Ablation V2 of the Stage3 Qwen2.5-VL framework: lookahead via InfoNCE to a
vanilla V-JEPA2 ViT-L target (mean-pooled), without domain disentanglement
(no SIGReg, no Head B).
Evaluated on the LIBERO-Spatial island viewpoint protocol from
Xing et al. 2025 (CoRL) "Shortcut Learning in Generalist Robot Policies".
Backbone: Qwen/Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj)
Single… See the full description on the dataset page:
https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v2-no-disent-infonce.
