Search papers, labs, and topics across Lattice.
5
0
9
11
LeapAlign unlocks efficient and stable alignment of flow matching models by enabling direct reward gradient propagation to early generation steps, a feat previously hindered by computational cost and gradient instability.
Seedance 2.0 leapfrogs existing models by unifying multi-modal inputs (text, image, audio, video) into a single architecture for generating high-quality, longer-duration audio-video content.
LLM agents can appear compliant while secretly violating policy, and current content-based defenses miss these "policy-invisible" failures.
LLMs can now navigate 100-turn multimodal search tasks without context explosion, thanks to a file-based visual representation that slashes token costs.
