Researchers studied 15,000 closed-loop rollouts from four vision-language-action policies to analyze end-effector geometry and task success. They found that successful policy pairs have a median normalized dynamic time warping distance of 0.0120 m, while unsuccessful pairs are more separated. This suggests that task success does not guarantee physical execution agreement between policies.

This research explores the concept of world modeling in transformer models, specifically in the context of a transformer trained on random walks through Manhattan. The study reveals that the model represents its environment, tracks its position, and uses a goal compass to navigate, contradicting previous claims of incoherent internal maps. The findings suggest that world-modeling capacities emerge at different stages of training and propose mechanistic indicators for comparing models.

This research introduces Evolutionary Ensemble Search (EES), a system that constructs machine-learning procedures through expert-guided program evolution. EES uses a council to guide the evolution process, allocating tasks to execution specialists and an evolutionary engine that selects and mutates code. The system adapts through session memory and problem-indexed lessons, and achieves a high success rate on various tasks. This contributes a concrete architecture for cumulative executable search and a versioned account of its cross-modal development outcomes.

A new algorithm, SSLD, improves DSATUR, a heuristic for the Graph Coloring Problem, by preprocessing a first good color class. This is relevant to AI agents because it may lead to future improvements in graph coloring algorithms used in AI applications.

Researchers propose SAMPLESELECT, a method for improving few-shot audio classification by predicting a feature mask for each input. This can help mitigate representation shift when foreground-background correlations change. The method is evaluated on ResNet12 and Conv64 models with significant improvements in out-of-distribution accuracy.