Search papers, labs, and topics across Lattice.
University of Edinburgh
4
0
6
51
Larger context sizes can amplify spurious signal reliance in tabular ICLs, leading to a 1.74x increase in routing errors when models are deployed in new environments.
Causal features can be identified post-hoc with 100% accuracy using the Normalised Sensitivity Ratio, even in complex environments.
VLMs struggle with emotion recognition not because they lack visual acuity, but because pre-training exacerbates dataset biases and sparse temporal sampling misses crucial micro-expressions.
Current text-to-image evaluation metrics miss critical clinical errors, but CSEval uses language models to catch them.
