Tag
本文提出一种逐步分解的评测方法，将语言模型在400步长程任务中的失败分为重复识别、局部处理与计数器追踪三类，发现重复识别错误随步数上升，而无关上下文注入会导致计数器追踪在所有种子中失败，外部精确提醒可恢复其对齐。
O'Reilly Radar 报道如何为生产环境构建可靠的 AI 智能体系统，核心观点是提示词调优和更强的模型只能改进响应、无法保证动作安全，必须依赖可强制执行的策略层、最小权限和可恢复状态。
A developer shares MnemoBrain, a memory system that enables AI agents to automatically remember context using hooks and semantic search, removing the need for explicit memory commands.
Gemini provides a systems-engineering analysis of Dex's architecture, highlighting its stateless transformer core augmented by a state-management system to solve LLM amnesia, and distinguishes engineering achievements from ontological claims.
This paper investigates whether context-injected fine-tuning improves small language models' ability to use retrieved law for legal question answering in Bangladesh. Using Qwen3.5 at 0.8B, 2B, and 4B scales, they find fine-tuning helps at smaller scales but not at 4B, and reduces language drift.
The author points out that the core essence of Agent/Harness is 'context injection', summarizing the key point of the technology in one sentence.
The article shares key insights from a workshop by Boris on using CLAUDE.md for context injection in Claude, highlighting three usage levels, specific commands like /loop, and plan mode to improve developer workflows.