Search papers, labs, and topics across Lattice.
Bar-Ilan university
3
0
6
18
LLMs' supposed introspection is likely just clever pattern matching on inputs, not genuine self-awareness of their internal states.
Discrete diffusion models can solve lookahead planning tasks more efficiently than autoregressive models by exploiting the asymmetry between forward and reverse generation.
LLMs encode truthfulness along a spectrum from general principles to narrow domain expertise, and exploiting domain-specific truth representations is key to steering model behavior.
