SCOUT localizes hangs, stragglers, and silent data corruption in LLM pre-training through strict-majority consensus among equivalent replicas, with replay-based checkpoint certification.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SCOUT: Symmetric Consensus Outlier Detection for Failure Localization in LLM Pre-Training
SCOUT localizes hangs, stragglers, and silent data corruption in LLM pre-training through strict-majority consensus among equivalent replicas, with replay-based checkpoint certification.