Research map / Large language models / Retrieval and knowledge
MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models
Chung-Ming Chien, Manu Orsini, Eugene Kharitonov, Neil Zeghidour, Karen Livescu, Alexandre Défossez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add asynchronous retrieval (MoshiRAG with Gemma 3 27B back end) to Moshi | Moshi 7B · HaluEval audio | accuracy (%): 10.5 -> 36.3 | Improved |
| fine-tune Moshi on RAG data without retrieval (ablation) | Vanilla Moshi · TriviaQA | accuracy (%): 22.8 -> 29.7 | Improved |
| swap retrieval back end from Gemma 3 27B to GPT-4.1 (ablation) | MoshiRAG · HaluEval audio | accuracy (%): 36.3 -> 51.3 | Improved |
| swap retrieval back end from Gemma 3 27B to Tavily web search (ablation) | MoshiRAG · TriviaQA | accuracy (%): 69.6 -> 77.5 | Improved |
| add RAG conversational template (lead portion before keyword) | MoshiRAG vs vanilla Moshi · keyword delay | E2EKD (s): 2.1 -> 3.1 | Hurt |
| fine-tune on RAG data and add retrieval | Moshi · Full-Duplex-Bench Pause | TOR: 0.99 -> 0.32 | Improved |
| fine-tune on RAG data with adversarial and distracting subsets (v2, v3) | Moshi · Full-Duplex-Bench User Interruption | GPT score: 0.77 -> 3.75 | Improved |
| summarize LLM reference before injecting it (ablation) | MoshiRAG · GSM8K (spoken) | accuracy (%): 33.9 -> 51.2 | Improved |
| use MoshiRAG with GPT-4.1 back end for out-of-domain math (ablation) | MoshiRAG · GSM8K (spoken) | accuracy (%): 33.9 -> 43.2 | Improved |
| retrieved reference vs final spoken response | MoshiRAG · QA benchmarks | accuracy (%): 83.0 -> 80.3 (LlamaQ) | Hurt |
Closest papers by meaning
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.86)
- MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation (NeurIPS 2026 · similarity 0.86)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.86)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.86)
- Language Model Can Listen While Speaking (AAAI 2025 · similarity 0.86)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
