wanshuiyin/aris-in-ai-offer

ARIS-in-AI-Offer is a study pack of interview cheat sheets aimed at students preparing for Chinese tech industry recruiting season, known locally as qiuzhao. The cheat sheets cover modern machine learning topics: large language models, multimodal models, diffusion models, agents, generative models, and the reinforcement learning techniques used to fine-tune them. Each sheet ships in both Chinese and English versions. Every cheat sheet follows the same three-pillar layout. The first part is a foundations section with formula derivations, intuition, and a short summary. The second part lists 25 high-frequency interview questions, stratified into three difficulty levels: L1 for essentials, L2 for advanced, and L3 for top-tier research lab questions. The third part is from-scratch PyTorch code, including realistic examples such as training a diffusion model with classifier-free guidance and DDIM sampling. The cheat sheets are not handwritten. They are generated by a workflow from a larger project the author maintains called ARIS, which stands for Auto Research in Sleep. ARIS is an agent-based platform with more than 74 research skills covering the lifecycle of academic work, from idea exploration through experiments, papers, rebuttals, and talk slides. Two of those skills, called /interview-cheatsheet and /render-html, produced everything in this repository. The README mentions that ARIS runs on multiple agent platforms including Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, and OpenClaw, and also has a standalone CLI called ARIS-Code. A central piece of the methodology, as described in the README, is cross-model adversarial review. The model that writes a tutorial and the model that audits it must come from different families, so no language model ever grades its own output. The README points to per-tutorial .review.json files as the audit trail. The output format is single-file HTML with MathJax for formulas, highlight.js for code colouring, a sticky table of contents, and a responsive layout designed to read cleanly on phones, tablets, and laptops with no backend. The tutorial index, grouped by area, includes attention mechanisms, KL divergence in RLHF, RLHF variants such as DPO, GRPO, and PPO, reasoning models including o1 and R1, on-policy distillation, and Mixture of Experts architectures from DeepSeek-V3, Mixtral, and Llama 4. The README also lists badges showing the parent ARIS repository has around 10 thousand GitHub stars and was a Hugging Face Daily Papers number one.