JAIR releases three fresh AI research papers on constraints, GPs, and games

In its latest issue, the Journal of Artificial Intelligence Research (JAIR) presents three novel studies that advance practical artificial intelligence techniques.

Efficient reformulations of half‑reified global constraints using auxiliary variables Authors: Ignace Bleukx, Hélène Verhaeghe, Dimos Tsouros, Tias Guns – published 2026/08/15. The paper tackles a common bottleneck in declarative constraint solving: when reified or half‑reified global constraints are flattened into non‑global forms, solvers lose access to specialized propagators and solve times can increase dramatically. By introducing auxiliary variables, the authors keep some global structure intact during decomposition, allowing constraint programming (CP) models to scale more effectively than before. This optimization is especially relevant for large data‑science applications such as logistics planning, where problems routinely involve thousands or millions of variables. ¹

Information‑based calibration of uncertainty quantification in product‑of‑experts Gaussian process models Authors: Yean Hoon Ong, Paolo Barucca, Wei Pan, Jun Wang – published 2026/08/15. Overestimation of posterior variances has been known since Murphy’s Bayesian Hierarchical Modelling (2003). The new GP‑pro‑c model addresses this by incorporating an information‑gain based calibration ratio that reduces each local expert’s variance without retraining all experts at every iteration—a step that would otherwise be O(n²) with n training samples. Early experiments show up to a 5 % improvement in negative log‑likelihood compared with baseline methods across several datasets. ¹

Variety‑seeking jump games on graphs Authors: Lata Narayanan, Jaroslav Opatrny, Shanmukha Tummala, Alexandros A. Voudouris – published 2026/08/15. The authors study a class of jump games where agents of different types occupy graph nodes and seek to maximize the variety of neighboring types. They prove that while the induced game can have improving response cycles, it is a potential game under four specific conditions—only two agent types; exactly one empty node; maximum degree two; or a 3‑regular graph with two empty nodes. Moreover, they establish equilibria existence on trees, cylinder graphs, and tori, and provide tight bounds on price‑of‑anarchy for social welfare and colorful edges. ¹