Search papers, labs, and topics across Lattice.
Affiliation:
5
0
9
Shared reasoning structures can significantly boost performance in continual reinforcement learning, with a novel replay mechanism achieving parity with multitask training.
MoEs, despite their scaling advantages, suffer from a surprising "spectral plasticity loss" in continual RL, but a simple Parseval penalty can recover performance.
Forget hand-crafted reward functions: MVR uses multi-view video and a frozen VLM to automatically shape RL rewards, teaching agents complex motions without getting stuck on static poses.
Momentum *can* accelerate single-pass stochastic gradient descent for generalized linear prediction, resolving a long-standing open question and outperforming variance reduction techniques.
SVD meets random projections: PRAC offers a surprisingly effective recipe for compressing LLM activations, slashing memory by 36% while barely impacting performance.
