Sebastian Raschka 讨论了大语言模型从预训练到后训练的转变、推理革命以及工具使用的重要性日益增加，回顾了一年来的变化和 2026 年初的发展。
Sebastian Raschka discusses the shift from pre-training to post-training, the reasoning revolution, and the growing importance of tool use in LLMs, reflecting on changes from a year ago and early 2026 developments.
要点 · TL;DR
基于可验证奖励的后训练推理是 LLM 的主要前沿。 Post-training reasoning with verifiable rewards is the main frontier for LLMs.
自一致性等推理缩放技术可在测试时提升性能。 Inference scaling techniques like self-consistency boost performance at test time.
智能体循环和工具使用使 LLM 适用于复杂多步任务。 Agentic loops and tool use make LLMs practical for complex multi-step tasks.
核心观点 · Key points
后训练推理是主要前沿，可验证奖励使数学和代码的可扩展强化学习成为可能。 Post-training reasoning is the main frontier, with verifiable rewards enabling scalable RL for math and code.
自洽性和自我改进等推理缩放技术可在测试时提升模型性能。 Inference scaling techniques like self-consistency and self-refinement boost model performance at test time.
智能体循环和工具使用正在成熟，使LLM更适用于复杂的多步骤任务。 Agentic loops and tool use are maturing, making LLMs more practical for complex, multi-step tasks.
LLM架构保持稳定；创新集中在效率上，如混合专家和稀疏注意力。 LLM architecture remains stable; innovations focus on efficiency like mixture-of-experts and sparse attention.
更长的上下文窗口减少了对持续学习的需求，但工具使用是获取最新信息的关键。 Longer context windows reduce need for continual learning, but tool use is key for up-to-date information.
反共识 · Contrarian takes
预训练的研发不再令人兴奋；大部分收益现在来自后训练和推理缩放。 Pre-training R&D is less exciting; most gains now come from post-training and inference scaling.
社交媒体上的一次编码声称往往具有误导性；实际使用需要多次迭代。 One-shot coding claims on social media are often misleading; real use requires multiple iterations.
多智能体系统被过度炒作；瓶颈是上下文工程，而非模型能力。 Multi-agent systems are overhyped; the bottleneck is context engineering, not model capability.
持续学习是一个遥远的梦想；实际更新是半自动的且受资源限制。 Continual learning is a distant dream; practical updates are semi-automatic and resource-constrained.
文本扩散模型可能取代Transformer以实现廉价、快速的推理，但无法达到顶级性能。 Text diffusion models may replace transformers for cheap, fast inference but not for top performance.
本期章节 · Chapters（共 25）
开场与欢迎Intro and Welcome
一年来的演变：推理与工具使用Evolution from a Year Ago: Reasoning and Tool Use
实用视角：2026年初发展Practical Perspective: Early 2026 Developments
LLM辅助编程的演变Evolution of LLM-assisted coding
渐进改进与推理努力Incremental improvements and reasoning effort
本地智能体如MultbotLocal agents like Multbot
Moldbot与LLM用例的兴奋Excitement around Moldbot and LLM use cases
LLM实际应用：开发与聊天界面Sports prediction side project with LLM
一次性声称与现实Practical LLM usage: development and chat interfaces
引言与创新关键领域One-shot claims vs reality
可验证奖励与推理训练Introduction and Key Areas of Innovation
奖励类型与过程奖励模型Verifiable Rewards and Reasoning Training
推理缩放与模型路由Reward types and process reward models
GPT OSS与推理努力Inference Scaling and Model Routing
自洽性与自我优化GPT OSS and Reasoning Effort
LLM的智能体用途Self-Consistency and Self-Refinement
智能体用例与多智能体系统Agentic Uses of LLMs
LLM架构演变Agentic Use Cases and Multi-Agent Systems
架构变化：效率与性能LLM Architecture Evolution
持续学习：梦想与现实Architecture changes: efficiency vs. performance
持续学习的挑战Continual learning: dream vs. reality
未来方向与预测Challenges of Continual Learning
个人工作与书籍更新Future Directions and Predictions
构建LLM的书籍与推理续篇Personal Work and Book Update
Book on Building LLMs and Sequel on ReasoningBook on Building LLMs and Sequel on Reasoning