aris-in-ai-offer

ARIS-in-AI-Offer (ARIS in 秋招)

Hoping to make your 秋招 (qiūzhāo, Chinese AI campus recruiting season) a little easier 🌱

📖 中文版 (Chinese version): README_CN.md

📱 One link for everything: easyaioffer.github.io (short address; the site itself is wanshuiyin.github.io/ARIS-in-AI-Offer) — every sheet on one page, searchable, 中/EN switch, dark mode, and a per-reader 已读 tracker. Save it on your phone.

📚 Jump to a topic — 35 first-party cheat sheets across 8 categories, plus a community-contributed question bank:

🧠 General / Foundations · 🎯 Post-Training & Reasoning · 🏛️ LLM Architecture & Systems · 🌊 Generative Models — Theory & Tokenizers · 🎨 Generation Systems (Image / Video / 3D / Diffusion Post-Training) · 👁️ Multimodal · 🤖 Agents · 🦾 Embodied AI & World Models / 具身与世界模型

Or browse the full 📚 Tutorial Index ↓ · jump to 🌐 ARIS-Homepage ↓.

· · · · ·

🏆 Built on a battle-tested foundation — the ARIS main repo has ~10k GitHub stars, was HuggingFace Daily Papers #1, won AI Digital Crew Project of the Day, and ships 74+ research skills across 7+ platforms. This isn't a vaporware preview — every cheat sheet here is the production output of the same /interview-cheatsheet + /render-html workflow used in academic-research production.

A curated, bilingual (中文 + English) collection of ML / LLM / multimodal / diffusion / agent / generative-model interview cheat sheets, auto-generated by the ARIS — Auto Research in Sleep /render-html workflow.

Each cheat sheet is a long-form Chinese tutorial with: formula derivations · from-scratch PyTorch code · 25 high-frequency interview questions (L1 essentials · L2 advanced · L3 top-tier lab).

📖 Preview (above): one snapshot per pillar, taken from the Diffusion Foundations cheat sheet — ① Foundations (formula derivations + intuition + TL;DR), ② Interview Q&A (25 high-frequency questions stratified L1/L2/L3), ③ From-Scratch Code (runnable PyTorch, including CFG training + DDIM sampling). Every cheat sheet in this collection follows the same three-pillar structure.

🌐 ARIS-Homepage preview — CV → fact-checked academic homepage (click to expand)

Same /render-html workflow turning a CV into a fact-checked academic homepage. Live demo at wanshuiyin.github.io. Details + pipeline diagram in the ARIS-Homepage section ↓.

📝 Long-form blog preview — A Survey on Continuous DLM (2026 H1) (click to expand)

A standalone hand-authored long-form technical survey. A Survey on Continuous DLM (2026 H1, 6 papers) — Chinese-language survey by Ruofeng Yang (SJTU), written end-to-end via cross-model discussion (Claude Opus 4.7 + Codex GPT-5.5 xhigh + Gemini auto-gemini-3). 📖 Read full blog ↗.

📱 HTML reads cleanly everywhere

Phone on the subway, iPad at a café, laptop in the library — same HTML link opens equally well:

• 🧮 MathJax renders all LaTeX formulas (not screenshots — scalable, copyable, selectable)
• 💻 highlight.js colors all PyTorch code blocks
• 📐 Responsive layout adapts to any window width — no overflow, no blur
• 📑 Sticky TOC for jumping around long documents
• 💾 Single-file HTML — download once, read offline, no backend required

📢 What's New

• 2026-09-11 — 🌍 #35 World Models + a 2025–26 refresh of 3D Generation and Quantization — from community feedback ("3D 和 WM 有点少/老", "quant 主流用 PTQ 和 QAT"). World Models is new and deliberately curated — influential, representative, major-lab, open-weight work only, with the selection rule stated in the sheet and the famous-but-excluded listed with reasons: latent dynamics (RSSM → Dreamer v1–4, MuZero, TD-MPC2) · JEPA (I-/V-JEPA, V-JEPA 2/2.1, DINO-WM) · generative/interactive (Genie → Genie 3, GameNGen, Cosmos 1→3, Matrix-Game 3.0, Atlas) · three embodied routes · evaluation and the definition fight. 3D gains feed-forward reconstruction (DUSt3R → VGGT → MapAnything / DA3, with the confidence-loss derivation), feed-forward 3DGS, native mesh generation (MeshGPT → Meshtron, tokenization budgets), TRELLIS.2 / Hunyuan3D 2.x on a latent-representation axis, part-level and sim-ready assets; five duplicated L3 questions replaced. Quantization §10 rewritten around QAT: PACT/LSQ derivations, ParetoQ, Gemma 3/4 official QAT weights, Apple's 2-bpw on-device stack, DeepSeek FP8 / NVFP4 / MXFP8 training recipes — and the 2026 PTQ-vs-QAT guidance with costs cited rather than asserted. All bilingual; new runnable code/world_models_toy.py (RSSM balanced-KL fixed point, JEPA shortcut, gridworld LAM — asserts against analytic answers). Research fanned out to Opus 5 with every arXiv ID fetched; drafts reviewed by gpt-6-astra xhigh in 2–3 rounds each. world_models_tutorial.html.
• 2026-09-10 — 🎯 #34 Modern Diffusion Post-Training: Flow-GRPO · DGPO · DiffusionNFT — the three 2025 online-RL methods for flow-matching T2I, organised around the one thing that separates them: how the reward enters the gradient — per-step policy gradient on a marginal-preserving SDE, group-level DPO with the $\log Z$ cancellation, or a reflected two-branch regression on the forward process. Flow-GRPO moves here out of the Diffusion Post-Training sheet, with its reverse-SDE sign and closed-form KL corrected on the way. Bilingual, with a runnable script (code/diffusion_online_rl.py — analytic checks of marginal preservation, the NFT update sign, and DGPO's balanced weights) and 25 高频题. Design-reviewed before drafting (64 guardrails) plus three review rounds, all gpt-6-astra xhigh. modern_diffusion_post_training_tutorial.html.

📋 Earlier updates (22)

• 2026-07-31 — ✂️ Interview-scannability overhaul for #30-33 + collection-wide table-scroll fix — responding to reader feedback ("comprehensive but unscannable before an interview"): Transformer Block restructured — the concrete block assembly + dataflow + runnable code moved from a buried §7 up to §1 (answer first, rationale after); all four de-hedged — lead sentence states the conclusion, qualifiers demoted to notes, one canonical home per caveat (inference visible prose -33.5%, transformer_block -17.2%, eval -17.4%), zero factual changes; a quantitative screen confirmed the older 29 tutorials healthy. Renderer wide-table overflow fixed, all 67 HTMLs re-rendered; all four EN editions retranslated + fidelity-reviewed. Gate PASS. (#37 · 28e4fb9)
• 2026-07-22 — 🧱 4 new cheat sheets (#30-33): Transformer Block · LLM Evaluation & Benchmarking · LLM Pretraining Pipeline · LLM Inference & Serving Stack — four foundational/systems topics in one batch: residual topologies & the MQA/GQA/MLA design axes, unbiased pass@k estimation & LLM-as-judge, Kaplan vs Chinchilla & the corpus-factory data pipeline, the request state machine & PagedAttention/KV lifecycle — 25-30 interview questions each. First batch to move the cross-model design review before drafting (90-105 numbered guardrails each), then 3-5 independent GPT-5.6-sol review batches; 3 real bugs were each caught independently by multiple batches (Bradley-Terry separation criterion, the best-of-n KV formula). All bilingual with runnable scripts, gate PASS. (#36 · 74644cf)
• 2026-07-13 — 🔍 Diffusion-cluster resweep completes the sweep — 8 tutorials, 66 fixes, all 28 tutorials now under GPT-5.6-sol — covers the diffusion/generative-media cluster deferred from the prior resweep (07-12). Same two-stage pipeline (GPT-5.6-sol finds → independent Claude adversarially verifies): 67 candidates → 66 fixed, 2 REFUTED left untouched; includes 5 real code bugs (iCT/FSQ/LFQ/DDPO) and a recurring "conditional path is straight" ≠ "marginal ODE trajectory is straight" confusion. Gate PASS. (#32 · bfae8f1)