Search papers, labs, and topics across Lattice.
Peking University
5
1
8
10
Adding depthwise convolutions to Transformers can boost accuracy on downstream tasks while barely increasing model size.
Harness-Native routing transforms model selection from a mere cost-control mechanism into a powerful data engine that continuously improves agent performance.
Semantic acceptance rates can be misleading, with up to 44.2% of models failing to prevent observable harm even when they pass initial checks.
Adapter design can make or break coding performance in OpenClaw-style agents, with a full adapter boosting success rates by over 50 percentage points.
Diffusion Language Models are being held back by auto-regressive thinking, and unlocking their true potential requires a complete paradigm shift.
