Video-Coder addresses the challenge of precise spatial structure, motion dynamics, and controllable scene composition in video generation by introducing a code-driven agent. The agent autonomously writes and executes code to construct a dynamic 3D scene (Draft Video) that guides downstream video generators. To enable this paradigm, the authors designed a dedicated data pipeline, curated two datasets (Video-Coder-SFT-6k and Video-Coder-RL-2k), and developed the Video-Coder Harness, an interactive environment for code execution, 3D rendering, and feedback. The model is trained using supervised fine-tuning followed by harness-native reinforcement learning with dual reward feedback and an annealing strategy that balances execution and video rewards. Experiments indicate significant performance improvements on benchmarks such as PhyGenBench and VBench-2.0. The code, data, and models are fully open-sourced, though the paper is currently a preprint.
Weight-Space Programming (WSP) allows reasoning agents to directly inspect and modify neural network parameters instead of relying on standard gradient-based adaptation. This preprint systematically investigates WSP using 16 editing trajectories across seven benchmarks and two model scales (Qwen3.5-0.8B and Qwen3.5-9B). The findings reveal three main insights: first, WSP can achieve substantial performance gains via highly localized interventions—such as improving accuracy by 29.45 percentage points on MATH-500 by modifying only 1,023 parameters. Second, WSP exhibits promising yet uneven cross-task generalization, where a math-edited checkpoint transfers well to some benchmarks (e.g., SciBench and IFEval) but leaves others unchanged. Third, effective editing remains challenging and inconsistent, with substantial variation across trajectories and several attempts retaining original weights due to unreliable validation. The paper provides open-source code and highlights the fundamental challenges of intervention discovery, reliability, and cross-task transfer in parameter-level editing.
