KAT-Coder-V2.5 Technical Report

2026-07-06

AI Summary

This work addresses the limited autonomous agent capabilities of large code models in real, executable repositories—stemming from scarce reproducible environments, verifiable rewards, and high-quality execution trajectories—by proposing an end-to-end agent post-training framework. The approach leverages AutoBuilder to reconstruct multilingual repository sandboxes, generate task specifications, and produce near-successful trajectories, while KwaiClawEnv synthesizes large-scale tool-use data. Key innovations include process-aware filtering, hindsight-enhanced value estimation, test-suite-driven reward mechanisms, and multi-teacher online policy distillation. Moving beyond single-turn code generation, this method achieves state-of-the-art tool-use performance on PinchBench and demonstrates repository-level development capabilities rivaling Opus 4.8 across six software engineering and agent benchmarks.

Abstract

We present KAT-Coder-V2.5, a coding-focused agentic model trained to act autonomously inside real, executable repositories rather than as a single-turn code generator. Its capability is bottlenecked less by model scale than by the scarcity of reproducible environments, verifiable rewards, and high-value trajectories, which we address with an end-to-end agentic post-training framework. AutoBuilder reconstructs multilingual repositories into sandboxed environments with fail-to-pass and pass-to-pass verification at scale, from which we regenerate self-contained task specifications, recover near-miss trajectories, and distill supervision through process-aware filtering, while KwaiClawEnv synthesizes large-scale tool-use trajectories from executable services and real task seeds. We further scale reinforcement learning with harness randomization, a reliability-hardened sandbox, an asymmetric actor--critic PPO with hindsight-augmented value estimation, and a harness-oriented reward framework, and unify SWE, Agent-Claw, and WebCoding experts via Multi-Teacher On-Policy Distillation. Across six software-engineering and agentic benchmarks, KAT-Coder-V2.5 delivers the best agentic tool-use result on PinchBench and ranks second only to the frontier Opus 4.8 on repository-level software engineering. Our service is available at https://streamlake.com/product/kat-coder.

Problem

• agentic coding
• reproducible environments
• verifiable rewards
• high-value trajectories
• software engineering

Innovation

• agentic coding
• executable repository environments
• process-aware distillation
• tool-use trajectory synthesis
• multi-teacher on-policy distillation