Search papers, labs, and topics across Lattice.
3
0
6
0
Projector fine-tuning, commonly used for aligning MLLMs, unexpectedly introduces backdoor vulnerabilities with activation mechanisms distinct from those in text-only LLMs.
Current video object segmentation models fall apart when forced to reason causally, highlighting a critical gap between research benchmarks and real-world applicability.
Current methods for on-the-fly category discovery are fundamentally flawed because they treat it as a static classification problem, but PACO demonstrates that a dynamic, calibrated approach yields substantial improvements.
