Search papers, labs, and topics across Lattice.
Palo Alto Networks
2
0
3
A new fingerprinting method for AI skills achieves 77x compression while maintaining per-component identity, enabling nuanced recognition of skill families and tampering.
Single-neuron interventions can control language model behaviors without collapsing outputs, but only if they respect a defined coherence budget.
