Reproduce “PEER: A Collaborative Language Model” 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

Timo Schick, Jane Dwivedi-Yu, Zhengbao Jiang, Fabio Petroni, Patrick Lewis, Gautier Izacard, Qingfei You, Christoforos Nalmpantis and 2 more

PEER, a collaborative language model framework, simulates human iterative writing through a "Plan, Edit, Explain, Repeat" cycle. It addresses data scarcity by leveraging self-training with specialized infilling models, which enables fine-grained control, improved factual grounding via reference documents, and strong generalization across diverse text editing tasks and domains.

Key insights

  • Modeling text generation as an iterative, human-like 'Plan, Edit, Explain, Repeat' cycle significantly enhances control and interpretability for collaborative AI writing.
  • The scarcity of complex edit history data can be effectively overcome by training specialized infilling models (e.g., PEER-Undo to decompose plain text into synthetic edit histories), enabling robust self-training and domain generalization.
  • Explicitly integrating natural language plans and grounding generations with retrieved reference documents improves factual accuracy, reduces hallucination, and increases the trustworthiness of language model outputs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.