NeMo-WM · Live Research Dashboard · Sprint 9 Final
Neuromodulated
World Models
All experiments complete · k_ctx=32 · AUROC 0.9997 · 26K params · 8W inference
Session start — EXPLOIT regime
No-VLM AUROC
0.9997
k_ctx=32 · hard neg k≥32
Parameters
26,561
No GPS · No vision
Neg dist (unit sphere)
1.035
+3.5% past sphere
VLM contribution
+0.000
Full = No-VLM exactly
Superlinear scaling confirmed through k=32
Hasselmo 1999
Acetylcholine temporal precision theory: low-ACh = broad temporal integration = optimal for slow 4Hz outdoor navigation. Each doubling of k_ctx gives superlinear AUROC gain. At k=16→32 negatives cross the unit sphere boundary.
V-JEPA 2 Scale Comparison
Negative distance across k
No VLM AUROC
0.9997
proprio only · ep11
Full AUROC
0.9997
VLM + proprio · identical
VLM only AUROC
0.8905
DINOv2 StudentEncoder
VLM contribution
Δ 0.000
Zero additional power
No VLM — proprio only ★
0.9997
1.035
+1.001
Full (VLM + proprio)
0.9997
0.531
+0.513
No proprio (VLM only)
0.8905
0.013
+0.012
0.9997
PROPRIO ONLY · 26K
0.9997
FULL (VLM + PROPRIO)
+0.000
VLM NET CONTRIBUTION
AIM Latent Probe — N=1,752 · 150 files · 16 particles · ep28
8 signals confirmed
Full baseline k=4
0.9577
hard neg no GPS
HD lesion drop
−0.227
Primary signal (Moser 2008)
Vel lesion drop
−0.009
Not primary
HD:vel ratio
25:1
∞:1 at k_pos=1
Path Integration Channel Ablation
Blind AUROC
0.9997
No camera · No GPS · No light
Sensor cost
~$10
IMU + wheel encoder
Sensor power
3mW
Total additional
Camera required
NO
Physics-only pathway
Sensor Requirements
VEL
Wheel encoder
~1mW · <$5
IMU
Gyroscope + magnetometer
~2mW · <$5
CTL
Wheel current draw
0mW · $0
Total: ~$10 · 3mW · No camera
Where It Enables Deployment
Biology → NeMo-WM mapping
0.9997
PROPRIO ONLY · NO VISION
0.9997
FULL (VLM + PROPRIO)
NeMo-WM · github.com/taylorjohn · Sprint 9 complete