Source
Zeng et al. (2022), LIT-054 — ARXIV-2210.02414.
Overlap with SOTA-070, stated plainly
This and SOTA-070 are the same recommendation at different resolutions: track gradient norms to detect instability, per layer there and in aggregate here. Both are sourced to LIT-054 and both arrived in the same import.
The distinct content is the aggregate statistic itself — the global norm, which is what clipping already computes (SOTA-071), so on a run that clips the signal is free. Its usefulness is in the trend rather than the value: a norm that has been falling for 50,000 steps and turns upward is the earliest of the signals in this cluster, ahead of the loss spike that SOTA-098 watches for and well ahead of the NaN that SOTA-072 catches.
Standing
Kept rather than merged into SOTA-070, because the free-with-clipping aggregate and the deliberate per-layer pass are genuinely different in cost. But the record should not pretend they are independent findings: they are one paper’s stability section split into bullets, which is the pattern #107 is about.