The paper introduces Vintix, a fixed cross-domain action model designed to learn behaviors at inference time through in-context reinforcement learning. It argues that Algorithm Distillation can scale this paradigm beyond toy tasks and single-domain settings, and reports that it provides a competitive alternative to expert distillation for building versatile action models. The work presents initial evidence that in-context reinforcement learning could support generalist decision-making systems, with code released and publication at ICML 2025.

The study proposes using offline reinforcement learning to improve generative drug design while avoiding costly interactions with a docking environment. Because protein-ligand docking is computationally expensive and online reinforcement learning typically requires hundreds of thousands of interactions, the authors investigate whether offline RL can optimize non-differentiable docking and chemical properties using existing data. The preliminary work focuses on conditionally generating compounds with improved docking performance and chemical characteristics.

A 3D hydrodynamic model was developed to estimate carbon dioxide fluxes over the non-uniform Staroselsky Moch peatland in the Tver region of European Russia. The model, which uses a 'one-and-a-half' closure scheme, effectively describes the spatial distribution of wind velocities and CO2 fluxes, accounting for surface heterogeneity. The study found significant effects of vegetation heterogeneity on wind and CO2 flux patterns, with maximum airflow disturbances at forest edges. The model showed good agreement with eddy covariance flux measurements, indicating its potential for practical applications in environmental assessments and greenhouse gas flux up-scaling.