CUHK and Microsoft's EchoChat hits 8.64 on EchoEval for empathetic speech dialogue
jiqizhixin · x · 2026-11-05
CUHK and Microsoft researchers present EchoChat (NeurIPS 2026), structuring empathetic spoken dialogue as a chain: perceive affective cues, reason about the speaker's state and needs, then choose a supportive response — scoring 8.64/10 overall on EchoEval with separate assessments of emotional perception, reasoning and spoken response.
- SFT: Acoustic-Anchored Attention keeps reasoning grounded in audio, catching cues a transcript misses (silence itself is expressive; only the response is speech-decoded).
- RL: Step-Decomposed Credit Assignment gives stage-specific rewards — gates mask later-stage rewards when earlier stages fail, instead of rewarding plausible-sounding answers built on a wrong foundation.
- Data/benchmark: trained on EchoDialogue-400K; EchoEval is built from human-recorded speech with expert annotation.
More from Research
- Meta AI proposes Off-Context GRPO to fix GRPO's zero-gradient 'learning cliff' on hard problems — ceciletamura · 2026-11-05
- Template-based multilingual safety pipeline shows English-only evals misjudge alignment — swiss-ai · 2026-11-05
- 'Grafting' paper: SFT on donor checkpoints + weight merging beats on-policy self-distillation — AdtRaghunathan · 2026-11-05
- Physical Intelligence's RLT uses RL tokens to make robot fine manipulation 3x faster — svlevine · 2026-11-05
- FrankenSim's physics-based flute synthesis gets graded against real recordings — doodlestein · 2026-11-05
- Randomized Algorithms Everywhere: Why Randomness Beats Determinism in Production — arpit_bhayani · 2026-11-05