KwaiKAT Releases KAT-Coder-V2.5: An Agentic Coding Model Trained on 100K+ Environments
The KwaiKAT team has unveiled KAT-Coder-V2.5, an agentic coding model backed by over 100,000 verifiable repository environments and advanced training architectures.
Stock photo for illustration only, not from the actual event
- KAT-Coder-V2.5 is an agentic coding model developed by the KwaiKAT research team.
- Trained on over 100,000 verifiable environments spanning 12 programming languages.
- Infrastructure updates reduced sandbox feedback error rates from 16% to under 2%.
- Achieves top performance on PinchBench with a score of 94.9, outperforming competitors.
The KwaiKAT research team has released its latest agentic coding model, KAT-Coder-V2.5. The research frames a verifiable task as a triplet consisting of a precise task description, an executable repository environment, and a set of validation tests, where a patch is correct only if it passes all validation criteria. Tasks are mined from real pull requests and commits following the SWE-bench lineage, with merged code changes supplying golden patches and test changes supplying test patches. Raw issue text is discarded in favor of regenerated problem statements, requirements, and interface constraints, followed by strict clarity checks to eliminate ambiguity.
For environment management, AutoBuilder analyzes repositories and writes configuration scripts to install dependencies and run tests from clean checkouts, while a verification agent executes scripts in isolated sandboxes. The acceptance rule parses structured test-framework outputs, accepting environments only when more than 90% of expected tests are collected and pass/fail outcomes reproduce consistently across multiple runs, feeding failures back as structured information for iterative repair.
Combining preconfigured base environments, build-system templates, and a retrievable library of build recipes raised construction success rates from 16.5% to 57.2%, yielding over 100,000 verifiable environments across 12 languages while stripping Git history and commit metadata to prevent leakage. KwaiKAT addresses trajectories by providing process-level hints for near misses, applying rule-based gates to filter out invalid or exploitative runs, and randomizing tool names, argument conventions, and prompt templates to prevent harness overfitting.
Training advanced coding agents with Reinforcement Learning often encounters hidden infrastructure bottlenecks, such as sandbox timeouts or memory exhaustion. KwaiKAT's systematic engineering fixes highlight how robust low-level sandbox management directly influences the stability, speed, and convergence of reinforcement learning algorithms in large AI models.
During training, three crucial infrastructure fixes—an early-release image eviction policy, correcting remote sandbox initialization environment variables, and bypassing mainstream chat endpoints via Gateway Server to call /generate directly—reduced sandbox feedback error rates from roughly 16% to below 2% and drastically cut training collapses. The team selected PPO with GAE paired with an asymmetric actor-critic framework and a three-tiered reward system, further enhanced by Multi-Teacher On-Policy Distillation.
Stock photo for illustration only, not from the actual event
Under a unified Claude Code harness, KAT-Coder-V2.5 leads its panel on PinchBench with a score of 94.9, beating Opus 4.8 at 93.5, and places second on SWE-Bench Pro. Additionally, the open-weight KAT-Coder-V2.5-Dev variant features a 35B-total / 3B-active Mixture of Experts architecture post-trained on Qwen3.6-35B-A3B using 127K SFT examples followed by reinforcement learning.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment