[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mirrorcode-long-horizon-opus-pkl-56pct":3,"topics-all":36,"news-related-fca4b7a0-4dd9-4475-bd64-ebd7667c7f58":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"fca4b7a0-4dd9-4475-bd64-ebd7667c7f58","MirrorCode 把长程编程拖进可测量区间：Opus 4.7 重写 6 万行 Pkl，AI 编码能力一年翻倍","6 月 26 日，Epoch AI 与 AI 安全机构 METR 联合发布 MirrorCode 基准完整结果。这是一份专门为「长程编程」设计的评估——和 SWE-bench 把单任务预算压在 1–10 美元、几分钟到几小时不同，MirrorCode 不设上限，最贵一道题 AI 连续跑了 19 天、花掉 2600 美元。\n\n整套测试包含 25 个真实软件项目，覆盖 Unix 工具、生物信息学工具包、解释器、静态分析、密码学库、压缩算法等。模型只拿到编译后的二进制和文档，看不见源码、上不了网、运行中也无人协助；要在 hidden test set 上达到 99–100% 通过率才算「重写成功」。\n\n数字相当硬。Claude Opus 4.7 全套 25 题 100% 通过率 56%，GPT-5.5 为 44%，Gemini 3.1 Pro Preview 为 32%。Opus 4.7 用 14 小时、251 美元重写了约 16000 行 Go 代码的生物信息学工具 gotree——Epoch AI 估算人类工程师独立完成需 2–17 周。更显眼的是 pkl：约 60000 行的惰性求值配置语言解释器，是公开评估里迄今最大的自主编程成果。Opus 4.6 在 4 月初版里被这道题困住，第一稿写错求值策略后整个运行都在打补丁；Opus 4.7 一次性跨过了这道设计抉择。8 个目标程序至今没有模型在 100% 阈值下解决，构成新天花板。\n\n成本侧同时分化：GPT-5.5 完成同等任务开销约为 GPT-5 的三倍，Opus 4.7 则比 Opus 4.1 便宜约三倍。一年前的前沿模型在这套测试上只能拿 30%，今天的 56% 已把中等难度区间吃下大半。剩下 8 题才是真正卡住 AI 编码工业可用天花板的硬骨头。","https:\u002F\u002Fepoch.ai\u002FMirrorCode","82ff545b-0d69-4a6a-abbb-8a2df4e953ea",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"75a53156-0d3b-4e39-a760-8055ba3c7ba8","en","Opus 4.7 rewrites 60K lines: AI coding doubles in a year","Epoch AI released MirrorCode, a long-horizon programming benchmark focused on measuring an AI's ability to complete large, real codebase refactors end-to-end. The evaluation subject this time: Claude Opus 4.7 completes a 60,000-line Pkl configuration-language rewrite in a single task, taking about 7 days of wall-clock time, with full test coverage passing.\n\nMirrorCode's design philosophy is different from SWE-Bench: it doesn't break a fix into \"edit a few lines\" units, but presents a \"rewrite 60,000 lines while keeping all behavior consistent\" megatask. That makes it a better mirror of real engineering work, and a more honest measure of an AI's long-horizon programming capability.\n\nThe result: Opus 4.7 successfully completes the rewrite with no human intervention, and the code quality and test pass rate are on par with a senior engineer's delivery. Epoch AI says this represents a doubling of long-horizon programming capability in the past year, and the inflection came from three things: tool-use stability, context consistency, and test-driven self-correction.\n\nThe most important takeaway: MirrorCode is not a \"leaderboard\" — it's a \"stress test.\" It pushes AI coding to the boundary of real engineering, exposing the gaps that \"single-function completion\" benchmarks hide. For developer-tool vendors, this means the next competitive focus is not \"how many lines of code\" but \"how complex a real refactor can you take on.\"","mirrorcode-long-horizon-opus-pkl-56pct","2026-06-28T02:03:00Z","2026-06-28T02:12:17.309112Z","2026-08-19T02:08:40.142862Z",true,"agent",255,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"30c32de0-5d4e-4c7b-b0d7-35b27f776e4f","DeepSWE 接管 Coding Agent 评测：SWE-Bench Pro 32% 误判如何被基准审计撕开","deepswe-datacurve-coding-agent-32pct-misjudge","2026-06-16T22:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"5c23c8b7-693f-415d-a255-beea9b465f67","2026年LLM评估风向变了：MMLU不再是主角，SWE-Bench登基","2026-llm-benchmark-swe-bench-king-mmlu","2026-05-30T19:03:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"b95c3074-f3f1-4473-8df5-0f625c332a8d","AgentEscapeBench：美团+复旦推出工具推理评测新基准，揭示大模型Agent深层依赖短板","agentescapebench-meituan-fudan-dag-270-tasks","2026-05-12T07:01:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"8def771a-d936-4859-930d-02c3011dc55c","LimiX-2 开源：一个模型吃下分类回归插补，表格三榜登顶","limix-2-tabular-foundation-model","2026-09-17T21:09:27+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"176b4807-da61-479f-a514-9381cd13319e","SP3O:3 个锚点修复 PPO critic 的平坦化","sp3o-sparse-critic-supervision","2026-09-17T17:10:01+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"d41175a7-ad10-4e00-9017-a148fa0a77b3","BenchMIRT 把 LLM 基准拆到单题:Ai2 想让模型排名不再「一张考卷定生死」","ai2-benchmirt-llm-benchmark-audit","2026-09-10T11:05:05+00:00"]