Search papers, labs, and topics across Lattice.
6
0
9
Fine-tuning just 0.14% of parameters, ENCORE boosts VLM accuracy by 1.43% through innovative entropy-guided cropping and attention techniques.
Pruning visual tokens based on head alignment can retain nearly all performance while drastically reducing computational costs.
Optimizing multiple moments of failure probabilities can dramatically enhance LLM reasoning performance, outperforming traditional single-moment approaches.
Allocating rollout budgets based on state informativeness allows LLM agents to achieve superior performance in complex decision-making tasks without increasing computational costs.
Forget hand-tuning: VisPCO automatically finds optimal visual token pruning configurations in VLMs, outperforming predefined strategies across diverse benchmarks.
Forget memorizing surface patterns: RADAR leverages reinforcement learning to teach LLMs genuine relational reasoning, boosting knowledge graph performance by 5-6%.
