Search papers, labs, and topics across Lattice.
Affiliation:
2
0
5
Leading sign language models lose over 80% of their BLEU score the moment they encounter an unseen signer, exposing widespread benchmark inflation driven by identity and sentence leakage.
Current embedding benchmarks hide critical failure modes: models exhibit surprisingly specific and decoupled robustness profiles when exposed to realistic variations in text style, length, and language.
