Search papers, labs, and topics across Lattice.
Affiliation:
2
0
4
Extended reasoning modes paradoxically degrade an LLM's ability to assess evidence sufficiency in dynamic diagnosis, driving models to prematurely lock into diagnoses that succeed only by chance.
Reward models can be made more robust to spurious cues like length and sycophancy by explicitly training them to understand the *intent* behind a prompt.
