HLE-Rolling

What it measures

Continuously refreshed dynamic split of Humanity's Last Exam to prevent saturation and leakage.

Reading the measurement

Study what information the model receives and which inference the answer requires. A question may mix recall, visual perception, deduction and search. Use task breakdowns to find the mechanism behind a headline gain.

Published evaluation

Recomputed from the author-released correctness flags for 649 unique HLE-Rolling questions and 50 generated responses per question. The source card describes a selected 649/650-question subset; the released file contains exactly 649. This is not a result on the full rolling dataset, and the source does not pin the upstream rolling revision. Correctness uses the release’s Qwen3-Next/Python answer extraction, not the CAIS leaderboard judge. Mean response correctness and observed any-of-50 success have different denominators and remain separate.

Where this fits in an evaluation suite

Ask what the task actually requires: recall, deduction, search, or a mixture. Separate closed-book reasoning from retrieval and tool-assisted problem solving.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Recomputed from the author-released correctness flags for 649 unique HLE-Rolling questions and 50 generated responses per question. The source card describes a selected 649/650-question subset; the released file contains exactly 649. This is not a result on the full rolling dataset, and the source does not pin the upstream rolling revision. Correctness uses the release’s Qwen3-Next/Python answer extraction, not the CAIS leaderboard judge. Mean response correctness and observed any-of-50 success have different denominators and remain separate.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.