LBW-Guard: A Bounded Training Control Governance Layer for Stable LLM Training Under Stress

Forum topic · 小凯 · 2026-05-21

Summary

LBW-Guard (Learn-by-Wire Guard) is a bounded autonomous training-control governance layer that operates above AdamW to address instability, degraded runs, and wasted compute in modern language-model training. Rather than replacing the optimizer update rule, it observes training telemetry, interprets instability-sensitive regimes, and applies bounded control to optimizer execution while preserving fixed training objectives. Evaluated on a Qwen2.5-centered stress-and-robustness suite using WikiText-103 (with Qwen2.5-7B as the anchor, size comparisons against 3B and 14B variants, learning-rate stress tests, gradient-clipping baselines, and a TinyLlama-1B full-parameter sanity check), LBW-Guard reduces final perplexity from 13.21 to 10.74 (18.7% improvement) while cutting end-to-end time from 392.54s to 357.02s (1.10x speedup) in the 7B reference setting. Under strong learning-rate stress where AdamW degrades to perplexity 1885.24 (LR=3e-3) and 659.76 (LR=1e-3), LBW-Guard remains trainable at 11.57 and 10.33 respectively; gradient-clipping baselines do not reproduce this effect. The results support a scoped systems conclusion: stability-sensitive LLM training can benefit from a governance plane above the optimizer.

Original Abstract

Modern language-model training is increasingly exposed to instability, degraded runs, and wasted compute, especially under aggressive learning-rate, scale, and runtime-stress conditions. This paper introduces Learn-by-Wire Guard (LBW-Guard), a bounded autonomous training-control governance layer that operates above AdamW. Rather than replacing the optimizer update rule, LBW-Guard observes training telemetry, interprets instability-sensitive regimes, and applies bounded control to optimizer execution while preserving fixed training objectives.

Evaluation Setup

Corpus: WikiText-103 stress-and-robustness suite
Anchor model: Qwen2.5-7B
Model-size comparisons: Qwen2.5-3B and Qwen2.5-14B
Stress tests: aggressive learning rates, gradient-clipping baselines
Sanity check: TinyLlama-1B full-parameter (no-LoRA) run

Key Results

In the 7B reference setting, LBW-Guard reduces final perplexity from 13.21 to 10.74 (an 18.7% improvement), while reducing end-to-end time from 392.54s to 357.02s (a 1.10x speedup).
Under strong learning-rate stress, AdamW degrades to a final perplexity of 1885.24 at LR=3e-3 and 659.76 at LR=1e-3, whereas LBW-Guard remains trainable at 11.57 and 10.33, respectively.
Gradient-clipping baselines do not reproduce this effect.

Conclusion

These results support a scoped systems conclusion: stability-sensitive LLM training can benefit from a governance plane above the optimizer. LBW-Guard provides evidence that bounded runtime control can preserve productive compute under stress while remaining distinct from optimizer replacement and local gradient suppression.