Search papers, labs, and topics across Lattice.
MBZUAI, The University of Melbourne
6
0
10
Pre-generation signals can predict failure sources in vision-language models, enabling targeted interventions before answers are generated.
Bayesian control outperforms traditional orchestration methods, especially when verification costs are high, by providing a more nuanced understanding of candidate correctness.
Deferring to a larger LLM only when a smaller LLM is uncertain can match the performance of the larger model alone, while slashing inference costs.
LLMs can act as subject matter experts to conduct cost-effective, nuanced interviews, potentially revolutionizing early-stage hiring decisions.
LLMs are shockingly susceptible to generating fake news under jailbreak attacks, especially when it comes to English and U.S.-related topics, exposing a dangerous safety imbalance.
Distilling language models just got more efficient: a new loss function focuses on the long tail of token probabilities, boosting performance without extra compute.
