Literature Survey

What this phase owes the paper

Two things: a set of real papers you have actually seen, and an honest account of where your idea sits among them. Everything downstream depends on it. A related-work section written from the model's memory of a field is where invented citations come from.

Doing the survey

Search, then read what came back. Use Semantic Scholar, arXiv, DBLP or CrossRef. Collect titles, authors, years, venues and abstracts from the API response, not from recall.

Record the bibliography as you go, through the API, following citation-discipline. Do not plan to reconstruct citations later from the prose. Reconstruction is where fabrication enters.

Look for the limitation, not just the topic. The useful sentence in a related-work section is "X does A but assumes B, which does not hold here". That requires reading abstracts, and often the limitations paragraph of the papers themselves.

Note what would make your idea unnecessary. If a paper already does what you propose, the honest move is to say so and narrow the contribution. Finding this out at writing time is cheap. Finding it out in review is not.

When the research tool fails

Deep-research APIs fail in ways that do not look like failures. A real case: the provider returned HTTP 200 with a body containing nothing but keep-alive padding, because the upstream model died mid-generation. The naive client raised a JSON parse error, the pipeline logged a warning, and then wrote an entire paper with zero literature behind it. Nobody was told.

Handle it explicitly:

1. Recognise the shapes. An empty body, a whitespace-only body, or a stream of comment lines are all "the provider produced nothing". Say that, rather than surfacing a parser traceback.
2. Retry. These are usually transient. Two retries with backoff.
3. Fall back, do not skip. If deep research stays unavailable, build a smaller survey from direct academic search. It is thinner, but it is real, and the paper still has ground under it.
4. Mark the degradation in the paper and in your report. "Related work was assembled from direct database search rather than a full survey; coverage is partial" is an acceptable sentence. Silence is not.

The principle generalises: a thinner foundation is acceptable, an invisible one is not.

Before moving on

• Every reference came from an API response you read
• You can state, in one sentence, what is new here relative to the closest prior work
• You have noted anything that weakens the novelty claim
• If the survey degraded, the paper says so