Reproduce “KAT-Coder-V2 Technical Report” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Fengxiang Li, Han Zhang, Haoyang Huang, Jinghui Wang, Jinhua Hao, Kun Yuan, Mengtong Li, Minglei Zhang and 38 more

The KwaiKAT Team at Kuaishou developed KAT-Coder-V2, an advanced coding agent that employs a "Specialize-then-Unify" post-training paradigm to excel in diverse agentic coding tasks. This model achieved 79.6% on SWE-bench Verified and demonstrated leading performance in frontend aesthetics generation, supported by a modular infrastructure and algorithmic innovations that yielded up to 6.2x training speedup.

Key insights

  • The "Specialize-then-Unify" paradigm effectively mitigates capability fragmentation, enabling a single model to achieve expert-level performance across diverse agentic coding domains without significant cross-domain interference.
  • A modular and decoupled infrastructure like KwaiEnv is crucial for scaling agentic RL research and development, allowing independent iteration of components and seamless integration of new benchmarks and agent scaffolds.
  • Algorithmic advancements such as MCLA for stabilizing MoE RL and Tree Training for efficiently handling tree-structured trajectories are essential for stable convergence and significant speedups in large-scale agentic RL training.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.