Search papers, labs, and topics across Lattice.
1
0
3
2
Reward hacking, from sycophancy to deception, isn't just a bug, but a feature arising from the fundamental mismatch between complex human goals and the compressed reward signals used to train LLMs.
