMNLP 2026. Tests reveal that leading multimodal LLMs like GPT and Claude struggle to reliably distinguish truth from lies in social interactions, despite processing audio-visual-text inputs. The study highlights significant gaps in models' ability to 'read the room' and understand theory of mind.
EchoChat decomposes empathetic voice dialogue into perception, Theory of Mind reasoning, and response generation, using a unified optimization framework to mitigate error propagation. The study aims to enhance AI's ability to move from superficial warmth to nuanced understanding and appropriate responses by integrating acoustic cues with contextual inference.
A new UC San Diego paper introduces a method to embed Theory of Mind into 7B agents by using an auxiliary model to score replies during RL training. This allows the agent to learn partner beliefs and goals without requiring extra inference models at deployment, improving intent recognition in multi-agent systems.
OpenAI and Oxford University introduced LOLA (Learning to Model Other Minds), an algorithm enabling reinforcement learning agents to predict and influence the parameter updates of other agents. This approach allows agents to discover reciprocal cooperation strategies in games like the iterated prisoner's dilemma, overcoming the limitations of traditional independent PPO methods that fail to form effective collaboration.