Search papers, labs, and topics across Lattice.
6
0
9
Training dynamics of Transformers can be reduced to a low-dimensional manifold, revealing how inductive reasoning emerges from data statistics and model initialization.
Language-adaptive tokenization can significantly enhance morphological alignment without the need for separate vocabularies for each language.
Evasive steganographic payloads in LLMs can retain high recoverability while eluding detection, but targeted data interventions can restore visibility to these hidden secrets.
Escape the greedy trap: Convex optimization yields tokenizers that compress better and come with optimality guarantees.
Achieve 70% language identification accuracy with just five labeled samples per language using a novel tokenization-based approach.
Post-training LLMs is more about finding the right "key" (a few kilobytes of parameters) to unlock pre-existing knowledge than learning new information.
