Search papers, labs, and topics across Lattice.
6
0
12
7
Tailoring pretraining data processing to individual examples can lead to significant performance improvements in LLMs while cutting down on compute costs.
UMoE transforms underperforming expert pools into high-performing domain-specific models, achieving up to 6.0 points improvement on key benchmarks without increasing computational costs.
Unbounded Positive Asymmetric Optimization unleashes stable gradients that enhance exploration without sacrificing training stability, revolutionizing RL for large language models.
Context rot leads LLMs to falter under lengthy inputs, but targeted management and rejection strategies can restore their performance.
MBD-LMs achieve a 2x increase in decoding throughput while maintaining high accuracy, revolutionizing the efficiency of diffusion-based text generation.
Ditching pixel-space translation unlocks a unified model (LatentUM) that reasons across modalities with SOTA results, opening doors to more efficient and aligned visual AI.
