记录一下这次跳槽面试过程中遇到的面试题，供大家参考。

背景：2yoe，多大CS研究生，有paper。跳槽的时候在银行做Applied ML Scientist，海投的职位也类似比如Applied Scientist，Machine Learning Engineer，Machine Learning Scientist，Research Engineer和AI Scientist。下面按公司总结一些遇到的面试题：

某V开头的投资公司，MLE岗：

• General ML
  • L1 and L2 regularization
  • data drift vs model drift, detection method (PSI)
  • end-to-end flow from ideation to productionization of ML model
  • model serving
• Agent related
  • agent-to-agent communication, MCP
  • short-term vs long-term memory, how to design the memory module
  • Agent RL
• MLOps
  • Docker (ML模型上线，docker因为访问量过大崩了，应该怎么处理), Kubernetes, MLflow

粉车，MLE岗：

• General ML
  • cross validation, ask you to look at training and validation curves and tell if the model is overfitting or underfitting, and how to fix it. What if in production the model is underperforming, how to handle it?
  • Mathematically, why multicolinearity is a problem for linear regression, how L1 regularization can help with it? (multicolinearity -> linear dependence -> not full rank -> not invertible -> no unique solution)
  • L1 and L2 regularization, how to choose between them?
• Coding
  • Numpy implementation of Kmeans and extended to Kmeans++
  • LC 九八一，七六

某红色银行，AI Scientist岗：

• Agent related
  • discuss my RAG project for 1h, ask about the details of the retriever and generator, and how to evaluate the performance of the RAG system
  • How to reduce the latency? How to improve the retriever performance?

某A开头的药企，Senior AI Scientist岗：

• Deep Learning
  • Transformer architecture, attention mechanism
  • Position encoding
  • Different inference sampling methods for diffusion (DDPM-Solver，DDIM)
  • Image tokenizer for autoregressive image generation
• Research Presentation
  • 30min presentation on my research, followed by 30min Q&A. Questions are mostly about the details of the model architecture and training process, i.e. what if the loss is not converging or goes to NaN, how to debug?
• Coding
  • implement a very simple agentic or LLM-based framework to generate questions and answers for a given topic, and then evaluate the quality of the generated questions and answers.

某O开头的养老金公司，MLE岗：

• General ML
  • lightgbm vs xgboost
• Deep Learning / LLM
  • Transformer architecture, attention mechanism
  • LoRA PEFT
  • preprocessing for LLM post-training data
  • how to preprocess long docs, OCR on long docs, how to align table data and text analysis if the content span multiple pages, how to handle the page breaks in the middle of a table, etc.
• Agent related
  • Design an agentic system to automate the process of financial report analysis, including data extraction, summarization, etc. Discuss the data pipeline, generation pipeline, and how to evaluate its performance.

最后祝大家跳槽顺利！