Search papers, labs, and topics across Lattice.
4
0
8
5
TRACE transforms user corrections into enforceable rules, slashing preference violations from 100% to as low as 2% in critical coding tasks.
Soft-gating with an "advisor" model can steer LLMs to be safer and more useful, reducing over-refusal without sacrificing detection accuracy.
93% of "reasoning steps" identified by keyword matching are actually noise, but a simple stability filter and content-subspace projection can boost steering vector performance by 5-6% and enable cross-model transfer.
LLMs can parrot numerical shortcuts, but they fundamentally lack the human-like "number sense" to know when and why those shortcuts actually work.
