A whole congress screened against only its own past, set next to what its reviewers
said about the same abstracts. Decision-support only; a human makes the final call.
What the process flagged
No reviewer ratings exist for this run yet, so the comparison with reviewers is hidden.
Each abstract counts once: first the screening's verdict, then — for those it found novel — the
larger language model's answer to “does it add anything new?”. Click a number to list those abstracts.
Does the process flag what reviewers called “not new”?
Every abstract was rated by several reviewers (usually 3–5). Abstracts are grouped by
how many of those reviewers ticked the “No new information” weakness; each bar is the share flagged
(see “Count as flagged” above). Rising bars mean the two agree.
Likely duplicatePossible overlapLLM: not / only partly new
Where the screener and the reviewers disagree
Each abstract counted once: was it flagged (as set under “Count as flagged”), and did
reviewers call it “not new”? Reviewers count as calling it not new when at least
of them ticked “No new information”. Click a number to list those abstracts below.
Screener plus the LLM check
The second step of the process: a larger language model
() is asked of each abstract the screening did not flag: “does it add
anything new to the earlier work?” Read across: the share caught by the screener, then also counting the
LLM's “nothing new”, then also its “only partly new”. Each abstract counts once. Compare the rows: the
same checks also flag abstracts that no reviewer called “not new” — a wide net that
narrows down what reviewers look at, not a verdict.
How reviewers treated each screener verdict
Reviewer grade is the mean of the reviewers' grades, from 1 (best) to 6 (worst) — a
lower grade is better (accepted abstracts average lower than rejected ones). Mind the
direction: for the novelty score, low means close to earlier work.