Reproduce “EffComp: Efficient Prompt Compression via Hybrid Reinforcement & Supervised Learning” 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
Romtham Tungsununtham, Piyawat Lertvittayakumjorn, Atiwong Suchato, Peerapon Vateekul
EffComp compresses retrieved context for retrieval-augmented generation question answering by reranking documents, then selecting whole source sentences with a trained selector. Its reinforcement-learning and supervised-learning procedure rewards correct answers while favoring shorter contexts. On HotpotQA with Gemma-2-9B-IT, it reports 39.5× compression, 36.3% accuracy, and a C1 score of 0.53, versus CompAct’s 0.47; results are setup-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.