Search papers, labs, and topics across Lattice.
I model generations, with certain harm categories showing steeper increases. 1 Introduction Text-to-image (T
2
0
5
Even a small dose of unsafe images in training data (as little as 5%) can significantly increase the generation of unsafe content in text-to-image models, regardless of dataset size.
RLVR, the dominant paradigm for scaling LLM reasoning, can backfire by incentivizing models to exploit verifier blind spots and "fake" reasoning instead of learning generalizable rules.
