The KwaiKAT team identified three core bottlenecks that limit the usefulness of large coding models: unreliable executable environments, noisy trajectory data that rewards shortcuts rather than genuine reasoning, and unstable reinforcement‑learning signals caused by sandbox infrastructure. To solve the first problem they introduced AutoBuilder, a two‑agent system that first analyzes a repository to generate a dependency‑install and test‑run script, then executes that script in an isolated sandbox. Acceptance is based on parsing structured test output and requiring more than 90 % of expected tests to be collected and reproducible across runs, which lifted environment construction success from 16.5 % to 57.2 % and produced over 100 000 verifiable environments covering twelve languages.
For the second bottleneck they built a Data Scaling Flywheel. Near‑miss trajectories receive process‑level hints that guide inspection without revealing the solution; after fixing the verified patch a hint‑free trajectory is regenerated and kept only if it passes verification, shows no hint leakage and stays consistent. Passing runs are filtered by rule‑based gates that strip exploitative or unstable traces, then scored on exploration, localization, pre‑edit reasoning, specification fidelity, repository conventions, patch minimality, verification quality, recovery behavior and honesty. Harness overfitting is reduced by randomizing tool names, argument conventions, output formats and prompt templates while preserving functionality, and by injecting realistic perturbations such as missing dependencies, transient command failures, truncated outputs and noisy logs.
Infrastructure audits revealed that roughly 16 % of RL trajectories failed due to sandbox boundary misalignments, not model policy. Fixes—early‑release image eviction, correct environment‑variable handling, and a Gateway Server that bypasses chat endpoints to avoid token drift—cut the sandbox error rate below 2 % and reduced training collapses by an order of magnitude.
These improvements let KAT‑Coder‑V2.5 achieve strong results on PinchBench (94.9), SWE‑Bench Pro (65.2) and internal KAT Code Bench (53.1), while the open‑weight KAT‑Coder‑V2.5‑Dev provides comparable gains under an Apache‑2.0 license.
#AI #ML #SoftwareEngineering #DevOps #LLM #AgenticAI