Research map / Large language models / Language, safety and interpretability
Don’t Stop the Multi-Party! On Generating Synthetic Written Multi-Party Conversations with Constraints
Nicolò Penzo, Marco Guerini, Bruno Lepri, Goran Glavaš, Sara Tonelli
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 |
|---|---|---|---|
| generate turn by turn instead of one long pass (ablation) | Qwen2.5-7B-Instruct · WMPC generation | % compliant with all constraints: 20.32 -> 77.72 | Improved |
| generate turn by turn instead of one long pass (ablation) | Llama3.1-8B-Instruct · WMPC generation | % compliant with all constraints: 15.16 -> 66.52 | Improved |
| generate turn by turn instead of one long pass (ablation) | OLMo-2-7B-Instruct · WMPC generation | % compliant with all constraints: 0.04 -> 19.39 | Improved |
| generate turn by turn instead of one long pass (ablation) | Ministral-8B-Instruct · WMPC generation | % compliant with all constraints: 4.34 -> 0.87 | Hurt |
| generate turn by turn instead of one long pass (ablation) | Llama3.1-8B-Instruct · WMPC language variability | Repetition Rate (lower better): 18.08 -> 11.07 | Improved |
| generate turn by turn instead of one long pass (ablation) | Qwen2.5-7B-Instruct · WMPC language variability | Semantic Coherence: 0.588 -> 0.604 | Improved |
| generate turn by turn instead of one long pass (ablation) | Llama3.1 and Qwen2.5 · Human evaluation (96 WMPCs) | quality ratings: Naturalness 4.46 -> 4.29 (Llama3.1); Stance Consistency 3.60 -> 4.21 (Qwen2.5) | Mixed |
| generate turn by turn instead of one long pass (ablation) | Llama3.1-8B-Instruct · WMPC generation | Number of Speakers compliance: 29.56 -> 97.00 | Improved |
| generate turn by turn instead of one long pass (ablation) | Llama3.1 and Qwen2.5 · Generation cost | computation time: TT 4 to 8 times longer than OL | Hurt |
| choose Qwen2.5 or Llama3.1 instead of Ministral or OLMo2 as generator (ablation) | Synthetic WMPC generation · WMPC generation | % compliant with all constraints: 77.72 vs 0.87 (best TT vs Ministral TT) | Improved |
Closest papers by meaning
- Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap (NeurIPS 2026 · similarity 0.86)
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.86)
- MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation (NeurIPS 2026 · similarity 0.85)
- LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset (ICLR 2024 · similarity 0.85)
- WildChat-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training (ICML 2025 · similarity 0.85)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.85)
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.
