Arash Ebrahim Abtahizadeh, Muna Aghamelu, Adil Ahmed, Juan C. Angeles Acuna, Alexis Audran-Reiss, Asaf Avrahamy et al.
arXiv 2610.24628 42 pages Read the paper
An end-to-end explanation of the model’s attention cache: separate histories for each layer, recent-window storage in local layers, full-history storage in global layers, and shared key-value heads.
How static allocations, Greedy-KL, and EvoPress distribute weight precision to fit an edge deployment budget, and how calibration and downstream evaluation check the resulting quality tradeoff.
Real users and tool histories are costly and sensitive to collect; this pipeline creates coherent simulated digital lives and filters their task trajectories for grounded, privacy-aware, and safe agent practice.
A compact agent must learn broad capabilities and recover from its own imperfect generation paths; Muse Glimmer uses teacher-provided token distributions first during pre-training and then on student-generated rollouts before reinforcement learning.