Overview

• ReFactX introduces constrained generation that limits an LLM’s output to a pre‑indexed prefix tree of verbalized knowledge‑graph facts, removing the need for external retrievers or tools.
• Authors report scalability to knowledge bases of roughly 800 million facts with minimal generation‑time overhead, and they release implementation code on GitHub.
• The method is evaluated on question answering and is presented as adaptable to domain‑specific data while aiming to reduce hallucinations and pipeline complexity.
• A separate study builds a modular framework for travel mode choice prediction with Retrieval‑Augmented Generation, testing basic retrieval, balanced retrieval, cross‑encoder re‑ranking, and their combination.
• Across OpenAI models GPT‑4o, o4‑mini, and o3, the best result comes from GPT‑4o with balanced retrieval plus cross‑encoder re‑ranking at 80.8% accuracy on the 2023 Puget Sound survey, highlighting the need to align retrieval strategy with model capability.