Search papers, labs, and topics across Lattice.
Queen's University Belfast
8
0
9
EP-NCO slashes service response times by up to 50% in edge-cloud systems, outpacing traditional optimization methods.
Dynamic rubrics that evolve with the policy can significantly enhance reinforcement learning performance, even without external supervision.
Pruned LLMs can ace multiple-choice tests while failing to generate correct answers, revealing a critical evaluation blind spot in model assessment.
DFlare achieves up to 5.52x speedup in LLM inference by allowing draft layers to independently leverage richer target knowledge, breaking through previous capacity constraints.
Robot RL training can be dramatically sped up (3-10x) by decoupling CPU-based simulation from GPU-based learning, challenging the assumption that GPU-resident physics is essential for efficiency.
LLMs might seem fluent in Chinese-English translation, but HardMTBench reveals their surprising struggles with domain-specific knowledge, exposing weaknesses hidden by standard benchmarks.
Current translation benchmarks miss critical real-world constraints: IFMTBench shows instruction following scales more sharply with model size than translation quality, and general instruction following rankings correlate weakly with translation behavior.
A 440MB multilingual translation model now rivals commercial APIs, opening the door for performant on-device translation.
