The 7,114 Papers That Aren't a Leak

Peer review in the largest AI conferences has quietly become a public forensic exercise. This week an anonymous GitHub repository, xll0328/NIPS26-, posted an HTML file claiming to enumerate every paper accepted to NeurIPS 2026, roughly four weeks before the conference's Sep 24 notification window. The community has already started its own triage.

The 7,114 entries inside xll0328's HTML are the load-bearing detail. If the count tracks NeurIPS 2026's historical accept rate and the ICLR 2026 breach, a real submission-list exposure the community has codified as its standing reference, the file looks like a process-integrity signal worth auditing. If the count and entries are scraped or padded, the same file becomes a public test of how the community catches fabricated lists before the official list drops.

The reading most readers will land on is "leak." The reading underneath is closer to a referee: a community stress-testing its own double-blind process in the open, with a fresh, named precedent for what a real exposure looks like. xll0328's value is not whether it is what it claims to be. It is that the test runs publicly, in plain HTML, with a mechanism anyone can replay.

The repeatable mechanism: an unverified list lands; researchers compare it against the named precedent and the notification window; the count, the anonymization pattern, and the timing do the diagnostic. What remains uncertain is whether NeurIPS treats this as a process-integrity event to investigate or as a distraction to wait out. The community's verdict arrives in four weeks either way.