Search papers, labs, and topics across Lattice.
School of Computer Sciences, Universiti Sains Malaysia, Penang, Malaysia
4
0
7
Only 14.9% of citation-context retrieval gains are based on genuine future evidence, revealing a critical flaw in current evaluation methods.
Charge-based ranking can recover 99.2% of the performance gap in legal case retrieval, challenging the reliance on complex models for relevance assessment.
Discriminative readout from an omni-modal LLM slashes MSA error rates and latency, outperforming traditional generative methods by a wide margin.
Achieve real-time, high-fidelity speech synthesis with a 48ms latency by directly generating compressed audio codec tokens, bypassing the neural vocoder bottleneck.
