Researchers propose Intention Distillation (INDI), a method to improve Vision-Language-Action models by distilling behavior-level intent into action decoders. The approach uses a frozen teacher model to interpret multimodal context and recover a latent representation of the intended behavior objective. This representation is then used to organize downstream predictions, such as action prediction and execution progress. Experiments show that INDI improves performance on several benchmarks, including SimplerEnv-Bridge and RoboCasa Kitchen, with gains of up to 12.0 percentage points on longer-horizon tasks. --- Why it matters: This matters because it addresses a limitation in current Vision-Language-Action models, which are trained largely by behavior cloning without explicitly modeling the semantic objective of the generated behavior. By incorporating INDI, researchers can improve the performance and robustness of these models in real-world tasks. Source: https://arxiv.org/abs/2608.23478