第1011期 小路和你读AI论文
本期精选5篇高质量AI论文：1) AI智能体生态研究揭示协作产生强Allee效应，种群规模存在起飞阈值，对智能体安全有重要意义；2) BrickBench提出文本条件乐高组装的智能体基准，融合物理仿真验证；3) 思考惯性研究显示LLM被告知不思考时仍会输出显式推理，挑战一条指令关闭思维链的假设；4) 从形式化验证角度分析智能体自我改进的理论基础与极限；5) 提出工具型智能体的主动遗忘方法，可逆地整理上下文，累积输入token约减少50%，无需训练即可降低成本。
全部文档 / 标签: 推理控制
2 篇
Qwen3.8-27B Practical Guide: Control Reasoning Depth and Extend Context to 1M Tokens
Qwen3.8-27B is a compact, deployment-friendly dense model with native vision-language understanding and flexible thinking control. It introduces two key features: reasoning_effort to adjust reasoning depth per request, and ultra-long context support (262K native, up to 1M with YaRN). This guide provides practical instructions on using these features, including code examples for OpenAI-compatible API, vLLM, SGLang, and TokenSpeed, along with tips for optimal configuration.
