Search papers, labs, and topics across Lattice.
Affiliation:
3
0
3
LLMs exhibit a surprising bias towards earlier evidence, which significantly impairs their ability to resolve contextual knowledge conflicts.
LLMs' reasoning chains are surprisingly fragile at logical connectives, but targeted interventions at these "forking points" can dramatically improve accuracy more efficiently than brute-force methods.
LLMs can be steered to better logical reasoning by exploiting a shared, low-dimensional "logical subspace" learned by aligning natural language and symbolic reasoning views.
