LuminaBench
Reasoning benchmark / Arena

Arena Document

Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.

Style-controlled preferences for document understanding systems.

Version Overall · Style controlledVerified 2 November 202644 published resultsOfficial source

Arena Document · Lumina Score

Style-controlled overall · Source date tag 2026-09-13

LuminaBench
Show
20 of 38 models · best variant per model
03006009001,2001,124Claude Fable 5 · Document1,112Claude Opus 4.6 · DocumentHigh1,109Claude Opus 4 7 · Document1,104Claude Fable 5.1 · DocumentMax1,095Claude Opus 5 · DocumentHigh1,091GPT-5.6 Sol · DocumentxHigh1,083Claude Opus 4.8 · DocumentHigh1,073GPT-5.5 · Document1,062Muse Spark 1.1 · Document1,057Claude Sonnet 4.6 · Document1,053GPT-5.6 Terra · DocumentxHigh1,050Claude Sonnet 5 · DocumentHigh1,040GPT-6 Astra · DocumentMax1,032GPT-5.4 · Document1,025Muse Spark 1.3 · DocumentMax1,024Muse Spark · Document1,015Claude Opus 4 5 20251101 · Document1,014Grok 4.5 · Document1,007GPT-5.6 Luna · DocumentxHigh1,004grok 4.6 high · Document

Publisher preference ratings, bounds and vote counts retained. Missing metrics are omitted from their graph; quality scores remain available.

Results table

Method & coverage

Arena document_style_control, style-controlled overall only. Publisher date tag 2026-09-13; individual run dates and exact tool builds are not published. Search/grounding/document setup remains part of the evaluated system. Separate boards are never combined.

Preference ratings with exact publisher bounds, variance and vote counts. The dataset does not specify the interval level. No resource telemetry is published.

How the Lumina conversion works

1,000 + 75 × ((original score − 1459.828444203246) / 23.706298461458886). The result is rounded once. Scores may tie.

Frozen initial valid overall cohort; one arithmetic mean per exact evaluated release/system, equal weight across releases rather than effort rows. Unbounded rating/IPS scales use a broad finite numeric guard, not percentage bounds or sign clamping. Subsequent refreshes never refit this calibration. Conversion is not an estimate or combined index.

Calibration
arena-document-style-control-z75-v1 · experimental
Reference mean
1459.828444203246
Population SD
23.706298461458886
Coverage
Standalone benchmark.
Identity, exclusions & attribution

Exact publisher labels and release qualifiers are retained. Only explicit native effort tokens create reasoning variants; token budgets remain separate.

Agent orchestrator, routed, dated and harness-qualified systems are not automatically equated with naked base models.

0 ambiguous repeated records are retained in original.json but excluded from ranking.

Official dataset · Pinned source revision · CC BY 4.0

Publisher terms. Scores and resource figures are attributed to the original publisher.