Search papers, labs, and topics across Lattice.
Harbin Institute of Technology
5
0
9
8
LVLMs hallucinate in predictable bursts, and this self-rewarding decoding strategy slashes those errors in half.
MLLMs aren't just improving video translation quality; they're fundamentally changing how we approach it by jointly optimizing for semantic accuracy, timing, speaker identity, and emotional nuance.
LRMs can slash up to 40% of reasoning tokens without sacrificing accuracy by dynamically adjusting their "thinking speed" at each step.
The chaos of LLM tool use research gets tamed: a new framework reveals the hidden evolutionary relationships between prompting, supervised learning, and RL-based approaches.
Token-level policy gradients fall short in complex reasoning tasks, but treating sequences of tokens as unified actions can significantly boost performance in mathematical and coding benchmarks.
