The Obsessed Encoder — Enigma on How Predictable Features Hijack Self-Supervised Models
— by Vivax
Enigma, a robotics startup, has published "The Obsessed Encoder" — an empirical study of a failure mode it calls latent-space misallocation in JEPA-style…
Enigma, a robotics startup building foundation models for manipulation, has published The Obsessed Encoder — an empirical study of a failure mode it calls latent-space misallocation in JEPA-style self-supervised learning. Because these models are trained to predict, they prefer what is predictable: an encoder with a fixed latent budget quietly spends that budget on easy, predictable features while starving the hard ones a downstream task actually needs. Training looks better than ever while the representation rots.
The core experiment adds a faint, image-seeded 12-bit watermark tile to DINOv3's training images, with a random pattern of identical pixel magnitude as control. The watermarked run's prediction loss drops below the clean baseline while its classification probe decays toward chance and representations cluster by watermark pattern instead of image content; the control stays healthy — predictability, not corruption, breaks the model. The same signature appears in LeJEPA and in the LeWorldModel robotics stack, where a 5×5 corner square or a randomized goal pose holds loss below baseline while planning success never rises above chance, and the anti-collapse regularizer reads normal because the encoder folds a low-dimensional sheet until it looks Gaussian.
The lesson travels: hospital data is saturated with predictable non-signal — scanner watermarks, template text, institutional formatting. The health check for any world model, clinical ones included, has to live outside the loss function, in grounded probes tied to the quantities you actually care about.