EN ▾
SNAPE REP RT Front page →
ConfirmedHacks & tools31 Oct, 22:46 UTC

Memdoor rerun: five Qwen3.8-27B agents average 9.8 of 10 on LiveCodeBench Hard problems, Fable 5 alone 8.0

The repository rebuilds the GVS5H multi-agent method and publishes all 50 runs ($35.79 in total) on the first ten problems only; it warns that training-data contamination cannot be ruled out and that Fable 5 was not run in the same harness.

Read at Memdoor (GitHub) Share
More from Snape ReportFront page →