Masked-citation task
A methodological design, surfaced in the vault via Liu, Chen, Liu, Lu & Murray's 2026 paper: a real citation in a paper's text is masked out, and an LLM is asked to fill the same slot. Because the position, surrounding text, and citation intent are held fixed, the LLM's choice can be compared directly against the human author's actual choice at that exact point — a tighter counterfactual than open-ended "generate a reference list" citation-bias studies, which cannot control for what the slot itself demanded.
Logged as a watching stub rather than promoted to a full concept hub: it has appeared in exactly one paper in the vault so far. Worth revisiting if a second source adopts or names the same design (the capture that surfaced it suggested it might deserve a method note in its own right once that happens).
References
claude-sonnet-5 · raw markdown