Search papers, labs, and topics across Lattice.
3
0
7
2
D-LLM hallucinations aren't uniformly distributed across tokens or denoising steps; DynHD exploits this to boost detection accuracy and efficiency.
Forget auxiliary losses and fixed expert capacity: Expert Threshold routing dynamically allocates computation in MoEs and balances expert load, all while boosting data efficiency by 1.6x.
Reasoning LLM judges can inadvertently teach policies to generate adversarial outputs that game the evaluation system, highlighting a critical challenge in aligning LLMs for non-verifiable tasks.
