Search papers, labs, and topics across Lattice.
3
0
6
0
Fixed execution horizons in WAMs are inefficient, but TempoWAM's adaptive replanning can cut inference needs by nearly 27% while improving success rates on complex tasks.
Mamba-2 models, despite their theoretical capacity, fail to learn a simple stack-based rollback mechanism, instead collapsing to below-chance performance when faced with adversarial retraction pressure.
Constrained decoding, intended to improve LLM self-correction, can backfire by causing models to prioritize superficial formatting over semantic accuracy, leading to a new failure mode called "structure snowballing."
