Research
Understand the frontier. Then use it.
The ideas behind today’s models, explained properly: deep reads of the papers that still matter, curated picks, and the frontier developments worth knowing.
Foundations
- 注意力就是全部:Transformer 如何改写了序列建模
- 规模律:为什么「把模型做大」从信仰变成了工程
- Language Models are Few-Shot Learners · At 175B parameters, GPT-3 learns new tasks from a few examples in the prompt, with no weight updates
Reasoning & agents
- 思维链提示:让模型「先想再答」为什么有效
- ReAct:推理与行动交替,Agent 的最小原型
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning · Reinforcement learning alone (GRPO with verifiable rewards) grows long chains of thought in a base model
Alignment & safety
- InstructGPT:用人类反馈把模型调教成助手
- Concrete Problems in AI Safety · Recast AI safety as five testable problems: side effects, reward hacking, scalable oversight, safe exploration and distributional shift
- Constitutional AI: Harmlessness from AI Feedback · The model critiques and revises its own outputs against written principles (a “constitution”)
Interpretability
- Toy Models of Superposition · Small, controlled models show networks storing far more features than neurons through superposition
- Locating and Editing Factual Associations in GPT · ROME uses causal tracing to locate stored facts in GPT’s mid-layer MLPs
- In-context Learning and Induction Heads · Evidence that a specific circuit, induction heads, drives much of in-context learning
Latest
注意力就是全部:Transformer 如何改写了序列建模SChinese
它把「按顺序读」换成「同时看全句、按相关性取信息」,让训练可以完全并行;此后所有大模型都长在这副骨架上。
Deep read思维链提示:让模型「先想再答」为什么有效SChinese
在示例里写出中间推理步骤,大模型的算术与常识推理成绩就跃升;它确立了「让模型先想再答」的范式,也是推理模型时代的概念起点。
PickInstructGPT:用人类反馈把模型调教成助手SChinese
监督微调、奖励模型、强化学习三步,让 13 亿参数的模型在人类偏好上胜过 1750 亿的 GPT-3;这是 ChatGPT 的直接技术蓝本。
PickReAct:推理与行动交替,Agent 的最小原型SChinese
让模型在「想一步、做一步、看结果」之间循环,把思维链和工具调用合为一个回路;今天几乎所有 Agent 框架的循环都从这里来。
Pick规模律:为什么「把模型做大」从信仰变成了工程SChinese
语言模型的损失随参数量、数据量与算力按平滑的幂律下降,三者之间有可预测的最优配比;它让扩大规模成为可以提前算账的决策。
Pick