Robot AI models will cheerfully ignore what you tell them — and researchers have a name for it now.
A team studying Vision-Language-Action models — the architecture behind robots that take natural language orders and translate them into physical movement — found that three leading systems, Pi0, Pi0.5, and OpenVLA OFT, routinely complete tasks even when the instruction is logically impossible given the scene. The researchers call this "linguistic blindness": the models default to visually plausible actions learned during training, overriding whatever the language input actually says. To expose the problem systematically, they built ICBench, a benchmark derived from the LIBERO dataset that injects deliberate instruction contradictions while leaving the visual environment unchanged, forcing models to choose between what they see and what they're told.
The fix, Instruction-Guided Attention Recalibration (IGAR), rebalances how a model's attention mechanism weights language versus vision at inference time — no retraining, no architectural surgery. That matters because the current path to better robot reliability has mostly run through expensive fine-tuning or full retraining cycles; a plug-in correction that works on already-deployed models is a meaningfully different kind of solution. Tested across 30 LIBERO tasks and on a physical Franka robotic arm, IGAR cut erroneous executions under contradictory instructions without degrading baseline performance.
Linguistic blindness is a sharper, more specific failure than the vague "hallucination" label that gets applied to almost everything in generative AI right now — and it points to a structural problem worth watching as VLA models move from research demos toward warehouse floors and operating rooms.