VLA papers make execution structure explicit, while tougher tests expose reasoning gaps

Long-horizon robot control is getting built as an execution system, not just a bigger policy. HELM is the clearest example. It adds episodic memory, a learned failure check before execution, and rollback-based recovery. On LIBERO-LONG, that lifts OpenVLA from 58.4% to 81.5%. The paper also shows that simply extending context helps much less, reaching 63.8% at H=32. ST-π pushes the same agenda at the planning layer. It predicts chunk-level prompts with semantic intent, target location, and duration, then hands each chunk to a step-level action generator. The reported STAR dataset also makes this structure explicit with 30 real-world tasks, 50 demonstrations per task, and about 300k interaction steps.

Reasoning is being trained into VLA models more directly, but the evaluation signal says current gains are still fragile. ReFineVLA adds teacher-written rationales during fine-tuning and reports better SimplerEnv results, including 47.7% average success on WidowX, 68.8% on variant aggregation, and 76.6% on visual matching. The same day also brings a harder check on what these models actually learn. BeTTER applies controlled interventions for grounding, recomposition, and temporal stress. Under those tests, strong models collapse on unseen subgoal composition: pi_0.5 falls to 5.0% on B->C, GR00T-N1.6 to 15.0%, and Being-H0.5 to 0.0%. That makes the current picture clear: reasoning supervision can improve benchmark performance, but robustness to changed task structure is still a weak point.

Spatial modeling is another strong thread, with papers trying to give VLA policies better control over geometry and action precision. OmniVLA-RL combines a three-expert transformer with explicit 3D scene features and online reinforcement learning. The reported headline is 97.6% average success on LIBERO, though the available excerpt does not include the full comparison table. ST-π also treats space and time as first-class inputs by predicting spatial targets and temporal durations for each sub-task. Across both papers, the emphasis is concrete: improve grasping, placement, and multi-stage control by making spatial structure visible inside the action pipeline rather than leaving it buried in fused embeddings.