agentics

Epoch AI 发布了一项评测报告，用自身真实工作任务测试六个模型，发现 Fable 5.1 与 GPT-6 Astra 虽然领先，但在开放式任务、隐性标准遵循和研究判断力上仍无法实现工作自动化。

NavGPT-3 提出一个将大型语言模型推理与动作策略连接的层级式导航运行时，其 NavGPT-VLA 动作策略在 R2R-CE 和 RxR-CE 上达到 SOTA，并首次使自主智能体在导航成功度和路径保真度上匹配人类水平。

OpenAI announced the Decisions API in limited preview, powered by GPT-6 Luna, enabling apps to define questions and answer options for real-time content classification, request routing, and agent action selection.

GPT-6 Astra achieves state-of-the-art scores on the ARC-AGI-3 benchmark, scoring 99.9% with a Provider Adapter harness and demonstrating fewer actions than human testers. The model exhibits the ability to convert unfamiliar environments into compact symbolic world models for efficient planning.

The article argues that memorizing session transcripts provides no performance benefit for AI agents on SWE tasks, and may actually harm performance by wasting tokens, as valuable information is already captured in code artifacts and documentation.