Research map / Large language models / Language, safety and interpretability
Prompt-MII: Meta-Learning Instruction Induction for LLMs
Emily Xiao, Yixiao Zeng, Ada Chen, Chin-Jou Li, Amanda Bertsch, Graham Neubig
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 RL meta-training of the instruction generator | Prompt-MII-Zero (Llama3.1-8B) · 90 datasets, n=20 | macro-F1: 0.343 -> 0.433 | Improved |
| add RL meta-training of the instruction generator | Prompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=20 | macro-F1: 0.383 -> 0.441 | Improved |
| replace ICL n=100 with Prompt-MII n=20 | Llama3.1-8B · 90 datasets | macro-F1: 0.430 -> 0.433 | No clear effect |
| replace ICL with Prompt-MII at n=100 | Qwen2.5-7B · 90 datasets, n=100 | macro-F1: 0.482 -> 0.424 | Hurt |
| replace APE with Prompt-MII | Llama3.1-8B · 90 datasets, n=100 | macro-F1: 0.288 -> 0.405 | Improved |
| replace GEPA with Prompt-MII | Qwen2.5-7B · 90 datasets, n=50 | macro-F1: 0.346 -> 0.432 | Improved |
| increase demonstrations from 20 to 100 | Prompt-MII (Llama3.1-8B) · 90 datasets | macro-F1: 0.433 -> 0.405 | Hurt |
| use meta-prompt template 1 instead of 2 (ablation) | Prompt-MII-Zero (Llama3.1-8B) · 90 datasets, n=50 | macro-F1: 0.301 -> 0.354 | Improved |
| use meta-prompt template 2 instead of 1 (ablation) | Prompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=50 | macro-F1: 0.356 -> 0.387 | Improved |
| use Prompt-MII template in APE (ablation) | APE (Llama/Qwen follower) · 90 datasets, n=100 | macro-F1: 0.356 -> 0.384 | Improved |
| train generator on Llama but follow with Qwen instead of same-model Qwen | Prompt-MII · 90 datasets, n=20 | macro-F1: 0.441 -> 0.438 | Hurt |
| use Llama-trained generator for Qwen follower instead of untrained | Prompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=20 | macro-F1: 0.383 -> 0.438 | Improved |
Closest papers by meaning
- InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.89)
- WizardLM: Empowering Large Pre-Trained Language Models to Follow Complex Instructions (ICLR 2024 · similarity 0.89)
- SPaR: Self-Play with Tree-Search Refinement to Improve Instruction-Following in Large Language Models (ICLR 2025 · similarity 0.89)
- Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following (AAAI 2024 · similarity 0.88)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.88)
- Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models (ICLR 2025 · similarity 0.88)
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.
