Research
Reasoning & agents
How models went from continuing text to reasoning to acting: chain of thought, verification, tool use, benchmarks. Every design choice in a coding agent has a source here.
Articles
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。
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。
PickIn the canon
- S Chain-of-Thought Prompting Elicits Reasoning in Large Language Models · 2022 · Writing out intermediate steps in the examples takes PaLM 540B from 17.9% to 56.9% on GSM8K
- S ReAct: Synergizing Reasoning and Acting in Language Models · 2022 · The model alternates reasoning, acting and observing in one loop
- S DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning · 2025 · Reinforcement learning alone (GRPO with verifiable rewards) grows long chains of thought in a base model
- A On the Measure of Intelligence · 2019 · Argues intelligence is the efficiency of acquiring skills, not the skills themselves
- A Measuring Massive Multitask Language Understanding · 2020 · MMLU’s multiple-choice questions across 57 subjects became the common yardstick of model knowledge from 2020 to 2024
- A Evaluating Large Language Models Trained on Code · 2021 · The Codex paper introduced HumanEval and popularized pass@k for measuring coding ability
- A STaR: Bootstrapping Reasoning With Reasoning · 2022 · The model writes its own reasoning and fine-tunes on the chains that reach correct answers
- A Toolformer: Language Models Can Teach Themselves to Use Tools · 2023 · The model teaches itself to insert API calls (calculator, search, calendar, translation) into its text
- A Reflexion: Language Agents with Verbal Reinforcement Learning · 2023 · Agents keep written self-reflections in memory and improve across attempts without weight updates
- A Generative Agents: Interactive Simulacra of Human Behavior · 2023 · 25 agents with memory, reflection and planning behave believably in a simulated town
- A Let's Verify Step by Step · 2023 · Supervising each reasoning step is more reliable than supervising only the outcome
- A Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena · 2023 · Tests strong models as judges and documents biases such as position and verbosity
- A SWE-bench: Can Language Models Resolve Real-World GitHub Issues? · 2023 · Tests whether models can resolve 2,294 real GitHub issues end to end
- A SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering · 2024 · File, search and edit commands designed for the model work far better than a raw shell
- A Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters · 2024 · Spending more compute at inference often beats making the model larger