Search papers, labs, and topics across Lattice.
Gensyn
5
0
9
Step-aligned feedback in self-distillation can boost model performance by over 16 points by precisely targeting reasoning failures without disrupting correct outputs.
IR3DE achieves near-optimal performance in reasoning tasks while enabling seamless integration of new domain experts, all without the need for retraining.
Training LLMs with diverse, collaborative models doesn't have to sacrifice convergence or generalization – F-TIS proves it.
Forget expensive finetuning: DUME lets you dynamically combine specialized language models into a powerful MoE *without* any additional training.
Even a single compromised pipeline stage can inject backdoors that drastically misalign LLMs, bypassing standard safety alignment.
