· 3 items
Nothing new in AI today — showing the latest AI picks, from 2026-09-18.
arXiv ↗
MTVA-Bench: Evaluating the Language Model Inside Cascaded Voice Agents
- What
- MTVA-Bench evaluates the language model component inside cascaded voice agents.
- Why
- It isolates LLM decision-making for voice from ASR and TTS components.
- Impact
- Improves voice agent core logic evaluation
arXiv ↗
Paint-Anything: Unified Any-Color Control for Image Generation and Ed…
- What
- Paint-Anything introduces unified any-color control for image generation and editing.
- Why
- It allows specifying target colors via 24-bit hex values in generation.
- Impact
- Adds precise color control to design tools
arXiv ↗
RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Ag…
- What
- RAFT introduces a stateful RAG framework for troubleshooting agents.
- Why
- It treats support cases as multi-stage, stateful workflows rather than static docs.
- Impact
- Better retrieval for stateful agent design