Yuchen Cui

A single-word instruction change can tank VLA performance by 98%, but a simple LLM-based rephrasing pipeline boosts success by up to 27% without retraining.

Meta-Ctrl cuts constrained decoding memory from over 107TB to under 2GB while achieving higher subgoal success rates than GPT-4.

Cloning synthetic demonstrations alone caps performance, but combining it with reinforcement learning unlocks superior real-world manipulation capabilities.