Make sense of what AI is becoming.

Researchers introduce DyMAB, a dynamic multi‑armed bandit algorithm that enhances ALNS-based MAPF solvers by adapting neighborhood selection over time.

Researchers show linear existential rules keep path query answering in NL data complexity while guarded rules push it to 2ExpTime combined complexity.

A new study warns that neuro‑symbolic models can reach high accuracy by misgrounding concepts, compromising trustworthiness.

JAIR’s July 27 2026 survey reviews symbolic XAI in deep neural nets, mapping hybrid methods and noting evaluation limitations.

A new JAIR study finds that synthetic datasets produced by large language models enable neural predictors to forecast human behavior in strategic persuasion games more accurately than when trained on actual experimental data.

A new reinforcement‑learning based recommender called GEPR uses group knowledge to improve resource suggestions in online learning platforms.

Researchers present a framework that ranks sets of abstract arguments based on plausibility, extending Dung's 1995 semantics.

Researchers propose generating reinforcement‑learning rewards directly from formal specifications using a new quantitative semantics of STL.

Singh and Saha demonstrate how Signal Temporal Logic specifications can be turned into reward functions for reinforcement learning agents.

Researchers use large language models to automatically craft better heuristics for SAT solvers.

A new hierarchical reinforcement learning framework uses flow‑based generative models for direct off‑policy correction, improving data efficiency and convergence speed.

Researchers show a data-guided SVD approach can shrink deep neural networks to under three times their original size without fine‑tuning, preserving performance.