Search papers, labs, and topics across Lattice.
3
0
7
Stop training your captioning models with blunt instruments: ClaimDiff-RL uses verifiable, typed claim differences to expose the faithfulness/coverage tradeoff and enable more balanced, controllable caption generation.
Existing power outage prediction models fail to account for spatial weather patterns, but this new approach leverages graph neural networks and contrastive learning to substantially improve prediction accuracy.
Exact sampling in large-vocabulary decoding can be sped up by 19% simply by fusing it into the LM-head matmul, turning a bandwidth bottleneck into a lightweight epilogue.
