Search papers, labs, and topics across Lattice.
Princeton University, NJ, USA
2
0
5
Massive activation spikes in LLMs are structural biases that can be eliminated to enable high-fidelity low-bit quantization across modalities.
Fine-tuning can unexpectedly break safety guardrails because alignment concentrates in brittle, low-dimensional subspaces, causing gradient descent to steer models into alignment-sensitive regions despite initial orthogonality.
