Research map / Large language models / Language, safety and interpretability
DiscoverLLM: From Executing Intents to Discovering Them
Tae Soo Kim, Yoonjoo Lee, Jaesang Yu, John Joon Young Chung, Juho Kim
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 SFT+DPO+GRPO training | Qwen3-8B · Creative Writing | Discovery score: 35.2 -> 45.2 | Improved |
| add DPO after SFT | Llama-3.1-8B SFT · SVG Drawing | Discovery score: 45.4 -> 51.6 | Improved |
| add DPO after SFT | Qwen3-8B SFT · Creative Writing | Discovery score: 34.9 -> 44.0 | Improved |
| add system prompt for intent discovery | Llama-3.1-8B Base · SVG Drawing | Discovery score: 45.6 -> 40.0 | Hurt |
| add system prompt for intent discovery | Qwen3-8B Base · Creative Writing | Discovery score: 35.2 -> 39.0 | Improved |
| fine-tune with SFT only | Llama-3.1-8B · Creative Writing | Interactivity score: 20.1 -> 92.3 | Mixed |
| add SFT+DPO on Technical Writing | Llama-3.1-8B · Technical Writing | Discovery score: 49.1 -> 49.0 | No clear effect |
| use DPO trained on Creative Writing for unseen tasks | Llama-3.1-8B · Unseen tasks | Discovery score: 47.8 -> 54.6 | Improved |
| use DiscoverLLM SFT+DPO on data visualization | Llama-3.1-8B · Data Viz (unseen) | Discovery score: 59.4 best baseline vs 58.8 | No clear effect |
| remove length penalty from reward (ablation) | Qwen3-8B offline DPO · Creative Writing | Discovery score | No clear effect |
| replace simulator backbone or rules (4 ablations) (ablation) | DiscoverLLM SFT+DPO · simulator ablations | Discovery score | No clear effect |
| use DiscoverLLM instead of Base in user study | Qwen3-8B · User study | share rating 8 or higher: 80% -> 84% | Improved |
Closest papers by meaning
- CollabLLM: From Passive Responders to Active Collaborators (ICML 2025 · similarity 0.89)
- IntentRL: Training Proactive User-intent Agents for Open-ended Deep Research via Reinforcement Learning (ICML 2026 · similarity 0.88)
- Eliciting Human Preferences with Language Models (ICLR 2025 · similarity 0.87)
- VibeCheck: Discover and Quantify Qualitative Differences in Large Language Models (ICLR 2025 · similarity 0.87)
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.87)
- Beyond Prompts: Dynamic Conversational Benchmarking of Large Language Models (NeurIPS 2024 · 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.
