Reproduce “AIR: Complex Instruction Generation via Automatic Iterative Refinement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wei Liu, Yancheng He, Hui Huang, Chengwei Hu, Jiaheng Liu, Shilong Li, Wenbo Su, Bo Zheng
The AIR framework generates complex instructions by iteratively refining simple instructions, drawing constraints from human-written documents using an LLM-as-judge mechanism. Fine-tuning models with the resulting AIR-10K dataset led to consistent performance improvements on complex instruction-following benchmarks, including CF-Bench and FollowBench, over existing automatic and human-curated datasets.
Key insights
- Human-written documents implicitly contain rich preferences and formatting conventions that can be extracted and converted into explicit constraints for instruction generation.
- Mimicking human iterative refinement by progressively adding constraints based on a model's performance gaps is effective for generating complex and challenging instructions.
- Using an LLM as a judge to compare model responses against human-preferred document references is crucial for identifying valuable constraints that truly highlight model insufficiencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.