Research map / Large language models / Alignment and preferences
Flipping the Dialogue: Training and Evaluating User Language Models
Tarek Naous, Philippe Laban, Wei Xu, Jennifer Neville
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 |
|---|---|---|---|
| use trained UserLM-8b instead of GPT-4o as user simulator | GPT-4o assistant · GSM8K and HumanEval simulated conversations | assistant score (%): 74.6 -> 57.4 | Hurt |
| use UserLM-8b instead of GPT-4o as simulator | user simulation · PRISM dialogue termination | F1: 1.38 -> 63.54 | Improved |
| train a user LM (UserLM-8b) instead of prompting GPT-4o | user simulation · Naturalness | Pangram human-likelihood (%): 3.31 -> 80.21 | Improved |
| train UserLM-8b instead of prompting GPT-4o | user simulation · First-turn diversity | 1-gram Jaccard (%): 74.42 -> 94.55 | Improved |
| train UserLM-8b instead of prompting GPT-4o | user simulation · Intent decomposition | 1-gram overlap (%): 7.68 -> 2.69 | Improved |
| train UserLM-8b instead of prompting GPT-4o | user simulation · User role adherence | accuracy (%): 38.85 -> 93.95 | Improved |
| train UserLM-8b instead of prompting GPT-4o | user simulation · Intent adherence | accuracy (%): 70.95 -> 94.65 | Improved |
| scale UserLM from 1b to 8b | UserLM · PRISM perplexity with intent | perplexity: 18.58 -> 14.92 | Improved |
| scale the prompted assistant from GPT-4o-mini to GPT-4o | prompted user simulator · Six intrinsic metrics | metric wins | No clear effect |
| train UserLM from instruction-tuned checkpoint instead of base (ablation) | UserLM · PRISM perplexity | perplexity | Hurt |
| start training from 1b instead of 8b instruction-tuned assistant (ablation) | UserLM from instruct checkpoint · PRISM perplexity | perplexity: 27.25 -> 18.45 | Improved |
| train with intent conditioning (ablation) | UserLM · PRISM perplexity | perplexity | Improved |
Closest papers by meaning
- Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap (NeurIPS 2026 · similarity 0.88)
- CollabLLM: From Passive Responders to Active Collaborators (ICML 2025 · similarity 0.88)
- MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback (ICLR 2024 · similarity 0.88)
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- ToolDial: Multi-turn Dialogue Generation Method for Tool-Augmented Language Models (ICLR 2025 · similarity 0.87)
- Broaden your SCOPE! Efficient Multi-turn Conversation Planning for LLMs with Semantic Space (ICLR 2025 · similarity 0.87)
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.
