research How Wrong Can a Good Predictor Be? Diverging updates with vanishing predictive KL for every fixed finite hidden-state count. GradES Gradient-based early stopping for efficient transformer fine-tuning. Achieves 1.70–1.77× faster full-parameter fine-tuning at neutral accuracy on Qwen3-0.6B, Llama-3.1-8B, and Qwen3-14B. OceanEnv A unified marine-robotics simulation environment integrating bathymetry, ocean physics, biodiversity, and pollution data for 3-DOF/6-DOF sim-to-real training. tools LLM Orchestra Multi-agent LLM system (v6.x) with Claude as orchestrator, five-mode operation, and structured state management. Built Paper Forge academic writing system on top.
