A Practitioner's Primer on RSI

Who This Is For

This assumes you've either read the two research pieces above or plan to — they're the map of what exists. This piece assumes you now want to actually touch the field: run something, break something, modify something, and eventually publish something. It's written for a researcher or strong engineer with general deep learning background who hasn't specifically worked in RSI yet, not for someone starting from zero in machine learning.

What to Read and Understand First

There is no single official RSI course or institutional syllabus — this was searched for specifically and not found, so treat any claim of one with suspicion. What exists instead is a small set of maps and one influential framing post:

Hands-On: What to Actually Run

These are real, found-directly repos, not inferred ones. "Official" means the paper's own authors published it; "community" means someone else reimplemented it and fidelity to the original paper is unverified.

Start with FunSearch or AutoML-Zero if you want something that runs in Colab or on a laptop within an afternoon. Start with SICA or DGM if you specifically want to work with the LLM-agent-self-modification pattern that most 2026 techniques build on.

What to Implement or Extend

The fastest path to a real contribution isn't a new architecture from scratch — it's picking one of the repos above and changing exactly one component, then measuring what happens. Three concrete starting projects:

Designing New Experiments

The July 2026 field-mapping survey (arXiv:2607.07663, covered in the 2026 update) frames the literature along two axes: what the system improves (behavior, policy, evaluator, or the research process itself) and degree of loop closure (human-in-the-loop through fully closed). A second, independently-surfaced finding worth building experiment design around: self-improvement strength tracks the strength of external verification — in practice, the quality of your benchmark or verifier matters more to your result than the sophistication of your agent.

In practice, most publishable 2026 work falls into one of three shapes: benchmark-driven (pick ARC-Bench, MLE-bench, or RE-Bench and beat a reported score), ablation-driven (take an open technique and change one component, as above), or new-domain-driven (apply an existing technique somewhere it hasn't been tried). No standalone "research methodology for RSI" paper was found distinct from the survey papers themselves — this framing is this article's synthesis of what the survey and the repos' own structure imply, not a quote from a single source.

Resources You'll Actually Need

Most of this field is LLM-API-bound, not GPU-bound. The exceptions: METR's RE-Bench sandbox runs on an NVIDIA L4 GPU per agent run, and full-scale AutoML-Zero-style search historically used large CPU clusters — a toy reproduction of either runs on far less.

Capturing and Publishing Your Work

Keeping Up With the Field

At roughly 500 relevant papers per quarter, reading everything isn't realistic. What's actually verifiable as a tracking mechanism:

Where to Find the Field

⚠️ Confidence Notes and Gaps

🔗 Full Reference List