腾讯混元发布 Hyra 2.0:为模型研发打造可复用工具Tencent releases Hunyuan Hyra 2.0: reusable tools for model R&D


AI free token/AI免费积分

IT之家 11 月 2 日消息,今日腾讯混元发布了 AI 科研平台 Hyra 2.0,将研发重点从直接用 AI 攻克科学难题,转向让通用 Agent 参与科学模型、研究工作流及软件工具的研发与升级,让一次研发的成果在后续研究中被反复使用,推动科研范式从“刷题解题”走向“造工具”。

腾讯混元发布 Hyra 2.0:为模型研发打造可复用工具

此前,Hyra 1.0 曾驱动开源模型应对加法组合学中悬置五十余年的开放难题,验证了“给 AI 明确评判标准的难题并循环尝试”这一路径的可行性。Hyra 2.0 在此基础上进一步提出,工具升级带来的收益可被下游任务反复复用,从而形成智能复利。

腾讯混元发布 Hyra 2.0:为模型研发打造可复用工具

IT之家注意到,架构上,Hyra 2.0 将研究工作拆分为 Overall、Mission 和 Worker 三层,Overall 负责人与人类研究者沟通并对齐目标,Mission 长期负责单一研究难题并判断路线取舍,Worker 则独立读代码、做实验、分析失败原因,并可质疑 Mission 的判断。平台还外观了项目状态、Agent 分层记忆与共享知识库 Books 三套分离的存储机制,项目复盘时可将可迁移方法整理为 skill,供后续任务直接调用。

本文来源:https://www.ithome.com/1/021/266.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT之家 November 2 news, today Tencent Hunyuan released the AI scientific research platform Hyra 2.0, shifting its R&D focus from directly using AI to solve scientific problems to engaging general-purpose Agents in the research and upgrading of scientific models, research workflows, and software tools, enabling the results of a single R&D effort to be repeatedly used in subsequent research, driving the scientific research paradigm from "solving practice problems" to "building tools."

Tencent releases Hunyuan Hyra 2.0: reusable tools for model R&D

Previously, Hyra 1.0 drove open-source models to tackle a long-standing open problem in additive combinatorics that had remained unsolved for over fifty years, verifying the feasibility of the path of "providing AI with clear evaluation criteria for problems and attempting cyclically." Building on this, Hyra 2.0 further proposes that the benefits brought by tool upgrades can be repeatedly reused by downstream tasks, thereby forming an intelligent compound interest.

Tencent releases Hunyuan Hyra 2.0: reusable tools for model R&D

IT之家 noted that, architecturally, Hyra 2.0 divides research work into three layers: Overall, Mission, and Worker. Overall is responsible for communicating with human researchers and aligning goals; Mission is responsible for a single research problem over the long term and judging the trade-offs of routes; Worker independently reads code, conducts experiments, analyzes causes of failure, and can question Mission's judgments. The platform also features three separate storage mechanisms: project status, Agent tiered memory, and the shared knowledge base Books. During project reviews, transferable methods can be organized into skills for direct invocation by subsequent tasks.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry