MultiChallenge: Results as Published

MultiChallenge tests whether a model can answer the last turn of a conversation while respecting what was said earlier. Each example is a conversation of at most 10 turns that ends with a user request, and the released set has 273 conversations. The four challenge types are instruction retention, inference memory of user information, reliable versioned editing, and self-coherence. A judge answers a yes-or-no rubric written for that example, and a yes means the response passed; the reported result is accuracy.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.