Papers

繁體中文

This folder connects the hands-on Agent Engineering Roadmap with research papers from major labs, universities, and open research communities.

The goal is not to collect every paper. The goal is to help learners understand the research lineage behind tools, memory, workflow, multi-agent systems, evaluation, safety, and production agents.

PDF Policy

This repository links to official paper pages and PDF URLs, but does not commit PDF files.

Reasons:

• paper licenses differ across arXiv, ACM, ACL, OpenReview, company sites, and conference proceedings
• external official links are easier to keep current
• original course notes are more useful than copied PDFs

When adding a paper, include the title, year, organization, official link, PDF link when available, related modules, and an original engineering summary.

Start Here

File	Purpose
Paper Reading Roadmap	Chapter-by-chapter paper map
Paper Notes	Short engineering notes for key papers
Paper Reading Roadmap zh	繁體中文論文閱讀路線
Paper Notes zh	繁體中文重點論文導讀

Recommended Papers

Topic	Paper	Year	Organization	Why it matters
Browser agent	WebGPT: Browser-assisted question-answering with human feedback	2021	OpenAI	Early reference for browsing agents, citation collection, and human feedback.
Reasoning and acting	ReAct: Synergizing Reasoning and Acting in Language Models	2022	Google Research / Princeton	Foundation for interleaving reasoning traces and actions.
Safety training	Constitutional AI: Harmlessness from AI Feedback	2022	Anthropic	Important background for safety principles, critique, revision, and AI feedback.
Tool use	Toolformer: Language Models Can Teach Themselves to Use Tools	2023	Meta AI	Shows how language models can learn when and how to call APIs.
Reflection	Reflexion: Language Agents with Verbal Reinforcement Learning	2023	Princeton / Northeastern	Connects feedback, episodic memory, and better future agent behavior.
Self-improvement	Self-Refine: Iterative Refinement with Self-Feedback	2023	Allen AI / CMU / UW / others	Background for reviewer agents and iterative output improvement.
Agent memory	Generative Agents: Interactive Simulacra of Human Behavior	2023	Stanford / Google	Classic observation, memory, reflection, and planning architecture.
Planning	Tree of Thoughts: Deliberate Problem Solving with Large Language Models	2023	Princeton / Google DeepMind	Introduces search over reasoning paths for harder planning tasks.
Lifelong agent	Voyager: An Open-Ended Embodied Agent with Large Language Models	2023	NVIDIA / Caltech / UT Austin / Stanford / ASU	Skill library, automatic curriculum, environment feedback, and self-verification.
RAG foundation	Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks	2020	Meta AI / UCL / NYU / FAIR	Core paper for parametric plus non-parametric memory.
Long context	Lost in the Middle: How Language Models Use Long Contexts	2023	Stanford / UC Berkeley / Samaya AI	Shows why long context is not the same as reliable retrieval or memory.
Multi-agent framework	AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation	2023	Microsoft Research	Connects multi-agent conversation patterns, tools, humans, and framework design.
Multi-agent software	ChatDev: Communicative Agents for Software Development	2023	Tsinghua / OpenBMB	Role-based multi-agent software development workflow.
Agent benchmark	AgentBench: Evaluating LLMs as Agents	2023	Tsinghua / Ohio State / UC Berkeley	Evaluates LLMs as agents across interactive environments.
Agent survey	A Survey on Large Language Model based Autonomous Agents	2023	Renmin University / others	Broad taxonomy for LLM-based autonomous agents.
RAG eval	RAGAS: Automated Evaluation of Retrieval Augmented Generation	2023	Exploding Gradients / CardiffNLP	Practical reference for RAG evaluation dimensions.
Software agents	SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering	2024	Princeton	Shows why the interface between agent and computer matters.
Prompt injection	Prompt Injection attack against LLM-integrated Applications	2023	Nanyang Technological University / Zhejiang University / others	Useful for security modules and tool-result injection threat modeling.
Deceptive behavior	Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training	2024	Anthropic / collaborators	Shows why safety testing should include hidden-trigger and persistence risks.
LLM-as-judge	Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena	2023	LMSYS / UC Berkeley	Foundation for using model judges while understanding bias and limits.

Reading Order

1. ReAct
2. Toolformer
3. WebGPT
4. RAG
5. Lost in the Middle
6. Generative Agents
7. Reflexion
8. Self-Refine
9. Tree of Thoughts
10. Voyager
11. AutoGen
12. AgentBench
13. RAGAS
14. SWE-agent
15. Prompt Injection
16. Constitutional AI
17. Sleeper Agents

Paper Note Template

Title:
Year:
Organization:
Official link:
PDF link:
Topic:
Problem:
Key idea:
System component:
Engineering takeaway:
Possible implementation:
Limitations:
Related roadmap modules:
Related examples:

Contribution Rule

When adding a new paper:

• link to the official paper page
• link to the official PDF only when available
• write your own summary
• explain the engineering implication
• connect it to at least one course module or example
• do not copy abstracts, figures, or PDF contents into this repository