Given partial context about a research paper — an abstract, a set of citations, related figures — an LLM can often reconstruct a plausible version of the full paper. This project asks a more precise question: which pieces of context does the model actually rely on to do that, and are different context types substitutes for each other (either one suffices) or complements (both are needed together)?

Answering that has direct implications for how much and what kind of context matters when using LLMs for literature synthesis, research assistance, or paper drafting — and where those systems are likely to hallucinate structure rather than genuinely infer it.