LatWeave organizes knowledge into a multidimensional lattice and compiles multi‑hop QA into three deterministic operators—meet (constraint intersection), compare (lattice‑order comparison), and abstain (structural abstention). LLMs are used only for one‑shot extraction during lattice construction and for query‑planning hints; the answer‑generation path contains no LLM calls and requires no task‑specific training, making the whole pipeline auditable and reproducible.
We evaluate the paradigm on six public benchmarks to delineate its operating envelope. When the knowledge base is complete (MetaQA, 39,093 questions), three‑hop meet chains achieve an any‑hit of 0.9975, on par with fully supervised KBQA. On the templated multi‑hop set 2WikiMultihopQA (held‑out n=1,258) we obtain EM 0.865, well above published structure‑augmented RAG results. Performance degrades on open‑text deep‑composition tasks (MuSiQue) and on extraction‑coverage gaps (HotpotQA), which we attribute to factors outside the lattice‑algebra layer. For incomplete information (IIRC) we achieve structural abstention with abstain accuracy 0.971 and leak rate 0.029. Within this envelope, deterministic execution incurs no performance penalty, and every step can be recomputed, providing end‑to‑end auditability.
Review: LatWeave proves that deterministic lattice algebra can replace probabilistic reasoning in multi‑hop QA, delivering reproducible, zero‑LLM answers while preserving strong performance on well‑covered domains.