Courses teach what's settled; papers are where the frontier lives — and the field moves monthly. This is a recurring skill, not a phase: learn to find what matters, read it so it sticks (the three-pass method), verify it by reproducing (the part everyone skips), fold it into your own work, and eventually write your own. The end state: take a brand-new arXiv paper, judge it in 10 minutes, reproduce its core claim, and explain it to someone else.
Pick one recent paper close to your interests (RAG, agents, fraud-GNNs, RL trading) and produce: (1) a one-page extraction summary, (2) a small reproduction of one result plus a ~50-line re-implementation of the core idea, (3) a short blog post explaining it simply — teaching is the deepest test — and (4) a paragraph on what you'd change or try next: the seed of your own paper. You've got this skill when a new arXiv paper takes you 10 minutes to judge and a weekend to verify.
Guide: research-skills.md ↗