TMT is a runtime backdoor detector for vision‑language‑action policies. It uses a Token Manifold branch and a latent Transition branch trained on benign rollouts. The authors also propose policy purification through self‑distillation using a frozen copy of the backdoored policy. In evaluation TMT outperforms ten baselines on unseen tasks across three VLA backdoor attacks.
Key points
Detector: two branches assess input‑token structure and prediction errors in adjacent‑layer latent dynamics
Purification: frozen copy provides benign‑input actions to supervise a student on paired benign and triggered observations without a separate clean reference policy
Evaluation: compared with ten baselines, achieves state‑of‑the‑art backdoor detection performance on unseen tasks across three VLA backdoor attacks
Sources
The Qeios Podcast episode on mindfulness training in Brazil
The Qeios Podcast released a new episode discussing mindfulness training as public health capacity in Brazil. It states that training instructors for the whole Unified Health System would cost under 0.1% of its budget. The episode is available on Spotify.
Key points
Cost claim: training instructors would cost under 0.1% of the Unified Health System budget
Availability: the episode can be listened to on Spotify
Sources
- Original post
- Linked resource - Linked in the post