Finding and tracing the literature

You cannot build on work you cannot find, and you cannot find it by reading everything. Finding the literature is a skill with its own moves: constructing a search, following citations, judging what a source is worth, and deciding, fast, what to actually read. Done well, an afternoon of searching saves months of reinventing what already exists. Done badly, you miss the one paper that would have changed your question.

A search is a claim about what you want, written in the database's language. Two levers do most of the work. Boolean logic: OR joins synonyms and widens recall (dropout OR "stochastic regularization"); AND combines concepts and raises precision (dropout AND overfitting). Controlled vocabulary, a database's indexed terms, catches papers that use words you didn't think of. Start broad, read what comes back, and tighten. A search that returns three papers is usually too narrow, not admirably precise.

Once you have one good paper, you rarely need to search again; you follow its citations. Backward chaining reads the paper's own reference list, moving toward older, more foundational work. Forward chaining reads the newer papers that cite it, moving toward developments, replications, and rebuttals. Alternating the two, snowballing, walks you through a whole line of research and is often faster and more complete than keyword search alone.

Not every source carries the same weight. A preprint is public but has not yet passed journal peer review, often current and valuable, but unvetted, so read it critically. A peer-reviewed article has been filtered by independent experts, which lowers (but does not eliminate) error. A predatory outlet charges a fee and guarantees fast publication with fake or absent review; its tell is pay-to-publish, not the fee alone; legitimate open-access journals charge processing fees too. Knowing which you are holding changes how much you should trust it.

You will find more than you can read, so reading is triage: spend the cheapest signals first. Screen the title, then the abstract, then the figures, then the methods, and only then the full text, discarding most candidates before you ever open the PDF. The same discipline underlies formal evidence synthesis, where explicit screening at title/abstract stage is what keeps a review tractable and honest.