Search papers, labs, and topics across Lattice.
3
0
7
Diffusion LLMs can achieve up to 6.1x higher throughput than autoregressive models by dynamically adjusting decoding granularity based on real-time load, a feat unattainable with fixed-block approaches.
Forget semantic relevance: this new keyframe sampling method, driven by information bottleneck theory, significantly boosts MLLM performance on long-form video understanding tasks.
Multimodal LLMs that ace single-turn safety can still be tricked into 90-100% ASR with multi-turn attacks, especially when modalities switch mid-conversation.
