Search papers, labs, and topics across Lattice.
Affiliation:
4
0
8
Sycophancy emerges early in LLM training and its specific vulnerabilities are largely shaped by post-training, not pretraining, dynamics.
LLM sycophancy is not an intrinsic property but a malleable behavior shaped by subtle variations
Answer order and wording, not moral judgment, drive the yes-no bias in LLMs, revealing a surprising disconnect between model behavior and human reasoning.