365 seed logic-and-reasoning problems expanded to 1,095 variants across eight categories, designed to test reasoning that needs only K-12 knowledge; even the top model reached 62.8% at release.
unassessed
| Category | reasoning |
|---|---|
| Subcategory | general logical reasoning independent of specialist domain knowledge, K-12 scope |
| Page status | active |
| Metric | accuracy, graded by a hybrid rule-based and LLM-judge scorer |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 1095 |
| Dataset licence | MIT (Hugging Face dataset card and GitHub repository both state this) |
| Publisher | Meituan (LongCat team), with a co-author at the University of Chinese Academy of Sciences |
General365 tests general reasoning ability -- constraint satisfaction, branching and enumeration, spatial and temporal reasoning, recursive and backtracking logic, resistance to semantic interference, implicit-information inference, optimal-strategy selection, and probability under uncertainty -- while deliberately keeping the knowledge each problem requires at or below K-12 level. The authors built it to isolate reasoning skill from specialist domain knowledge, arguing that benchmarks like competition mathematics or graduate-level science exams conflate the two: a model can fail a hard problem because it lacks the reasoning strategy, or because it lacks the domain fact, and most benchmarks cannot tell which. General365 tries to remove the second failure mode by using only material an educated non-specialist already knows.
Free-form question answering with mixed answer types (a number, a short text answer, or a choice from given options), English, testing 365 hand-written seed problems expanded via controlled variation into 1,095 total problem instances.
No model card in ModelSpec reports this benchmark yet.