Search papers, labs, and topics across Lattice.
2
0
3
7
LongRCA Bench reveals that diagnosing failures in long-horizon agents requires distinct metrics for responsible roles and root causes, with a new method achieving over 50% accuracy in role attribution.
Rethinking GPU reliability, this study shows that ranking nodes by failure risk can outperform traditional predictive maintenance methods, capturing 64% of failures in the top 5% of at-risk GPUs.
