Search papers, labs, and topics across Lattice.
School of Electronic and Computer Engineering, Peking University
8
0
11
EFlow's innovative separation of evidence retrieval and reasoning processes leads to substantial improvements in long-video reasoning performance, addressing critical biases in existing frameworks.
Achieving a 4x faster training time and up to 6.4x cost reduction for RL post-training of DiTs by effectively utilizing idle spot GPUs.
SpecGen accelerates GPU kernel optimization by generating candidate kernels in parallel with reasoning, cutting down latency and boosting resource efficiency.
A new 3D representation and BEV alignment framework drastically enhance the generalization of robotic manipulation policies across varied environments and robot types.
DisagFusion unlocks up to 20x higher throughput for diffusion model serving by intelligently splitting the workload across heterogeneous GPUs and dynamically adapting to workload shifts.
Achieve 100% agent recovery correctness with near-zero overhead by intelligently checkpointing only the OS state that actually matters.
Unlocking the potential of compute-in-memory accelerators for LLMs requires carefully navigating a complex dataflow design space, and AccelCIM provides the first systematic framework to do so.
LLMs can translate free-form natural language feedback from users with paralysis into safe and personalized robot control policies, drastically reducing the burden of preference learning.
