· 3 items
arXiv ↗
MSI-Bench: Evaluating Multi-Speaker Voice Interaction for Collaborati…
- What
- MSI-Bench evaluates multi-speaker voice interaction for collaborative AI agents.
- Why
- Voice agents need robust multi-speaker support for real-world collaborative settings.
- Impact
- Informs voice agent multi-speaker architecture design
arXiv ↗
"MeBo Leaves a Piece of You Behind": Designing a Relational Voice-Bas…
- What
- MeBo is a fully functional relational voice-based memory companion designed for older adults.
- Why
- Demonstrates practical design patterns for long-horizon relational voice memory.
- Impact
- Provides patterns for relational voice memory features
arXiv ↗
onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs an…
- What
- onPanda is an interactive tool for efficiently annotating LLM alignment data via token-level correction.
- Why
- Token-level correction improves alignment and agent trajectory fine-tuning efficiency.
- Impact
- Speeds up agent behavior fine-tuning workflows