Satori · autoresearch
SATORI  ·  MoE autofrontier ten primitives · one crown-gated frontier each · crowned only on real gains
█████  █████  █████  █████  █████  █████
█      █   █    █    █   █  █   █    █
█████  █████    █    █   █  █████    █
    █  █   █    █    █   █  █  █      █
█████  █   █    █    █████  █   █  █████
the fleet · primitives as experts
Each vision primitive is its own crown-gated frontier: an agent proposes and codes angles, and a gain counts only if it beats the reigning king on a held-out split it never trained on. One board, one metric per primitive — never compared across.
primitives run
10 / 10
crowned
5
crowns total
8
beat sota bar
5
hard tracks
2

Each row is scored on its own metric; the track runs from genesis (○) to the current king (●), with the specialist SOTA as a dashed ◆ bar — shown only where the benchmark+metric match ours. Where a specialist reports a different metric or no standard bench exists, the row says so rather than draw a false bar. Positions self-normalized per metric; lengths are within-row progress, not cross-row comparisons.

descriptionCOCO-Cap · CIDEr
0.401.42+1.02sota 1.45
countingCountBenchQA · acc
0.790.98+.20sota 0.94
detectionCPPE-5 · F1@.5
0.560.77+.21bar: F1≠mAP
groundingRefCOCO · Acc@.5
0.850.92+.07sota 0.90 · beat
ocrDocVQA · ANLS
0.943holdssota 0.95 · −.4pt
reasoningA-OKVQA · acc
0.900holdssota 0.90 · tie
behaviourUCF101 · top-1
0.8750.980+.105sota ~0.98 · tie
spatialVSR · acc
0.840beatsota 0.78 · beat
pointingPixMo · point-acc
0.660.790+.12sota 0.79 · met
segmentationRefCOCOg · cIoU
64.166.5+2.4sota ~78.7
○ genesis● king · crowned ● king · held (grey)★ crown ◆ specialist sota— the climb
PrimitiveBench · metric GenesisKingΔSOTA GateWhy
countingCountBenchQA · acc0.7870.982+.1950.936♛ crownedtargeted_data — LCB +0.102 ≫ δ; the fleet’s biggest single jump
descriptionCOCO-Cap · CIDEr0.401.42+1.021.45♛ crownedstrong_sft crowned @ fleet δ0.02 (LCB +.029, holdout +.146) — tops generalist VLMs, nears BLIP-2 1.45; product stays rich
detectionCPPE-5 · F1@.50.5640.770+.206
F1≠mAP
♛ crownedstrong_sft re-crowned genesis→0.770 (LCB +.200); tops prior 0.743; no fitting generalist to set an F1@.5 bar
groundingRefCOCO · Acc@.50.8500.920+.0700.900beats barsame-harness re-baseline: 72B teacher scores 0.900 here (paper 0.927 doesn’t reproduce); our 8B 0.920 beats it +2.0pt. Coord-convention artifact caught (72B would false-score 0.0 naively)
reasoningA-OKVQA · acc0.9000.9000.900ties barsame-harness re-baseline: 72B teacher = 0.900, identical to our 8B (paper 0.928 doesn’t reproduce). Wall was a paper artifact
ocrDocVQA · ANLS0.9430.947near-tiesame-harness re-baseline: 72B teacher 0.947 vs our 8B 0.943 = 0.4pt (within noise); paper 0.964 doesn’t reproduce. Only primitive with real (tiny) headroom → distill target
behaviourUCF101 · top-10.8750.980+.105~0.98♛ crownedbuilt a Satori video expert: 0.875→0.9796 TTA (windowed-training fix, full split-1 test, no leakage); +10.5pt, a statistical TIE with the 0.98 bar (2 clips short)
spatialVSR · acc0.8400.840~0.78beats bar0.840 at genesis already clears the best-VLM VSR line (~0.775) — born above bar, no climb needed
pointingPixMo-Points · point-acc0.660.790+.12~0.79bar metdedicated aug build carried holdout 0.66→0.790 (public 0.795) — both meet the ~0.79 Molmo bar, split-consistent +.12; crown gate held on n=200 public underpower, not a null
segmentationRefCOCOg · cIoU64.166.5+2.4~78.7decoder builttrained text-conditioned SAM2 decoder: box→SAM2 64.1→66.5; frozen-encoder wall ~67. VLM-in-the-loop tested (SEG-token → SAM2, H100): frozen-encoder near-parity 0.61, but UNFREEZING regressed to 0.45 (overfit at RefCOCOg scale) — hypothesis disconfirmed, 66.5 holds. 78.7 needs far more mask data, not encoder unfreeze
Satori · autoresearch — a pure function of each primitive’s crown-gate trail · every row scored on its own held-out split, never compared across metrics · king checkpoints backed up locally.