Tencent made the HyperLogic benchmark repository and dataset public on October 28, 2026. It tests Chinese logical reasoning across three tiers, with answers derived by running formal models through logic solvers rather than written by hand.
The release includes 69 Seed items, 59 Base items and 30 Hard items, plus scoring tools and a sandbox setup for testing agents on Hard problems. The other 70% remains private. In the team’s tests of seven models, none scored above 16% on Hard by direct answering; using a code sandbox raised scores, but the top result was 57% with the logic library.
Why it matters
Researchers can use the released problems and tools to test reasoning systems, but the private 70% is needed for full-benchmark evaluation and to limit exposure of test items.
