AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: In-context learning analysis · 10 citations (OpenAlex)

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 changedWhereMeasuredEffect
add RL meta-training of the instruction generatorPrompt-MII-Zero (Llama3.1-8B) · 90 datasets, n=20macro-F1: 0.343 -> 0.433Improved
add RL meta-training of the instruction generatorPrompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=20macro-F1: 0.383 -> 0.441Improved
replace ICL n=100 with Prompt-MII n=20Llama3.1-8B · 90 datasetsmacro-F1: 0.430 -> 0.433No clear effect
replace ICL with Prompt-MII at n=100Qwen2.5-7B · 90 datasets, n=100macro-F1: 0.482 -> 0.424Hurt
replace APE with Prompt-MIILlama3.1-8B · 90 datasets, n=100macro-F1: 0.288 -> 0.405Improved
replace GEPA with Prompt-MIIQwen2.5-7B · 90 datasets, n=50macro-F1: 0.346 -> 0.432Improved
increase demonstrations from 20 to 100Prompt-MII (Llama3.1-8B) · 90 datasetsmacro-F1: 0.433 -> 0.405Hurt
use meta-prompt template 1 instead of 2 (ablation)Prompt-MII-Zero (Llama3.1-8B) · 90 datasets, n=50macro-F1: 0.301 -> 0.354Improved
use meta-prompt template 2 instead of 1 (ablation)Prompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=50macro-F1: 0.356 -> 0.387Improved
use Prompt-MII template in APE (ablation)APE (Llama/Qwen follower) · 90 datasets, n=100macro-F1: 0.356 -> 0.384Improved
train generator on Llama but follow with Qwen instead of same-model QwenPrompt-MII · 90 datasets, n=20macro-F1: 0.441 -> 0.438Hurt
use Llama-trained generator for Qwen follower instead of untrainedPrompt-MII-Zero (Qwen2.5-7B) · 90 datasets, n=20macro-F1: 0.383 -> 0.438Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy