kaifaai.app AI 快讯 · 2026-10-10

联想TianxiCode代码智能体登顶SWE-bench-Live全球第一

联想天禧自研代码智能体TianxiCode以71%问题解决率获SWE-bench-Live全球第一，为专业代码智能体框架。

联想天禧AI自主研发的TianxiCode代码智能体，在SWE-bench-Live（当前最具现实意义的代码Agent评测基准）中以71%的问题解决率位列全球第一。该结果来自量子位报道，数据明确、无修饰词，代表其在真实软件工程场景（如GitHub issue修复、跨文件调试）中的实际交付能力。不同于纯学术榜单，SWE-bench-Live强调端到端闭环：从理解模糊需求、检索代码库、生成补丁，到通过CI测试并合并PR——这正是AI原生开发工具链落地的核心瓶颈。对AI应用创业者而言，此成绩揭示两个硬事实：第一，垂直领域智能体无需通用AGI即可建立壁垒——TianxiCode聚焦代码而非多模态，但解决率超多数通用Agent；第二，评测标准正在迁移：投资人和客户已不再满足于‘能写Hello World’，而要求‘能修生产环境Bug’。团队若仍在用HumanEval或MBPP做核心指标，将面临POC阶段即被客户质疑真实效能的风险。建议将SWE-bench-Live或类似生产级基准纳入研发流程，并公开关键子任务（如test-pass rate、merge success rate）数据。

Lenovo’s TianxiCode tops SWE-bench-Live with 71% solve rate

Lenovo Tianxi AI’s TianxiCode code agent achieved a 71% solve rate on SWE-bench-Live—the most realistic, production-oriented coding benchmark today—which evaluates full-cycle tasks like interpreting GitHub issues, modifying multi-file codebases, passing CI tests, and merging PRs. This result, reported by Qbit News, reflects tangible engineering delivery—not synthetic accuracy. For AI application startups, it signals a decisive shift in evaluation: investors and enterprise buyers now demand proof of closed-loop productivity, not just code generation fluency. Building a general-purpose agent is no longer the default path to defensibility; domain-specialized agents (like TianxiCode for software engineering) can dominate benchmarks and win contracts by excelling at high-stakes, narrow workflows. Teams still relying on HumanEval or MBPP for validation risk failing customer POCs where real repos, legacy dependencies, and test-suite failures expose gaps. Integrating SWE-bench-Live—or similar production-grade benchmarks—into development and sales cycles is now table stakes.

来源：量子位 — 联想天禧自研代码智能体TianxiCode斩获SWE-bench-Live全球第一