Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
2
0
5
0
Model safety isn't about whether adversarial content is seen, but whether it spreads: Claude strips injections at write_memory, while GPT-4o-mini propagates them flawlessly.
Diffusion-based planners in discrete environments like Tetris can actually *degrade* performance with longer planning horizons or naive DQN-based action selection, highlighting the need for careful consideration of uncertainty compounding and critic alignment.