Phi-4-multimodal-instruct

Method version: benchmark-meta-score-v3

The index includes active measurements from reviewed primary sources when at least five models are compared in the same documented comparison slice across dataset, harness, agent, tools, and trial count. Each benchmark series and evaluator contributes one shared setup slice from the newest capture. When one source and series include several explicitly documented reasoning tiers, the highest documented tier (none through max) is selected, never the best result value. The explicit source label “Non-reasoning” is treated as none. Unknown or ambiguous tiers exclude the model from that series. Within each series, percentages and rates are logit-transformed, other units stay unchanged, and all values are then standardized. A one-factor model then estimates one shared strength per model and, per series, how sharply it separates the models. Area scores move away from that strength only as far as the model's own results in the area support. The overall score weights the core areas at 75% and finance, law and health at 25%. It requires at least eight test series from two sources and three areas. The 0 to 100 scale refers to a fixed reference group of 29 language models whose average sits at 50.

Published model-specific parameters can differ. This is not a controlled experiment; comparisons with such differences are marked provisional.

Unmeasured series are not entered as zero. Where a model has no own results in a core area, its overall strength stands in, and the note “estimated without own results in …” names the area. Fewer test series lead to a larger standard error.

The 95% interval follows from the standard error of 200 recalculations with randomly redrawn test families. It describes how precisely the recorded test series determine the value. It is no prediction for future tasks or for later recalculations with other participants.

The source count is not a count of independent experiments.