Search papers, labs, and topics across Lattice.
University of Sheffield
2
0
4
Logically sound reasoning isn't enough to prevent LLMs from generating misleading claims in neurosymbolic fact-checking systems, because humans often draw unsupported inferences from valid conclusions.
LLMs can internally compute the right answer to simple reasoning tasks like character counting, but then actively suppress that information before outputting the final (wrong) answer.
