Learning to Collaborate: An Orchestrated-Decentralized Framework for Peer-to-Peer LLM Federation

Fine-tuning Large Language Models (LLMs) for specialized domains is constrained by a fundamental challenge: the need for diverse, cross-organizational data conflicts with the principles of data privacy and sovereignty. While Federated Learning (FL) provides a framework for collaboration without raw data exchange, its classic centralized form introduces a single point of failure and remains vulnerable to model inversion attacks. Decentralized FL (DFL) mitigates this risk by removing the central aggregator but typically relies on inefficient, random peer-to-peer (P2P) pairings, forming a collaboration graph that is blind to agent heterogeneity and risks negative transfer. This paper introduces KNEXA-FL, a novel framework for orchestrated decentralization that resolves this trade-off. KNEXA-FL employs a non-aggregating Central Profiler/Matchmaker (CPM) that formulates P2P collaboration as a contextual bandit problem, using a LinUCB algorithm on abstract agent profiles to learn an optimal matchmaking policy. It orchestrates direct knowledge exchange between heterogeneous, PEFT-based LLM agents via secure distillation, without ever accessing the models themselves. Our comprehensive experiments on a challenging code generation task show that KNEXA-FL yields substantial gains, improving Pass@1 by approx. 50% relative to random P2P collaboration. Critically, our orchestrated approach demonstrates stable convergence, in stark contrast to a powerful centralized distillation baseline which suffers from catastrophic performance collapse. Our work establishes adaptive, learning-based orchestration as a foundational principle for building robust and effective decentralized AI ecosystems.

Constructing Custom Thermodynamics Using Deep Learning

One of the most exciting applications of AI is automated scientific discovery based on previously amassed data, coupled with restrictions provided by the known physical principles, including symmetries and conservation laws. Such automated hypothesis creation and verification can assist scientists in studying complex phenomena, where traditional physical intuition may fail. Of particular importance are complex dynamic systems where their time evolution is strongly influenced by varying external parameters. In this paper we develop a platform based on a generalised Onsager principle to learn macroscopic dynamical descriptions of arbitrary stochastic dissipative systems directly from observations of their microscopic trajectories. We focus on systems whose complexity and sheer sizes render complete microscopic description impractical, and constructing theoretical macroscopic models requires extensive domain knowledge or trial-and-error. Our machine learning approach addresses this by simultaneously constructing reduced thermodynamic coordinates and interpreting the dynamics on these coordinates. We demonstrate our method by studying theoretically and validating experimentally, the stretching of long polymer chains in an externally applied field. Specifically, we learn three interpretable thermodynamic coordinates and build a dynamical landscape of polymer stretching, including (1) the identification of stable and transition states and (2) the control of the stretching rate. We further demonstrate the universality of our approach by applying it to an unrelated problem in a different domain: constructing macroscopic dynamics for spatial epidemics, showing that our method addresses wide scientific and technological applications.