Search papers, labs, and topics across Lattice.
5
0
6
0
Relying on a single cleanliness score can lead to compounding errors in noisy-label learning, but TRACE uncouples the assessment of observed and pseudo labels for more reliable supervision.
Textual shortcuts in VLMs can significantly undermine reasoning accuracy, but a simple intervention can restore fresh visual computation.
Explicitly modeling object orientation can drastically reduce systematic errors in multimodal spatial reasoning, outperforming traditional methods by a significant margin.
Dissimilarity Invariance reveals a stable relationship between unrelated samples that can be harnessed to combat the challenges of noisy labels in deep learning.
LLMs can reason more efficiently by sharing intermediate thoughts during parallel search, achieving better accuracy with less computation.
