Search papers, labs, and topics across Lattice.
Shanghai Jiao Tong University
4
0
6
FlowTracer reveals that optimizing token-level rewards based on attention-induced information flow can dramatically enhance reasoning performance in LLMs.
Even the best MLLMs struggle to meet user requirements, achieving only 66% coverage of essential task functions.
By explicitly encoding the structural rigidity of ships, SDF-Net achieves state-of-the-art cross-modal ship re-identification, demonstrating the power of incorporating physical priors to overcome radiometric discrepancies.
Autonomous driving's next leap hinges on reasoning, not just perception, but current LLM-based approaches are too slow for real-time control.
