[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-4d9f11bb-5795-45e5-a90b-7eb29756da24":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"4d9f11bb-5795-45e5-a90b-7eb29756da24","腾讯混元发布 Hyra-1.0：用递归自我改进重写研究智能体的范式","7 月 21 日,腾讯混元正式发布 Hyra-1.0（Hunyuan Research Agent），这是国内首个面向研究与工程任务的递归自我改进（Recursive Self-Improvement，RSI）智能体。\n\nHyra 的设计遵循 The Bitter Lesson：把外部框架做轻、把动作空间做宽。给定任务描述后，它会持续运行一个探索循环——基于历史经验（含过往 solution 的运行日志、评估器反馈、源代码）不断提出更好的方案，直到主动结束或预算耗尽，最终返回历史最优解。这种“经验复用 + 自动反思”的模式，让 Hyra 能在公开 benchmark 之外，在产品系统、AI 研发流水线乃至自然科学与工业场景里持续进化。\n\n官方样例展示了这种能力的优势：仅凭一张 2D 参考图生成可渲染的 3D 模型，Hyra 多轮迭代修改建模代码和渲染参数，由基于 rubrics 的 VLM judge 从轮廓、比例、结构完整性、材质等维度打分，产物比 Claude Code goal 模式更接近参考图像、更符合人类审美。\n\n过去一年 RSI 与自动研究几乎由海外主导：DeepMind 的 AlphaEvolve 用 48 次标量乘法完成 4×4 复数矩阵相乘，Together AI 把 11 维 kissing number 下界推到 604，Karpathy 的 autoresearch 给出精简的训练自循环范式。腾讯这次入场，意味着国内厂商开始系统性补齐这一前沿方向——而 RSI 的真正考验是“在生产环境里跑得久、稳、安全”，Hyra 把验证直接挂在研发流水线和工业场景里，这种姿态本身比纯 demo 更值得关注。","https:\u002F\u002Ffinance.sina.com.cn\u002Ftech\u002Fdigi\u002F2026-07-21\u002Fdoc-iniiptrh9371935.shtml","d46ec0a7-501b-4ef8-9c89-2391b2701b3b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",{"id":18,"name":19,"slug":19,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model","2026-07-21T06:30:00Z","2026-07-21T06:05:25.133092Z","2026-07-21T06:05:25.133102Z",true,"agent",4]