人大高瓴

赵鑫课题组

一面

  • 6分钟内代码相关项目介绍
  • 基础问题
    • 什么是贝叶斯公式、有什么应用
    • 求数组第 kk 大数
  • 保研和直博意向相关(为什么加入课题组和直博等等)

二面

基础论文阅读

任选下面论文列表中的论文,做一个严格控制在10分钟的展示,重点展示你的思考、猜想、探究等等(内容不限),可以有侧重性,自由选择论文即可。

论文列表(个人归类):

  • 强化学习与复杂推理

    • R1-searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning
    • R1-Searcher++: Incentivizing the Dynamic Knowledge Acquisition of LLMs via Reinforcement Learning
    • DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
    • What’s Behind PPO’s Collapse in Long-CoT? Value Optimization Holds the Secret
    • An Empirical Study on Eliciting and Improving R1-like Reasoning Models
    • DAPO: An Open-Source LLM Reinforcement Learning System at Scale
  • 预训练与模型架构

    • Attention Is All You Need
    • LoRA: Low-Rank Adaptation of Large Language Models
    • RoFormer: Enhanced Transformer with Rotary Position Embedding
    • Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
    • DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models
    • Mamba: Linear-Time Sequence Modeling with Selective State Spaces
    • How Does Sequence Modeling Architecture Influence Base Capabilities of Pre-trained Language Models? Exploring Key Architecture Design Principles to Avoid Base Capabilities Degradation
    • Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence
    • AttentionInfluence: Adopting Attention Head Influence for Weak-to-Strong Pretraining Data Selection
  • 多模态

    • Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs
    • DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning
  • 推荐系统

    • A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems
    • Self-Attentive Sequential Recommendation
    • Recommender Systems with Generative Retrieval

领域技能考察

根据组内的研究方向,任选一个或多个(不限)方向的相关材料,我们会针对性地提问。因为要求会略高一些,所以可以不做,根据个人时间和兴趣自由决定即可。

  1. 强化学习与复杂推理

    a. 材料:verl代码库 [https://github.com/volcengine/verl

    b. 大致考察内容:ppo,GRPO基本原理与实现;RL时Environment交互实现方法;RL训练指标含义(reward、length等)

  2. 预训练

    a. 材料:Megatron代码库 https://github.com/NVIDIA/Megatron-LM

    PyTorch自定义操作 https://zhuanlan.zhihu.com/p/344802526

    b. 大致考察内容:如果想改进Attention结构,具体应该如何修改;用autograd.Function写一个激活函数是ReLU的FFN,带重计算

  3. 信息检索

    a. 材料:RAG Survey https://dl.acm.org/doi/pdf/10.1145/3637528.3671470

    b. 大致考察内容:RAG系统搭建需要什么组成模块,有什么核心的可提升点;如何设计多轮搜索规划流程?(训练时和测试时)

  4. 推荐系统

    a. 材料:RecBole代码库 https://github.com/RUCAIBox/RecBole

    b. 大致考察内容: DeepFM,SASRec,LightGCN基本原理与实现;BPR损失原理与实现;推荐训练指标(AUC,NDCG,Recall等)

  5. 模型架构

    a. 材料:A Survey on Inference Optimization Techniques for Mixture of Experts Models

    b. 在GPU显存不足以存放完整的MoE的情况下,如何再保证性能和吞吐率的同时,实现加载和推理,请给出几种解决方案

  6. 多模态

    a. 材料:Why are Visually-Grounded Language Models Bad at Image Classification?; https://github.com/LLaVA-VL/LLaVA-NeXT

    b. 大致考察内容:如果研究多模态模型的知识边界,应该如何定义问题,设计实验?;如何实现图像编码的any resolution 方法?