ARIS-in-AI-Offer: Bilingual ML Interview Cheat Sheets Built by a Claude Code Workflow

The gap this fills between a textbook and a question bank

Chinese AI campus recruiting, 秋招, compresses a broad ML curriculum into a few weeks of interviews. The usual preparation path is a textbook plus a scattered question list, and the two rarely line up. ARIS-in-AI-Offer addresses that mismatch by fixing one structure across every topic: formula derivations, from-scratch PyTorch code, and 25 high-frequency interview questions stratified into L1 essentials, L2 advanced, and L3 top-tier lab. The README states there are 34 first-party cheat sheets across 7 categories plus 1 community-contributed category, covering foundations, post-training and reasoning, LLM architecture, generative model theory and tokenizers, generation systems for image, video, 3D and diffusion post-training, multimodal, agents, and embodied AI. The intended reader is a candidate preparing for Chinese-language interviews who wants the derivation and the implementation in the same document. It is not a course. Nothing in the material suggests it teaches a topic from zero, and the L1/L2/L3 split assumes you already know which layer you are weak at.

How the ARIS pipeline turns a topic into a rendered cheat sheet

The repository is the output side of a generation workflow, not a hand-written book. According to the README, each cheat sheet is produced by the ARIS /interview-cheatsheet command and then rendered by /render-html into a single-file HTML document. ARIS itself, Auto Research in Sleep, lives in a separate repository, and the README describes this collection as the production output of the same workflow used in academic-research production. The rendering step is where the practical value sits. MathJax renders LaTeX as selectable text rather than screenshots, highlight.js colors the PyTorch blocks, the layout is responsive, and a sticky table of contents handles long documents. Because the result is one HTML file with no backend, a downloaded copy opens offline on a phone, tablet, or laptop. The repository also carries two adjacent outputs built on the same renderer: a CV to fact-checked academic homepage generator, with a live demo at wanshuiyin.github.io, and hand-authored long-form blogs, including a Chinese survey on continuous diffusion language models credited to Ruofeng Yang (SJTU) and described as written through cross-model discussion.

Repository layout and the commands the README gives

The material shows a docs/tutorials directory holding the rendered HTML and a docs/tutorials/code directory holding the runnable scripts. The clearest example is the newest sheet, number 34, Modern Diffusion Post-Training: Flow-GRPO, DGPO, and DiffusionNFT. Its script is code/diffusion_online_rl.py, and the README describes it as analytic checks of marginal preservation, the NFT update sign, and DGPO's balanced weights. That is a useful pattern: the script does not train a model, it verifies the algebra the sheet derives. The rendered file for that topic is modern_diffusion_post_training_tutorial.html. The README does not give a clone command, a dependency list, or an install step for these scripts, so treat the code directory as a set of standalone files you read and run individually rather than a package you install. The generation commands themselves, /interview-cheatsheet and /render-html, belong to the ARIS toolchain in the other repository; this repository does not document them as local entry points.

The one structural idea worth copying: reward entry into the gradient

The newest sheet is organised around a single axis rather than a list of methods. The README says the three 2025 online-RL methods for flow-matching text-to-image are separated by how the reward enters the gradient: a per-step policy gradient on a marginal-preserving SDE for Flow-GRPO, group-level DPO with log Z cancellation for DGPO, and a reflected two-branch regression on the forward process for DiffusionNFT. The changelog also notes that Flow-GRPO was moved out of the earlier Diffusion Post-Training sheet and that its reverse-SDE sign and closed-form KL were corrected during the move. That correction is the most informative detail in the material. It tells you the collection is revised rather than frozen, and it tells you revisions happen at the derivation level, not just at the formatting level. It also means a sheet you downloaded earlier can be wrong in a way that a later version fixes, which is an argument for re-reading the What's New section before an interview rather than trusting a cached copy.

Where the collection is thin or hard to use

The README is candid about at least one failure mode. The entry for 2026-07-31 describes an interview-scannability overhaul for sheets 30 through 33 plus a collection-wide table-scroll fix, and it quotes reader feedback that the material was comprehensive but unscannable. A cheat sheet that cannot be skimmed under time pressure is doing half its job, and the fix arrived only after the complaint. Two other limits are visible. First, the depth is Chinese-first: the README leads with the Chinese version and describes each cheat sheet as a long-form Chinese tutorial, so a reader who needs English-depth explanations will find the bilingual framing thinner than it sounds. Second, there are no retrieved releases, so there is no version number to pin, no changelog artefact outside the README, and no way to declare which revision of a sheet you studied. The generation approach has a related cost: a sheet is only as good as the review rounds described in the changelog, and the material reports design review plus three review rounds for sheet 34 without describing what those rounds check for in general.

What to compare it against, and how the approaches differ

The obvious alternative is a general interview preparation repository such as Deep-ML-style problem sets or a question bank like the machine learning interview collections on GitHub. The difference is in the unit of content. A question bank gives you prompts and answers, and you supply the derivation and the code yourself. ARIS-in-AI-Offer gives you the derivation, the code, and the questions bundled per topic, which is more work per sheet and fewer sheets overall. A second alternative is a conventional textbook plus a paper reading list, which is deeper and slower and gives you no rendering guarantees. The trade is coverage against integration. If your weakness is recall under time pressure, a question bank is the better tool. If your weakness is explaining why a method works and writing it from memory, the three-pillar structure is the closer fit. A third comparison point is the ARIS main repository itself, which the README describes as shipping 74 or more research skills across 7 or more platforms; this repository is a consumer of that toolchain's output rather than a substitute for it.

Licence, maintenance, and what a fork costs you

The repository is MIT licensed, which permits reuse and modification with the licence and copyright notice retained. That matters for a candidate who wants to fork the HTML, restyle it, or translate a sheet, and it matters for anyone embedding the material in an internal study group. It does not settle the provenance of the underlying figures or quoted paper content, and the README does not address third-party assets, so check individual files before republishing anything beyond personal use. On maintenance, the changelog shows activity concentrated in bursts: a large entry on 2026-07-31 covering sheets 30 through 33, and a new sheet 34 on 2026-09-10. That is a fast-moving topic list, and the cost of keeping a fork current is a manual diff per sheet, since there are no releases to track. The cheapest upgrade path is to watch the What's New section and re-download only the HTML files whose topics changed, rather than mirroring the whole docs/tutorials tree.

Editorial conclusion

Adopt ARIS-in-AI-Offer if you need Chinese-language, offline-readable preparation material that pairs formula derivations with runnable PyTorch and a stratified question bank, and if you are willing to treat it as one input among several. Do not adopt it if you need English-depth content, a maintained question bank with versioned releases, or answers you cannot verify against primary sources; the repository has no releases and the README itself notes a reader complaint that the sheets were comprehensive but hard to scan. Before relying on any sheet, open the specific HTML file for your topic, check the date in What's New against the paper it cites, and run the accompanying script under docs/tutorials/code/ to confirm the derivation matches the code.

Official sources

1. Issues
2. License: MIT
3. README
4. wanshuiyin/ARIS-in-AI-Offer on GitHub