Track synthesis, materials and catalysis as it is published.
Search papers, labs, and topics across Lattice.
3
0
6
A novel two-stage distillation framework achieves superior generalist reward model performance by first cultivating domain specialists and then integrating them, even surpassing individual teacher capabilities.
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.