NeSy-RAG Compiles Retrieved Chunks Into Executable Prolog, Lifting ShARC Accuracy From 42.8% to 61.1% With Auditable Traces

arXiv 2608.06292·medium signalXBlueskyLinkedInCopy link

Rather than feeding retrieved text to an LLM and hoping the reasoning is sound, NeSy-RAG (arXiv 2608.06292, Aug 6) synthesizes a Prolog module per retrieved chunk, generating predicates that encode Boolean claims which may depend on user-specific facts, then retrieves and composes them into queries using joint natural-language/code embeddings. A symbolic knowledge-gap detector identifies missing user facts whose truth value would change the query outcome and automatically triggers a follow-up question — a concrete answer to the 'agent guesses instead of asking' problem. On ShARC with no domain-specific training it scores 61.1% versus 42.8% for a same-model RAG baseline, and every answer ships with an execution trace linking each step to its source.