Research map / Large language models / Agents, code and math
Training LLM Agents to Empower Humans
Evan Ellis, Vivek Myers, Jens Tuyls, Sergey Levine, Anca Dragan, Benjamin Eysenbach
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 |
|---|---|---|---|
| train with Empower logit-threshold completions instead of SFT-20 | Llama3.1-8B Instruct assistant · LiveCodeBench (Llama-3.3-70B simulated human) | Pass@1: 0.097 -> 0.282 | Improved |
| train with Empower instead of untrained Base | Llama3.1-8B Instruct assistant · LiveCodeBench (Llama-3.3-70B simulated human) | Pass@1: 0.170 -> 0.282 | Improved |
| train with Empower instead of untrained Base | Qwen3-14B assistant · LiveCodeBench (Gemma simulated human) | Pass@1: 0.086 -> 0.170 | Improved |
| train with Empower instead of untrained Base | Qwen3-8B assistant · LiveCodeBench (Gemma simulated human) | Pass@1: 0.092 -> 0.178 | Improved |
| cap suggestion length at 10 tokens (Base-10) instead of Base (ablation) | Qwen3-8B assistant · LiveCodeBench (Llama-3.3-70B simulated human) | Accept Ratio: 0.351 -> 0.592 | Improved |
| cap suggestion length at 10 tokens (Base-10) instead of Base (ablation) | Qwen3-8B assistant · LiveCodeBench (Llama-3.3-70B simulated human) | Pass@1: 0.183 -> 0.198 | Mixed |
| cap suggestion length at 10 tokens (Base-10) instead of Base (ablation) | Llama3.1-8B Instruct assistant · LiveCodeBench (Gemma simulated human) | Pass@1: 0.064 -> 0.064 | No clear effect |
| cap suggestion length at 10 tokens (Base-10) instead of Base (ablation) | Qwen3-14B assistant · LiveCodeBench (Gemma simulated human) | Pass@1: 0.086 -> 0.062 | Hurt |
| supervised finetune on next 20 human tokens (SFT-20) instead of Base | Llama3.1-8B Instruct assistant · LiveCodeBench (Llama-3.3-70B simulated human) | Pass@1: 0.170 -> 0.097 | Hurt |
| use Empower assistant instead of Base-20 with real users | Llama-3.1-8B assistant · Human study | acceptance rate: 6.18% -> 8.08% | Improved |
| use Empower assistant instead of Base-20 with real users | Llama-3.1-8B assistant · Human study | characters deleted per accepted suggestion: 12.91 -> 9.56 | Improved |
| use Empower assistant instead of Base-20 with real users | Llama-3.1-8B assistant · Human study | most relevant suggestions preference: 61% (p = 0.240) | No clear effect |
Closest papers by meaning
- Learning to Assist Humans without Inferring Rewards (NeurIPS 2024 · similarity 0.90)
- Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · similarity 0.87)
- Estimating the Empowerment of Language Model Agents (ICML 2026 · similarity 0.86)
- Self-Rewarding Language Models (ICML 2024 · similarity 0.84)
- Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis (AAAI 2024 · similarity 0.84)
- Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.84)
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.
