快手KwaiKAT团队发布KAT-Coder-V2.5：基于10万+可验证仓库环境训练的智能体编程模型

快手的 KwaiKAT 团队推出了 KAT-Coder-V2.5：https://arxiv.org/abs/2607.05471。它是一个经过训练，可以在真实可执行仓库中操作的编码模型，而不是生成单轮代码。可通过 StreamLake 使用该服务模型：https://streamlake.com/product/kat-coder。一个开源权重版本 KAT-Coder-V2.5-Dev：https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev，已在 Hugging Face 上以 Apache-2.0 许可单独发布。

该研究将可验证任务定义为三元组。它需要一个精确的任务描述、一个可执行仓库环境以及一组验证测试。补丁只有在通过所有这些测试时才被认为是正确的。

任务从真实的拉取请求和提交中挖掘，遵循 SWE-bench 系列：https://arxiv.org/abs/2509.16941。合并的代码更改提供黄金补丁，随附的测试更改提供测试补丁。原始问题文本被舍弃，不作为规范。相反，描述被重新生成为三个部分：基于黄金补丁的问题陈述、由测试补丁派生的需求，以及从两者推断的接口约束。随后进行清晰度检查，删除任何模糊、不完整、未充分说明或内部不一致的内容。

AutoBuilder 负责环境端。构建代理分析仓库并编写配置脚本，从干净的检出开始安装依赖并运行测试。验证代理在隔离的沙箱环境中执行该脚本。

验收规则是关键部分。验证不读取退出代码或 grep 日志模式。它解析结构化测试框架输出，仅当收集到超过 90% 的预期测试并且运行结果在不同执行中可复现时，才接受环境。失败结果会作为结构化信息反馈，用于迭代修复。

通过结合预配置的基础环境、构建系统模板以及可检索的精炼构建方案库，构建成功率从 16.5% 提升到 57.2%。最终得到超过 100,000 个可验证环境，覆盖 12 种语言。Git 历史、提交元数据及其他可利用痕迹被剥离，以确保代理无法从仓库中读取参考解决方案。

仅根据最终测试成功来筛选轨迹具有误导性。一些通过测试的运行依赖硬编码、机制绕过或面向测试的捷径。一些失败的运行包含有价值的搜索、本地化和修复行为。

KwaiKAT 处理这两个方向。对于接近通过的情况，有针对性的过程级提示指示需要检查或验证的内容，而不泄露解决方案。仅这一点就能将以前零通过任务的通过率提升到约 20%。由于提示轨迹包含推理时不可用的信息，经验证的补丁随后被修复，并从原始任务上下文重新生成无提示轨迹。仅保留通过验证、不泄露提示且与补丁一致的样本。

对于通过的运行，基于规则的门控机制会移除无效、不稳定或剥削性的轨迹。评分阶段随后对探索、本地化、预编辑推理、规范遵守、仓库惯例、补丁最小性、验证质量、恢复行为和诚信进行评分。

第三种机制针对测试框架过拟合。工具名称、参数约定、输出格式和提示模板在功能保持不变的情况下被随机化。由于验证是基于测试结果而非测试框架痕迹，一个任务可以在多种测试框架配置下重新提供。还注入了现实的扰动：缺失依赖、瞬时命令失败、输出截断和日志噪声。

在 KAT-Coder-V2：https://arxiv.org/abs/2603.27703 训练期间，奖励曲线缓慢最初被归咎于RL算法。然而，审计发现约16%的轨迹失败是由于沙箱基础设施问题，而非模型策略，边界错位有时导致约40步的观测为空并破坏奖励。

随后进行了三项基础设施修复。首先，早期释放的镜像驱逐策略将磁盘使用率从95%降低到60%，将因超时导致的无效上线从6–7%降低到不到1%。其次，在远程沙箱初始化期间纠正环境变量，停止了会翻转6–7%样本奖励的系统覆盖，错误率降至1%以下。第三，网关服务器绕过了主流聊天端点（通过重新应用 apply_chat_template 并重新分词，导致约200轮交互时标记漂移达40%），直接调用 /generate 以确保上线标记对齐。

这些更新共同将沙箱反馈错误率从约16%降至不到2%，并将训练崩溃次数减少了一个数量级。

研究团队选择使用 PPO：https://arxiv.org/abs/1707.06347 并结合 GAE，而非无评论者轨迹方法，因为生产环境将会话拆分为结构上不同的样本，这使得分组基线变得复杂。

使用非对称 Actor–Critic：https://arxiv.org/abs/1710.06542 时，Critic 获取特权训练上下文（奖励、测试、覆盖、补丁、元数据、未来轮次），而 Actor 仅看到 rollout 状态。推理时，Critic 和额外上下文会被丢弃。

奖励分为三层：核心任务评分要求所有 fail_to_pass 和 pass_to_pass 测试通过；标准行为约束对重复、错误工具调用和调试残留进行惩罚；失败轨迹激励通过 F2 对文件检索打分，并给予部分测试积分。

五位专家通过使用逆 KL 的多教师在策略蒸馏（Multi-Teacher On-Policy Distillation）、离策略起点以及 Prune-OPD：https://arxiv.org/abs/2605.07804 的漂移感知截断来融合。

在统一的 Claude Code 训练环境下，KAT-Coder-V2.5 在 PinchBench：https://pinchbench.com/ 上以94.9领先其面板，超过 Opus 4.8 的93.5。在 SWE-Bench Pro 上排名第二（65.2 对 69.2），在内部 KAT Code Bench 上也排名第二（53.1 对 57.3）。

然而，它在 Terminal-Bench 2.1 上表现落后，最终以60.7 排名末位，低于 GLM-5.1（61.8）和 Opus 4.8（84.6）。在 SciCode 上得分为50.3，与 GLM-5.2 持平。

值得注意的是，开源权重 KAT-Coder-V2.5-Dev：https://huggingface.co/collections/Kwaipilot/kat 是一个独立的 35B 总量 / 3B 活跃 MoE 模型，在 Qwen3.6-35B-A3B 上使用 127K SFT 示例进行后训练，然后进行 RL。在单独的内部评估协议上评估，其结果无法与主要旗舰表进行比较。

查看论文：https://arxiv.org/abs/2607.05471，Hugging Face 上的模型权重：https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev，以及产品页面：https://streamlake.com/product/kat。所有研究成果的功劳归项目的研究人员所有。

Michal Sutter 是一名数据科学专业人士，拥有帕多瓦大学的数据科学硕士学位。凭借统计分析、机器学习和数据工程的坚实基础，Michal 擅长将复杂的数据集转化为可操作的洞察。

构建一个自主代理活动场地运营者 [完整代码]：https://pxllnk.co/twdn5

谢谢！我们的团队将很快与您联系 🙌

The KwaiKAT Team at Kuaishou has introduced the KAT-Coder-V2.5：https://arxiv.org/abs/2607.05471. It is a coding model trained to operate inside real, executable repositories rather than emit single-turn code. The served model is available through StreamLake：https://streamlake.com/product/kat-coder. An open-weight variant, KAT-Coder-V2.5-Dev：https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev, was released separately on Hugging Face under Apache-2.0.

The research frames a verifiable task as a triplet. It needs a precise task description, an executable repository environment, and a set of validation tests. A patch is correct only if it passes all of them.

Tasks are mined from real pull requests and commits, following the SWE-bench：https://arxiv.org/abs/2509.16941 lineage. The merged code change supplies a golden patch and the accompanying test change supplies a test patch. Raw issue text is discarded as a specification. Instead, descriptions are regenerated into three parts: a problem statement grounded in the golden patch, requirements derived from the test patch, and interface constraints inferred from both. A clarity check then drops anything ambiguous, incomplete, underspecified, or internally inconsistent.

AutoBuilder handles the environment side. A build agent analyzes the repository and writes a configuration script that installs dependencies and runs tests from a clean checkout. A verification agent executes that script in an isolated sandbox.

The acceptance rule is the interesting part. Verification does not read exit codes or grep log patterns. It parses structured test-framework output, and accepts an environment only when more than 90% of expected tests are collected and pass/fail outcomes reproduce across runs. Failures are fed back as structured information for iterative repair.

Combining a preconfigured base environment, build-system templates, and a retrievable library of distilled build recipes raised the construction success rate from 16.5% to 57.2%. The result is over 100,000 verifiable environments spanning 12 languages. Git history, commit metadata, and other exploitable traces are stripped so agents cannot read the reference solution out of the repo.

Filtering trajectories by final test success is misleading. Some passing runs rely on hard-coding, mechanism bypassing, or test-oriented shortcuts. Some failing runs contain valuable search, localization, and repair behavior.

KwaiKAT addresses both directions. For near misses, targeted process-level hints indicate what to inspect or verify without revealing the solution. That alone raises the pass rate of previously zero-pass tasks to roughly 20%. Because hinted trajectories contain information unavailable at inference, the verified patch is then fixed and a hint-free trajectory is regenerated from the original task context. Only samples that pass verification, show no hint leakage, and stay consistent with the patch are retained.

For passing runs, rule-based gates remove invalid, unstable, or exploitative trajectories. A scoring stage then rates exploration, localization, pre-edit reasoning, specification fidelity, repository conventions, patch minimality, verification quality, recovery behavior, and honesty.

A third mechanism targets harness overfitting. Tool names, argument conventions, output formats, and prompt templates are randomized while functionality is preserved. Because verification is anchored to test outcomes rather than harness traces, one task can be re-served under many harness configurations. Realistic perturbations are injected too: missing dependencies, transient command failures, truncated outputs, and noisy logs.

During KAT-Coder-V2：https://arxiv.org/abs/2603.27703 training, slow reward curves were initially blamed on the RL algorithm. However, an audit revealed that ~16% of trajectories failed due to sandbox infrastructure issues rather than the model policy, with boundary misalignments sometimes emptying observations for ~40 steps and corrupting rewards.

Three infrastructure fixes followed. First, an early-release image eviction policy lowered disk usage from 95% to 60%, reducing timeout-induced invalid rollouts from 6–7% to under 1%. Second, correcting environment variables during remote sandbox initialization stopped system overrides that flipped rewards on 6–7% of samples, cutting errors below 1%. Third, the Gateway Server bypassed mainstream chat endpoints—which caused 40% token drift at a ~200-turn scale by re-applying apply_chat_template and re-tokenizing—and called /generate directly to ensure rollout token alignment.

Together, these updates reduced the sandbox feedback error rate from roughly 16% to below 2% and cut training collapses by an order of magnitude.

The research team chose PPO：https://arxiv.org/abs/1707.06347 with GAE over critic-free trajectory methods because production harnesses split sessions into structurally distinct samples, complicating group baselines.

Using asymmetric actor–critic：https://arxiv.org/abs/1710.06542, the Critic gets privileged training context (rewards, tests, coverage, patches, metadata, future turns), while the Actor sees only rollout state. The Critic and extra context are discarded at inference.

Rewards are three-tiered: Core Task Scores require all fail_to_pass and pass_to_pass tests to pass; Standard Behavior Constraints penalize duplication, bad tool calls, and debug remnants; Failed Trajectory Incentives score file retrieval via F2 and give partial test credit.

Five experts fuse via Multi-Teacher On-Policy Distillation using reverse KL, an off-policy start, and drift-aware truncation from Prune-OPD：https://arxiv.org/abs/2605.07804.

Under a unified Claude Code harness, KAT-Coder-V2.5 leads its panel on PinchBench：https://pinchbench.com/ with 94.9, beating Opus 4.8 at 93.5. It places second on SWE-Bench Pro (65.2 vs 69.2) and the internal KAT Code Bench (53.1 vs 57.3).

However, it lags on Terminal-Bench 2.1, placing last with 60.7 behind GLM-5.1 (61.8) and Opus 4.8 (84.6). On SciCode, it scores 50.3, matching GLM-5.2.

Notably, the open-weight KAT-Coder-V2.5-Dev：https://huggingface.co/collections/Kwaipilot/kat is a separate 35B-total / 3B-active MoE post-trained on Qwen3.6-35B-A3B using 127K SFT examples, then RL. Evaluated on a separate in-house protocol, its results are not comparable to the main flagship table.

Check out the Paper：https://arxiv.org/abs/2607.05471, the Model Weight on Hugging Face：https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev, and the Product Page：https://streamlake.com/product/kat-coder. All credit for this research goes to the researchers of this project.

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.

Build an Agentic Event Venue Operator [Full Codes]：https://pxllnk.co/twdn5

Thanks! Our team will contact you soon 🙌