ENTERPRISE · ALL
项目实战与企业级 · 全部题目
共 787 篇，本页第 481-528 篇。← 回到学习路径视图
No.1507What is the role of the context window during LLM inference面试官想考察的远不止“上下文窗口是模型能看到的 token 数”这种概念背诵。真正的意图是：你是否理解上下文窗口在推理时…→No.1508Do you prefer DLMs or LLMs for latency-sensitive applications面试官想考察你在延迟敏感场景下的工程决策能力，而非单纯背诵模型定义。这是一道系统设计 + 工程取舍题，刁钻点在于：DLM…→No.1510What is mixed precision (e.g., FP16) and why is it used during inference面试官想考察你对推理阶段性能优化的底层理解，而非单纯背诵“FP16省显存”。刁钻点在于：为什么推理时不用纯FP16？混合…→No.1511What are the various bottlenecks in a typical LLM inference pipeline when running on a modern GPU面试官想看你是否具备系统级性能分析的硬核能力，而非只背几个“显存不够”的常识。这道题考察类型是工程取舍 + debug，…→No.1512How do you measure LLM inference performance面试官想考察你对 LLM 推理性能的系统性理解，而非零散背指标。核心是区分在线服务（延迟敏感）与离线批处理（吞吐优先）场…→No.1513What are the different LLM inference engines available面试官想看的不是引擎列表，而是你在真实部署场景下的选型判断力。这道题属于系统设计+工程取舍类型，刁钻点在于：多数候选人只…→No.1514What are the challenges in LLM inference面试官想看你能否系统性地拆解 LLM 推理的整条链路瓶颈，而不是只背几个“显存大、速度慢”的泛泛之谈。考察类型是系统设计…→No.1515What are the possible options for accelerating LLM inference面试官想看你是否具备系统级优化思维，而非只背几个加速名词。这道题考察类型是工程取舍 + 系统设计，刁钻点在于：候选人常只…→No.1516What are the different approaches for choosing examples for few-shot prompting面试官想考察你对 few-shot prompting 的工程化理解，而非简单背诵“选几个例子”。刁钻点在于：多数人只会…→No.1517Why is context length important when designing prompts for LLMs面试官想考察你是否真正理解 LLM 的“上下文窗口”不是一个可无限扩展的容器，而是一个受注意力机制、推理成本和性能衰减共…→No.1518What is In-Context Learning (ICL), and how is few-shot prompting related面试官想考察你对大模型核心能力——上下文学习（ICL）的底层理解，而不仅仅是背概念。这属于“概念+工程取舍”型问题。刁钻…→No.1519What is self-consistency prompting, and how does it improve reasoning面试官想考察你对 LLM 推理增强技术的深度理解，不只是背概念。这道题是“工程取舍 + 系统设计”混合型：自一致性（Se…→No.1521What is catastrophic forgetting, and why is it a concern in fine-tuning面试官想看你是否真正理解“微调”的代价，而不是只会跑脚本。这题表面是背概念，实则考察 模型稳定性与泛化性的工程取舍。刁钻…→No.1522What are the strengths and limitations of full fine-tuning面试官想看的不是“全微调能调所有参数”这种表面答案，而是你能否在资源、数据、性能三角中做出工程取舍。这道题属于工程取舍分…→No.1523When might prompt engineering be preferred over task-specific fine-tuning面试官想看的不是“prompt engineering 省钱，fine-tuning 效果好”这种教科书式二分法。真正考…→No.1524| Q4 | How does quantization affect inference speed and memory requirements面试官想考察你对模型部署核心瓶颈的理解深度，而非单纯背诵量化定义。这是典型的工程取舍类问题，刁钻点在于：量化看似简单（F…→No.1525| Q46 | How does Beam Search improve upon Greedy Search, and what is the role of the beam width parameter面试官想考察你对解码策略的工程理解深度，而非仅仅背诵概念。这是典型的“工程取舍”题，刁钻点在于：你能否从“搜索空间 vs…→No.1526| Q47 | When is a deterministic strategy (like Beam Search) preferable to a stochastic (sampling) strategy这道题考察的是工程取舍与任务适配能力，而非单纯背概念。面试官想看你能否跳出“Beam Search 好还是 Sampli…→No.1527| Q50 | How is Beam Search fundamentally different from a Breadth-First Search (BFS) or Depth-First Search (DFS)面试官想考察你对搜索算法本质的理解，而非单纯背诵定义。这道题是典型的“概念对比+工程取舍”类型，刁钻点在于：Beam S…→No.1528| Q53 | What is the role of the context window during LLM inference面试官想看你是否真正理解 context window 不是“内存条”，而是 LLM 推理的硬边界和性能瓶颈。考察类型是…→No.1529| Q59 | Do you prefer DLMs or LLMs for latency-sensitive applications面试官想看你是否具备工程选型的硬核能力，而非只会背模型参数。这道题表面是“DLM vs LLM”，实则考察：① 你是否能…→No.1530| Q62 | What are the challenges in performing distributed inference across multiple GPUs面试官想考察你能否从系统层面拆解分布式推理的瓶颈，而非只背概念。这是典型的工程取舍 + debug 类型问题，刁钻点在于…→No.1531| Q65 | What is continuous batching, and how does it differ from static batching面试官想考察你对 LLM 推理引擎底层调度机制的理解深度，而非简单背概念。这是典型的“工程取舍+系统设计”题，刁钻点在于…→No.1532| Q69 | What are the various bottlenecks in a typical LLM inference pipeline when running on a modern GPU面试官想看你能否系统性地拆解 LLM 推理管线，区分计算密集（compute-bound）和内存密集（memory-bo…→No.1533| Q72 | What are the challenges in LLM inference面试官想看你是否真正理解LLM推理的“瓶颈在哪里”，而不是背一堆优化名词。核心考察三个维度：系统性思维（能否从算法、系统…→No.1534| Q73 | What are the possible options for accelerating LLM inference面试官想看你是否真正理解 LLM 推理的瓶颈（显存带宽 > 算力），而非只会罗列技术名词。考察类型是系统设计 + 工程取…→No.1535| Q79 | What are the different approaches for choosing examples for few-shot prompting面试官想考察你对少样本提示（few-shot prompting）中示例选择策略的深度理解，而非仅背诵“选几个例子放进去…→No.1537| Q82 | What is In-Context Learning (ICL), and how is few-shot prompting related面试官想考察你是否真正理解 ICL 的机制本质，而不仅仅是背定义。这是典型的“概念 + 工程取舍”题，刁钻点在于：很多人…→No.1538| Q83 | What is self-consistency prompting, and how does it improve reasoning面试官想考察你对 LLM 推理增强技术的理解深度，尤其是从“单次推理”到“多次采样+聚合”的范式跃迁。这并非简单的概念背…→No.1539| Q86 | How would you structure a prompt to ensure the LLM output is in a specific format, like JSON面试官想考察的不是你会不会写“请输出JSON”，而是你对LLM输出格式控制的工程级理解。这是典型的系统设计+工程取舍题，…→No.1540| Q88 | What are the different phases in LLM development面试官想看你是否具备 LLM 开发的全局视野，而非只懂某个环节（如只会调 API 或跑微调）。考察类型是系统设计 + 工…→No.1541| Q89 | What are the different types of LLM fine-tuning面试官想考察你对 LLM 微调生态的全局认知，而非死记硬背分类。刁钻点在于：能否区分“全参数微调”与“参数高效微调（PE…→No.1542| Q90 | What role does instruction tuning play in improving an LLM’s usability面试官想考察你是否真正理解指令微调（Instruction Tuning）的本质，而非仅背诵“在指令数据上微调”的定义。…→No.1543| Q92 | How do you prevent overfitting during fine-tuning面试官想考察你对微调过拟合的系统性工程思维，而非单纯背诵正则化名词。刁钻点在于：大模型时代，过拟合的根源已从“参数过多”…→No.1544| Q93 | What is catastrophic forgetting, and why is it a concern in fine-tuning面试官想看你是否理解灾难性遗忘的底层机制，而不仅仅是背定义。考察类型是“工程取舍+debug”，刁钻点在于：多数候选人只…→No.1545| Q94 | What are the strengths and limitations of full fine-tuning面试官想看你能否辩证分析全量微调（Full Fine-Tuning）的工程取舍，而非单纯背诵概念。这是典型的“系统设计+…→No.1546| Q96 | When might prompt engineering be preferred over task-specific fine-tuning面试官想考察你对大模型应用落地中“成本-性能-灵活性”三角权衡的实战理解，而非单纯背诵概念。刁钻点在于：多数人只会说“数…→No.1547| Q100 | When should you prefer task-specific fine-tuning over prompt engineering面试官想考察你对模型定制化手段的工程决策能力，而非单纯背诵概念。核心是判断你能否在数据、性能、成本、维护四个维度上做量化…→No.1548| Q105 | How would you handle LLM fine-tuning on consumer hardware with limited GPU memory面试官想考察你在资源受限场景下的工程化能力，而非单纯背诵微调步骤。这是典型的系统设计+工程取舍题，刁钻点在于：候选人常只…→No.1549| Q108 | What are the possible options to speed up LLM fine-tuning面试官想考察你对大模型微调全流程的瓶颈感知和系统优化能力，而非简单罗列技术名词。刁钻点在于：多数候选人只提LoRA或混合…→No.1550| Q112 | In the context of LLM pretraining, what is scaling law面试官想考察你是否真正理解 LLM 预训练的核心规律，而非仅背诵“参数越多越好”。这是典型的工程取舍 + 概念理解题，刁…→No.1551| Q114 | What is model parallelism, and how is it used in LLM pre-training面试官想考察你对分布式训练核心策略的理解深度，尤其是模型并行（Model Parallelism）在 LLM 预训练中的…→No.1552How does the Faithfulness metric help diagnose this issue面试官想考察你对 RAG 评估的深度理解，而非单纯背诵 Faithfulness 定义。刁钻点在于：你是否能把一个抽象指…→No.1553Explain how hallucinations in LLMs specifically impact the Faithfulness metric. What techniques could you implement to improve the Faithfulness metric score面试官想考察你对 LLM 幻觉与评估指标之间因果关系的深度理解，而非简单背诵定义。刁钻点在于：Faithfulness …→No.1554How can including the faithfulness metric improve reliability面试官想考察你对 RAG 系统可靠性保障的工程落地能力，而非背诵“Faithfulness 就是忠实度”的定义。刁钻点在…→No.1555What are the pros and cons of query transformation techniques面试官想考察的不是你能否背诵“查询重写、分解、扩展”这几个名词，而是你是否理解每种变换技术背后的工程取舍：为什么用、什么…→No.1556When do you opt for hybrid search instead of semantic search面试官想考察你对检索系统的工程判断力，而非单纯背诵概念。这道题属于工程取舍类型，刁钻点在于：语义搜索（如 DPR、Col…→No.1557Why is the cross-encoder typically used as the re-ranker rather than the bi-encoder面试官想看你是否真正理解双编码器（Bi-Encoder）与交叉编码器（Cross-Encoder）在架构上的本质差异，以…→