Presents a unified black-box reinforcement learning framework to train and optimize agents running inside complex execution harnesses. Uses sandbox-parallel rollouts, a serving proxy that captures model calls and reconstructs multi-turn trajectories as prefix trees, and adapted GRPO/PPO optimizers to achieve stable, scalable RL across heterogeneous harnesses.
Enables interactive serving of large Mixture-of-Experts (MoE) models on personal machines by adapting offload and execution to measured device bandwidth and agentic workload patterns. Key features include bandwidth-adaptive execution, semantic-aware caching of recurrent state, and an elastic GPU expert cache; supports 20+ MoE models and runs models from ~35B to 753B on consumer/workstation GPUs.
Fine-tunes long-horizon LLM agents with evolution strategies so full-model updates run at inference-level GPU memory. Emphasizes trajectory-level credit via black-box rewards, online prompt–parameter co-evolution, and a cosine decay for perturbation scale to balance exploration and adaptation; suited for limited-GPU settings.
AI Model2026
A 35B mixture-of-experts LLM tuned for agentic coding and end-to-end self-improvement: it jointly generates tasks, scaffolds, and solution rollouts. Activates ~3B params/token, supports 256K context (extendable), and emits chain-of-thought plus OpenAI-style tool calls.