AI Daily — April 20, 2026
2026-04-20
rss.arxiv.org
LACE: Lattice Attention for Cross-thread Exploration
LACE introduces cross-thread attention that allows multiple parallel reasoning paths in an LLM to share intermediate states and correct each other during inference, rather than running as fully independent samples. The authors address the lack of training data for this collaborative reasoning pattern by constructing a synthetic pipeline that teaches models to communicate across threads. This is a meaningful architectural departure from standard parallel sampling (e.g., majority voting) and could improve both efficiency and accuracy on complex reasoning tasks.
rss.arxiv.org
Aletheia: Gradient-Guided Layer Selection for Efficient LoRA Fine-Tuning Across Architectures
Aletheia proposes replacing uniform LoRA adapter placement with gradient-probe-guided layer selection and asymmetric rank allocation, applying adapters only to layers most relevant to the downstream task. Evaluated across 14 models from 8 architecture families (0.5B–72B parameters, including dense and MoE), it achieves 15–28% training speedup without degrading task performance. The breadth of the benchmark and practical speedup make this a notable contribution to parameter-efficient fine-tuning tooling.