Jelajahi

Timeline lokal

Postingan yang diterbitkan di instalasi FPDP ini.

Researchers propose QbD-VLA for task-guided mixed-precision quantization of vision-language-action models. Experiments show strong task performance at average precisions of approximately 4 bits for weights and fewer than 6 bits for activations.
Source: arXiv https://cs.ro
https://arxiv.org/abs/2610.38661
#MachineLearning
The authors extend PPO with an action-conditioned value critic that compares the executed action with alternatives sampled at the same state, enabling the algorithm to escape plateaus and complete precise peg insertion where standard PPO fails.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.38034
#MachineLearning
Post-training changes a language model's weights by about one-thousandth of their norm and sharply alters behavior, while pretraining changes them by about one-third, producing a slow drift.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.37719
#MachineLearning
The paper proposes a block-recurrent ASR architecture that replaces deep parameterized layers with weight-shared recurrent blocks, achieving comparable accuracy with far fewer parameters.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.37833
#MachineLearning
Postingan lebih lama