km1994/LLMs_interview_notes

What is this?

LLMs_interview_notes (also referred to in the README as the "AIGC Interview Book") is a Chinese-language interview question bank for large language model (LLM) algorithm roles. It contains 800+ high-frequency interview questions organized by topic, not by company or article like typical interview repositories.

Key features

• Organized by exam point: Each question has a difficulty rating (⭐, ⭐⭐, ⭐⭐⭐)
• Structured answers: Every question follows the same three-part format:
  1. A one-sentence answer
  2. A deeper explanation of why
  3. What follow-up questions the interviewer is likely to ask
• Focus on reasoning, not memorization: Questions emphasize "why," "how to weigh trade-offs," and "how to implement" rather than definitions
• Up-to-date: Covers topics through 2026, including Harness Engineering, Loop Engineering, GraphRAG, and MCP
• Community-driven: Users can submit new questions via Issues and fix answers via Pull Requests

Content modules

Module	Question count	Detailed explanations available
RAG (Retrieval-Augmented Generation)	160	40 questions fully explained
Agent	100	24 questions fully explained
RL post-training (RLHF/PPO/DPO/GRPO)	120	40 questions fully explained
Hermes Agent	100	22 questions fully explained
Loop Engineering	100	100 answer key points
GraphRAG (knowledge graph)	100	100 answer key points
Object detection & computer vision	100	100 answer key points
Classic topics (30+ subtopics)	Hundreds	Topic lists + per-topic answer links

Classic topics include: LLM basics, fine-tuning, LangChain, RAG, PEFT, inference, distributed training, Agent, position encoding, tokenizer, hallucination, model comparison, CoT, MoE, distillation, multimodal, KV Cache, role-play, DeepSeek-R1, MCP, context engineering, Qwen3, and more.

126 questions with full detailed explanations are already open-sourced in the repository, covering RAG foundations, chunking strategies, embeddings, Agent basics, RLHF, PPO, DPO, and Harness Engineering.

Who is it for?

• Students or professionals preparing for LLM algorithm or AI Agent interviews (campus, experienced, or career-change)
• Engineers transitioning from CV, NLP, or backend into AIGC
• Interviewers who need difficulty-tiered questions they can use directly
• Anyone wanting a systematic review of LLM knowledge

How to use it effectively (from the README)

1. First pass: Scan ⭐ questions to build a mental map — about 20 questions per day for two weeks
2. Practice speaking: Close the screen and explain ⭐⭐ questions aloud for one minute, then check against the answer
3. Mock interviews: Use ⭐⭐⭐ scenario questions (from Loop Engineering, GraphRAG, object detection) to design solutions and discuss trade-offs — each includes "what's being tested" and "what earns points"

Suggested paths by role:

• Algorithm roles: Classic topics (basics → fine-tuning → distributed training → memory optimization), then RL post-training (120 questions)
• Agent/application roles: Agent (100) → Hermes → Loop Engineering → MCP/context engineering
• Transitioning from CV/NLP: Basics + position encoding + tokenizer, then RAG and Agent
• Only one week: Just ⭐ questions plus the classic topic section for your target role

Example question format

【⭐⭐】Why is it said that "70% of RAG effectiveness is determined by chunking"?

💡 What's tested: Whether you understand chunking's decisive impact on retrieval quality, not just tuning chunk_size
🎯 What earns points: Explaining the chain from chunking → retrieval → generation; giving concrete chunk size ranges
✅ Answer key points: Chunking determines the information completeness of retrieval units. Too small breaks semantics; too large drowns in noise...

Example real interview question (from the README)

"Can long-context models (e.g., Gemini 1.5 Pro with 1M token context) replace RAG?"

The README says the interviewer doesn't want a simple "yes" or "no" — they want you to break it down across cost, timeliness, and traceability dimensions. Being able to identify those dimensions is what earns a high score.

Contribution model

• Issues: Submit new questions
• Pull Requests: Fix or improve answers
• The maintainer reviews every submission

Current status

• Actively maintained with recent commits
• Badges show stars, forks, and last-commit status
• PRs are welcome (per the README badge)