Search papers, labs, and topics across Lattice.
1
0
3
Complex, multi-step instructions can cause LLMs to completely ignore question content and instead rely on positional shortcuts when asked to underperform, revealing a critical vulnerability in adversarial evaluation.
