Mechanistic Interpretability Fundamentals Certification

The Mechanistic Interpretability Fundamentals Certification is a free advanced-level, 25-minute skill test from Science of Apps. It tests mechanistic interpretability fundamentals — the residual stream view of transformers, attention heads and circuits such as induction heads and indirect object identification, superposition and polysemantic neurons, sparse autoencoders and dictionary learning, activation patching and causal interventions, probing and the logit lens, attribution graphs and circuit tracing, and the limits of interpretability for AI safety — written to the Transformer Circuits thread, ROME and IOI papers for 2026 and 2027. Score 70% or more on its 25 questions to earn a verifiable Mechanistic Interpretability Fundamentals certificate (SOA-MIF). Every question is based on published standards and research — including Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet (2024) (Anthropic, Transformer Circuits), A Mathematical Framework for Transformer Circuits (Elhage et al., Anthropic (2021)), Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small (Wang et al., 2022) (arXiv) — and each answer is explained with its source.

Science of Apps offers a free Mechanistic Interpretability Fundamentals Certification, written for 2026 and 2027. It tests transformer anatomy for interpretability, circuits and algorithms, features, superposition and dictionary learning, causal methods and probing and interpretability for safety and practice with 25 original scenario questions in all, each naming its published source; you get an instant score and a verifiable certificate with a credential ID you can add to LinkedIn.

Read the five skill areas on this page, then take the 25-minute test online — no fee, course or sign-up. You see every answer explained when you finish. Score 70% or more and you get a Mechanistic Interpretability Fundamentals certificate straight away, on screen, as a PDF and by email, with a credential ID anyone can verify.

Yes. The questions are original scenarios written by Science of Apps, and each one is based on a public source you can read: Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet (2024) (Anthropic, Transformer Circuits); A Mathematical Framework for Transformer Circuits (Elhage et al., Anthropic (2021)); Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small (Wang et al., 2022) (arXiv); Locating and Editing Factual Associations in GPT (ROME, Meng et al., 2022) (arXiv); Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (Bricken et al., Anthropic (2023)); Probing Classifiers: Promises, Shortcomings, and Advances (Belinkov, 2022) (arXiv). The source is shown under each question and linked in the answer review. Science of Apps doesn't copy questions from other tests.

Five skill areas with five questions each, drawn from a bank of 10 per area so a retake brings new questions: Transformer Anatomy for Interpretability (Residual stream, Attention heads and MLP layers); Circuits and Algorithms (Induction heads, IOI circuit and Circuit discovery); Features, Superposition and Dictionary Learning (Superposition, Sparse autoencoders and Feature interpretation); Causal Methods and Probing (Activation patching, Ablations and Probes and their limits); Interpretability for Safety and Practice (Safety applications, Limits and Research practice). Every question is an original scenario written to the 2026–2027 standard, with the published source named under it.

Interpretability researcher, AI safety engineer, ML research engineer, Graduate student in machine learning and Technical AI governance analyst. It suits students at leading universities, career switchers and experienced professionals who want quick, verifiable proof of Mechanistic Interpretability Fundamentals knowledge in 2026 and 2027.

CS and AI students at MIT, Stanford, CMU, Berkeley, IITs, Tsinghua and ETH are hired on depth beyond tool use: model internals, systems and hardware. Passing earns a founder-signed Science of Apps certificate with a credential ID and QR code that recruiters can verify.

Mechanistic interpretability tries to reverse-engineer the algorithms a neural network has learned, identifying the features it represents and the circuits of weights and activations that compute its behaviour, rather than only explaining which inputs mattered.

Neurons are often polysemantic because models represent more features than they have dimensions (superposition). Sparse autoencoders learn an overcomplete dictionary of sparsely active directions that are often more interpretable than individual neurons.

Yes. The test is online, in English and free everywhere — people take it from India, the US, the UK, Europe, Africa, Latin America and Asia-Pacific. Your certificate is dated with your own local date and can be verified from anywhere at scienceofapps.com/certification/verify/.

25 questions — 5 from each of 5 skill areas — with a 25-minute limit. You need 70% (18 of 25) to pass and 90% for distinction.

We rate it Advanced: assumes years in the role or deep specialist knowledge. It has 25 scenario questions — 5 from each of 5 skill areas — and a 25-minute limit. You need 70% (18 of 25) to pass and 90% for distinction.

Nothing. There is no fee to take a test, to receive the certificate, to have it emailed or to verify it, and no sign-up.

On our server, not in your browser. The answers never reach the page until you submit, so a certificate can only be earned by passing. Each certificate gets a unique credential ID stored by Science of Apps.

You see your score for each area and every question with the correct answer and why, with the published source behind it. You can retake after 30 minutes, and the retake draws questions you haven't seen yet.

No. Every certificate is dated, so it always shows when you passed. Retake a test any time for a newer date or a higher score.

Yes. After you pass, Add to LinkedIn opens the Licenses & certifications form with the certificate name, Science of Apps as the issuer, the issue date, your credential ID and the verification link filled in. You can also download it as a PDF or PNG, and it's emailed to you as a PDF.

Every certificate has a credential ID and a QR code. Scanning it, or entering the ID at scienceofapps.com/certification/verify/, shows the holder's name, the certification, the score and the issue date from our records.

No. They are Science of Apps' own assessments of applied skill. They aren't issued by a university or standards body, and they work best alongside a portfolio that shows the skills in real work.