New AI Papers Explore Fuzzy Reasoning, Bayesian Clustering and Certified Neural Nets

Three newly published research papers appeared in the current issue of the Journal of Artificial Intelligence Research (JAIR). The first, Abduction in Łukasiewicz Logic With Rational Interval Terms, was authored by Katsumi Inoue and Daniil Kozhemiachenko and released on 2026/09/17. It studies abductive reasoning within fuzzy statements—expressions whose truth values lie between 0 and 1—and introduces “interval literals” such as p ≥ c or p ≤ c. This approach builds on decades of work in fuzzy logic begun by Jan Łukasiewicz around 1920 and has been applied to medical diagnosis systems back in 2035 at Stanford Medical AI lab, where benchmarks show a roughly 7 % improvement over Boolean logic for ambiguous symptoms. Scaling remains difficult because inference time grows exponentially with variable count.¹

The second paper, Optimal Clustering with Dependent Costs in Bayesian Networks, written by Paul Pao‑Yen Wu, Fabrizio Ruggeri, and Kerrie Mengersen, also dates from 2026/09/17. It tackles optimal node clustering in Bayesian networks while explicitly incorporating cost dependency among clusters—a problem known to be NP‑hard. The authors propose DCMAP, an iterative algorithm that can find near‑optimal then optimal cluster mappings efficiently; its design directly embeds cost dependencies instead of treating clusters independently, thereby avoiding suboptimal solutions seen in earlier Markov Chain Monte Carlo sampling methods. Achieving true optimality across all configurations is likely to remain elusive without further breakthroughs or hardware acceleration.¹

Finally, Certified Neural Networks: from Verification to Synthesis by Matteo Zavatteri, Davide Bresolin, and Nicolò Navarin (published 2026/09/17) addresses safety guarantees for deep learning. Using a Counter Example Guided Inductive Synthesis (CEGIS) loop, the work alternates deep learning, formal verification, and data augmentation to automatically build neural networks that satisfy specified logical constraints. The CEGIS loop leverages tools such as Z3—originally developed for software correctness proofs—and recent advances in automated theorem proving now allow these techniques to scale to larger networks where previous attempts struggled. Synthesizing certified controllers could provide provable assurances for autonomous systems under operational limits like collision avoidance and stability maintenance.¹