disentangled-vla/

libero-island-ablation-v8-sigreg-low

Ablation V8 of the Stage3 Qwen2.5-VL framework: identical to V3 (single AttentiveLatentHead A + L1 + SIGReg with our V-JEPA domain branch) BUT sigreg_weight reduced from 0.1 (V3) to 0.00025. In V3 the SIGReg contribution (≈2.1) dominated the action L1 (≈0.13) at 16:1, suppressing the action signal and collapsing in-domain performance (36%). V8 reverses the balance — action contribution (≈0.13) is ~25× the SIGReg… See the full description on the dataset page: https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v8-sigreg-low.