Michelangelo (Latent Structure Queries)

What it measures

Evaluates long-context synthesis beyond surface retrieval via CodeRepoQA, BugLocalization, and IDK abstention.

Reading the measurement

Length is a test condition, not a capability score. Locate the relevant evidence, then examine whether the model resolves conflicts and combines details correctly. Keep document extraction, input length and answer grading fixed.

Published evaluation

Chart-derived readings of the cumulative-average score at the final 128K-region point, rounded to 0.01 on the paper’s 0–1 scale. Values were recovered from the published vector curves using the labeled linear y axis and checked against the rendered figure; they are approximate readings, not an original numeric result table. Cumulative average through the length range is not accuracy only on 128K questions. This covers MRCR, one of Michelangelo’s three distinct task families, and does not synthesize a suite-wide score.

Where this fits in an evaluation suite

A large context window is a capacity specification. Successful retrieval and cross-document reasoning are observed capabilities. Test both at the actual lengths your workflow needs.

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

Chart-derived readings of the cumulative-average score at the final 128K-region point, rounded to 0.01 on the paper’s 0–1 scale. Values were recovered from the published vector curves using the labeled linear y axis and checked against the rendered figure; they are approximate readings, not an original numeric result table. Cumulative average through the length range is not accuracy only on 128K questions. This covers MRCR, one of Michelangelo’s three distinct task families, and does not synthesize a suite-wide score.

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.