Search papers, labs, and topics across Lattice.
1
0
3
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.