Search papers, labs, and topics across Lattice.
5
0
9
0
RL models not only outperform SFT counterparts in reasoning tasks but also develop a hierarchical architecture that enhances representational quality and processing efficiency.
Spatial attention in VLMs is nearly irrelevant to accuracy, with self-consistency emerging as the true indicator of reliability.
LLMs can learn to reason more effectively by breaking down the reasoning process and optimizing each step individually.
Cutting LLMs' reasoning token budget can backfire spectacularly, tanking performance even below that of models with *no* reasoning at all.
Spot poisoned LoRA adapters without running them: a weight-space analysis achieves 97% accuracy in detecting backdoors, even when the trigger is unknown.
