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

📚 Jump to a topic — 29 first-party cheat sheets across 7 categories + 1 community-contributed category:

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

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.

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 ↓.

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

• 2026-07-13 — 🔍 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)
• 2026-07-12 — 🔍 Full-collection resweep — 20 tutorials, 236 fixes, cross-model review upgraded to GPT-5.6-sol — after the reviewer moved from GPT-5.5 to GPT-5.6-sol, re-audited training fundamentals / attention / RLHF / inference systems / PEFT / agents / RAG·VLM (20 files). Same two-stage pipeline: 240 candidates → 236 fixed, 4 REFUTED left untouched; includes 5 real code bugs and StarPO's acronym settled from the paper's own abstract. Gate PASS. (#31 · 59636aa)
• 2026-07-11 — 🔤 #29 Tokenization cheat sheet — fills the foundation the collection had been taking for granted: BPE / WordPiece / Unigram (training ≠ encoding trap) · byte-level BPE vs byte fallback · the easily-confused vocab quantities · the tokenizer-free frontier. Bilingual, with a runnable script (code/tokenization.py — exact GPT-2 bytes_to_unicode, Unigram DPs hand-checked, pure stdlib, verified on a real box) and 25 高频题. First tutorial with cross-model review moved up to the design stage (81 guardrails before drafting + 5 review batches + an ultra final pass, corrections flowing both ways). tokenization_tutorial.html.
• 2026-06-29 — ⚙️ #28 Optimizers & LR Schedules cheat sheet — the optimizer/schedule training-mechanics gap: SGD·Momentum·Nesterov · AdaGrad→RMSProp→Adam (the adaptive lineage) · Adam bias correction (the uncorrected first step is ~3.16× too large, not small — the direction trap) · AdamW decoupled weight decay (Adam's L2 ≠ weight decay) · frontier (Muon / Lion / Shampoo / SOAP / Adafactor / LAMB / Sophia) · LR schedules (warmup / cosine / Noam / WSD / one-cycle) · weight decay + LR-batch scaling + no-decay groups · grad clipping + LLM hyperparams (β₂=0.95). Bilingual (中文 + EN), with a runnable script (code/optimizer_lr_schedule.py — SGD/Adam/AdamW from-scratch vs torch.optim, the executable AdamW≠Adam+L2 proof, the bias-correction direction, cosine-warmup, verified on a real box) and 25 高频题. optimizer_lr_schedule_tutorial.html.

• 2026-06-19 — 🧱 #27 Normalization / Residual / Init cheat sheet — the foundational training-mechanics hole: BatchNorm / LayerNorm / RMSNorm / GroupNorm · Pre-vs-Post-LN (Xiong's gradient argument — why Post-LN needs warmup) · DeepNorm / Sandwich / QK-Norm · residual connections + scaling (LayerScale / ReZero / GPT-2's 1/√(2N)) · Xavier vs Kaiming (the ReLU factor-2, and why the preserved quantity is the second moment E[y²], not Var) · μP (width-invariant HP transfer) · Fixup / NFNets / DyT (norm-free) · plus the covariate-shift debunk (Santurkar). Bilingual (中文 + EN), with a runnable script (code/normalization.py — from-scratch LN/RMSNorm vs torch + Pre/Post-LN gradient + Kaiming second-moment + GPT-2 residual scaling, verified on a real box) and 25 高频题. The cross-model review caught a real methodological confound in the Pre/Post-LN gradient demo (an output-normalization artifact) and redesigned it to a loss-robust top/bottom weight-grad ratio. normalization_init_tutorial.html.
• 2026-06-18 — ⚡ #26 Linear / Sparse Attention cheat sheet — the sub-quadratic / efficient-attention hole the collection kept referencing but never derived: linear attention (kernel φ + associativity → matrix-state RNN) · SSM / Mamba (selective S6) · Mamba-2 / SSD (1-semiseparable duality ≡ structured masked linear attention) · DeltaNet / Gated DeltaNet (overwrite update) · chunkwise-parallel training · trainable sparse (NSA three-branch / MoBA / Lightning / DSA) · hybrid (Jamba / Hymba / Qwen3-Next / Kimi-Linear / MiniMax-01). Bilingual (中文 + EN), with a runnable script (code/linear_sparse_attention.py — chunkwise ≡ recurrent equivalence + delta-rule + block-sparse, verified on a real box) and 25 高频题, settled through multi-batch Codex GPT-5.5 xhigh citation / math / code / answer / overall + render-fidelity review. linear_sparse_attention_tutorial.html (#22 · b84e913).
• 2026-06-13 → 06-14 — 🧩 Two must-know cheat sheets shipped: #24 LoRA / PEFT + #25 RAG + Embedding / Retrieval — the two glaring holes the collection used everywhere but never derived. LoRA/PEFT: B=0 identity start · α/r vs rsLoRA √r · zero-latency merge · QLoRA (NF4 / double-quant / paged) · DoRA · the family vs Adapter / Prefix / Prompt / BitFit. RAG: bi-encoder · InfoNCE + hard negatives · Matryoshka · BM25 · HNSW · RRF hybrid · cross-encoder vs ColBERT late interaction · HyDE / Self-RAG / CRAG · GraphRAG · RAGAS. Both bilingual (中文 + EN), each with from-scratch PyTorch, a runnable script (code/lora.py · code/rag_embedding.py, verified on a real box) and 25 高频题, settled through multi-round Codex GPT-5.5 xhigh math/code + render + EN translation-fidelity review. lora_peft_tutorial.html · rag_embedding_retrieval_tutorial.html.
• 2026-06-08 — 🔧 tools/render_html.py now strips a leading UTF-8 BOM before frontmatter detection (6cc4876).
• 2026-06-05 — 🔭 Community Showcase: an online tutorial-collection site by @QiZishi — merges all 23 Chinese tutorials here with Datawhale Hello-Agents' LLM interview Q&A into one card-index reading site (read online · from #3). The README gains a Community Showcase section listing community derivatives — build something on top of these tutorials and open an issue to get it listed (f8e7d33).
• 2026-06-02 — 📝 Two new blogs — Continuous DLM (representation perspective, v2) + Diffusion × Representation × Manifold — the Continuous DLM survey is upgraded to a representation-perspective expansion (replacing the earlier v1), and a companion blog traces the "borrow representation / use the manifold" threads across image & video diffusion (SSL / Consistency / REPA / RAE / JiT / V-JEPA2). Both hand-authored, cross-model reviewed; the two cross-reference each other. Live: continuous_dlm_representation_perspective.html · diffusion_representation_manifold.html. See the Blog Index ↓.
• 2026-06-01 — 📝 New blog: a deep-dive guide to NVIDIA Cosmos 3 (omnimodal world model · MoT architecture) — a long-form Chinese popsci walkthrough of the 138-page Cosmos 3 technical report (15 sections · 12 figures): how understanding and generation are stitched into one Transformer (Mixture-of-Transformers), the training recipe, scaling / serving, and the three model sizes. Reviewed end-to-end across 5 rounds of Codex GPT-5.5 xhigh cross-model audit (numbers + mechanisms checked, overclaims trimmed). Self-contained single-file HTML at docs/blogs/cosmos3_mot_guide.html — a hand-authored guide (ELF blog format), not part of the audited /render-html pipeline. All figures and numbers are from the NVIDIA Cosmos 3 report (github.com/nvidia/cosmos); © the original authors, used here as an attributed popsci guide.
• 2026-05-31 — 🔍 Cross-model audit pass — real errors caught & fixed across the whole collection, now CI-enforced — a fresh Codex GPT-5.5 xhigh re-review of every cheat sheet (中文 + EN) surfaced genuine technical mistakes the first pass had missed, and fixed them: DeepSeek-V3's FP8 GEMM accumulates in FP32, not bf16; Qwen2-VL M-RoPE base is 1e6; Molmo's vision tower is CLIP, not SigLIP; TensoRF complexity is O(N³); a broken StreamingLLM render callout; plus 10 EN translation-fidelity fixes (meaning reversals, a dropped framework list). Every docs/ artifact now carries a traceable cross-model review, and a new CI gate (tools/verify_reviews.py --mode strict --reproduce, wired in review-audit.yml) blocks any PR whose HTML isn't a reviewed view that byte-reproduces from its source — so no un-reviewed or hand-edited content can land. Also merged today: a Focus-dim reading mode for --blog-mode pages (dim non-current sections · floating 🎯 toggle · ↑/↓ section nav), reflowed from the dllm blog effort into the academic template — dormant for the cheat sheets (gated under .aris-blog), it activates in blog-mode, ready for the first blogs rendered through it (landing soon). (5aae952 · 498ecf2)
• 2026-05-28 — 📝 First blog shipped: A Survey on Continuous DLM (2026 H1, 6 papers) — long-form Chinese technical survey by Ruofeng Yang (SJTU); superseded by the v2 rewrite at the same path, see the 2026-06-02 entry, written end-to-end through cross-model discussion (Claude Opus 4.7 + Codex GPT-5.5 xhigh + Gemini auto-gemini-3). Compares ELF, ByteDance Cola-DLM, and Flow-Matching family across discrete-DLM problems, the "known-unknown" continuous space idea, training pipeline, architecture / params / shapes, inference grids + Tab 6/7 numerical results, denoising trajectories, and a Field Landscape against Cola-DLM. Lives at docs/blogs/continuous_dlm_representation_perspective.html (1.7 MB self-contained, no build) — a hand-authored long-form HTML, outside the audited /render-html pipeline. Preview strip + live link at the top of this README. (8475a2d)
• 2026-05-28 — ⚡ render-html P0 polish + all 23 tutorials regenerated — academic template gained 7 interactive features: print degradation fix (PDF no longer loses <details> content), TOC sidebar scrollspy (current section auto-highlights as you scroll), figure lightbox (native <dialog> with focus trap + Esc), long-code auto-collapse (<pre> ≥30 lines wrapped in <details class="code-card">, per-block override via ```python {collapsed} / {open} fence flags), paper citation popover (new [[key]] MD syntax + --papers <papers.json> sidecar), eyebrow cleanup (marketing uppercase → body-serif gray), --blog-mode infrastructure (opt-in aris-blog body class). XSS-hardened script injection via json_for_script() (escapes </script> break-out). All 23 bilingual tutorial pairs (= 46 HTMLs) regenerated to pick up the new template shell — source MDs untouched. Codex GPT-5.5 xhigh 4-round review (design × 2 → code × 1 → spot-check × 1). Try it: scroll attention_tutorial.html and watch the TOC sidebar follow (b79c57d, 8793f40).
• 2026-05-26 — 🐍 5 runnable PyTorch tutorial scripts — first runnable-code contribution in docs/tutorials/code/: mha.py (MHA + causal mask) · axial_attention.py (H/W axial + complexity table) · flow_matching.py (Rectified Flow on 2D moons) · mmdit_block.py (double-stream MMDiT block) · toy_mmdit_t2i_pipeline.py (end-to-end T2I skeleton). Pure PyTorch, CPU-runnable in seconds, every script ships with built-in assert sanity checks (shape parity, numerical agreement with nn.MultiheadAttention where applicable). Pairs with attention_tutorial.md / flow_matching_tutorial.md / image_generation_systems_tutorial.md (f63f468).
• 2026-05-24 — 🐙 ARIS-Homepage v1.1: --from-repos — snapshot user-selected owner/repo list via gh CLI; LLM agent merges repo timelines into homepage News + featured_projects[].github. Private repos skipped by default. Closes #2 by @Yafei-Liu99 (cdcf9a2).
• 2026-05-23 — 🌐 ARIS-Homepage v1 shipped — CV → fact-checked academic homepage (DBLP / arXiv audit blocks wrong venue / year / author). Single-file HTML; Codex / Gemini reviews optional. Live demo: wanshuiyin.github.io. Skill: skills/homepage-generator/SKILL.md (b818c1d).
• 2026-05-22 — 🦾 Featured community contribution: 具身智能高频面试题库 by @WinstonJQ — 413 questions across 8 卷 (VLA / 模仿学习 / RL / 世界模型 / 工程落地 / 腿足控制 / 3D 感知 / 系统设计). Hosted externally; linked from the new "🦾 Embodied AI" category in the Tutorial Index (b1ebb6f).
• 2026-05 — 📚 4 new bilingual cheat sheets: KL Divergence in RLHF (k1/k2/k3 · placement gradient bias), LLM On-Policy Distillation (MiniLLM / GKD / Qwen3 / Tinker), Diffusion Post-Training (DDPO / DPOK / DRaFT / AlignProp / Diffusion-DPO / Flow-GRPO), Diffusion / Flow Distillation (CM / iCT / sCM / CTM / LCM / DMD/DMD2 / ADD/LADD). Total now: 23 first-party cheat sheets.
• 2026-05 — 📖 README restructure — preview-strip banner, ARIS credentials at top (badges + 10K-star foundation paragraph), shared WeChat community QR with the main ARIS repo.

📝 Blog Index

Long-form technical blogs — hand-authored, cross-model reviewed; outside the audited /render-html pipeline (figures © their original authors, used with attribution).

📚 Tutorial Index

🌐 Bilingual editions: every cheat sheet ships with both a Chinese (default) and an English HTML — filenames are *_tutorial.html (CN) and *_tutorial_en.html (EN). HTML columns below link to both.

🧠 General / Foundations

🎯 Post-Training & Reasoning

🏛️ LLM Architecture & Systems

🌊 Generative Models — Theory & Tokenizers

🎨 Generation Systems — Image / Video / 3D / Diffusion Post-Training

👁️ Multimodal

🤖 Agents

🎉 23 tutorials live (bilingual) (2026-05) — each ships with both Chinese and English HTML. Seven buckets: General · Post-Training · Architecture · Generative · Multimodal · Agents · Diffusion Post-Training. This round adds 4 new sheets: KL Divergence in RLHF, LLM On-Policy Distillation, Diffusion Post-Training, Diffusion Distillation. More (Flow-OPD / Audio Gen / further SOTA updates) coming — PRs welcome (see CONTRIBUTING).

🦾 Embodied AI / 具身智能

🌟 Community contribution by @WinstonJQ — hosted externally on a separate repo, generously shared with the community. If it helps your interview prep, please ⭐ the source repo to thank the author 🙏

🤖 How These Are Generated

Every tutorial uses ARIS's /interview-cheatsheet skill:

1. Plan — 12-14 sections (TL;DR · Intuition · Formulas · Code · Variants · Complexity · 25 Q&A)
2. Draft — 600-1000 lines of Chinese tutorial + runnable from-scratch PyTorch
3. Cross-model review — fresh-thread codex GPT-5.5 xhigh audit on 10 properties (formula correctness · code runnability · citation accuracy · table-pipe escapes · callout style · personal-info leak · ...)
4. Fix loop — trajectory-based; keep going if FAIL set is shrinking, stop if same issue recurs or ~6 rounds without convergence
5. /render-html — single-file HTML render + 13-property render audit (information fidelity · TOC · math · code highlight · safety · privacy · ...)
6. .review.json — full audit trail saved next to each tutorial

Cross-model adversarial review (executor ≠ reviewer family) is ARIS's core invariant: an LLM auditing its own output is no audit.

🌐 ARIS-Homepage — fact-checked academic homepage from CV

**The only personal-site gen