You're extracting research methodology details from academic papers. Some papers describe methods inline in narrative paragraphs ("We collected 1,200 survey responses over six weeks…"); others use a structured "Methodology" section with subheadings; others scatter relevant details across multiple sections.
Your extraction quality is inconsistent across these formats.
Which intervention will most improve it?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Few-shot examples are particularly effective for handling format variation because they demonstrate to the model how to extract from different structural patterns. The model learns: "given a narrative-style paper, here's what extraction looks like; given a structured paper, here's what extraction looks like." This is one of few-shot prompting's strongest use cases — teaching the model to generalize across surface variations to find the underlying information.
Converting papers to uniform Markdown is a major undertaking that doesn't necessarily preserve the structural cues the model uses. And the same kinds of variation (scattered vs. structured information) will exist in the converted output. You're shifting the problem, not solving it.
Larger models help with many things, but format variation is largely a prompting problem, not a capability problem. Few-shot examples cost less and address the issue more directly.
Running three extractions and taking the consensus is expensive and addresses output variance, not the underlying format-handling problem. If the model handles narrative-style papers poorly, three runs will probably all be wrong in the same way.