Goto

Collaborating Authors

 Large Language Model





Factuality-Aware Alignment for Large Language Models

Neural Information Processing Systems

This makes SFT less factual as it trains on human-labeled data that may be novel to the LLM. Furthermore, reward functions used in standard RL often inadequately capture factuality and favor longer and more detailed responses, which inadvertently promote hallucination.


PowerPM: Foundation Model for Power Systems Shihao Tu

Neural Information Processing Systems

Deep learning models have advanced ETS modeling by effectively capturing sequence dependence. However, learning a generic representation of ETS data for various applications is challenging due to the inherently complex hierarchical structure of ETS data.





Larry Summers to leave positions at Harvard and OpenAI after Epstein emails

The Japan Times

Former U.S. Treasury Secretary Larry Summers says he will step back from all public commitments, adding the move is to allow him to rebuild trust and repair relationships with the people closest to me. Former U.S. Treasury Secretary Larry Summers is stepping down from a teaching post at Harvard University and as a director of one of its business and government schools, a spokesperson said on Wednesday, after Congress released documents showing Summers shared close ties with the late convicted sex offender Jeffrey Epstein. A spokesperson for Summers, Steven Goldberg, said Summers' co-teachers would complete the semester for three ongoing courses. Mr. Summers has decided it's in the best interest of the center for him to go on leave from his role as director as Harvard undertakes its review, he said. Summers, also a former president of Harvard University, is a director of the Mossavar-Rahmani Center for Business and Government at the Harvard Kennedy School. Summers has been under fire since the U.S. House Oversight Committee released documents detailing an ongoing personal correspondence between Summers and Epstein, who died by suicide in a Manhattan prison in 2019 as he faced sex-trafficking charges.


MoV A: Adapting Mixture of Vision Experts to Multimodal Context

Neural Information Processing Systems

We conduct extensive experiments to evaluate the effectiveness of the proposed approach. Without any bells and whistles, MoV A can achieve significant performance gains over current state-of-the-art methods in a wide range of challenging multimodal benchmarks.