CUHK and Microsoft unveil EchoChat: cascaded gated rewards curb error propagation in empathetic voice reasoning

新智元 · wechat · 2026-11-07

Researchers from CUHK and Microsoft introduced EchoChat (NeurIPS 2026), tackling error propagation in multi-stage empathetic voice dialogue, where early perceptual mistakes become the premise for later theory-of-mind reasoning and response generation, leading to "empathetic mimicry" that sounds warm but misses the speaker's needs.

Key techniques

Cascaded gated rewards: rewards for mental-state inference and strategy planning are conditioned on prior-stage reliability, preventing later plausible analysis from masking upstream errors;
SDCA (Step-Decomposed Credit Assignment): computes group-relative advantages per stage and assigns them to the corresponding token segments, replacing GRPO's single trajectory-level signal;
Acoustic-Anchored Attention: attention constraints during SFT keep perception grounded in acoustic cues (tone, rhythm, pauses) as long CoT drifts away from raw input.

Resources: a 400K-sample structured dataset (EchoDialogue-400K) and EchoEval, a hierarchical benchmark with human-recorded speech and psychology-expert annotations covering implicit emotion, semantic-tone conflicts, and emotional shifts. Ablations show the gated rewards and step-level credit assignment improve inter-stage alignment and reduce error propagation.

Original post →

CUHK and Microsoft unveil EchoChat: cascaded gated rewards curb error propagation in empathetic voice reasoning

More from Multimodal