# Auditing Invisible Weight Updates with Reference Traces

arXiv: https://arxiv.org/abs/2607.09800 · HTML: https://mlsanity.com/paper/2607.09800/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 32 of its 51 references that are in the map.
- First posted on arXiv 9 Jul 2026.

## Builds on (most cited first, 30 of 32)

- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [GLU Variants Improve Transformer](https://mlsanity.com/paper/2002.05202/) (2020)
- [Root Mean Square Layer Normalization](https://mlsanity.com/paper/1910.07467/) (2019)
- [Mixed Precision Training](https://mlsanity.com/paper/1710.03740/) (2017)
- [HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision](https://mlsanity.com/paper/1905.03696/) (2019)
- [Deep Learning with Limited Numerical Precision](https://mlsanity.com/paper/1502.02551/) (2015)
- [FP8 Formats for Deep Learning](https://mlsanity.com/paper/2209.05433/) (2022)
- [Training Deep Neural Networks with 8-bit Floating Point Numbers](https://mlsanity.com/paper/1812.08011/) (2018)
- [FP8-LM: Training FP8 Large Language Models](https://mlsanity.com/paper/2310.18313/) (2023)
- [Quartet: Native FP4 Training Can Be Optimal for Large Language Models](https://mlsanity.com/paper/2505.14669/) (2025)
- [Training LLMs with MXFP4](https://mlsanity.com/paper/2502.20586/) (2025)
- [COAT: Compressing Optimizer states and Activation for Memory-Efficient FP8 Training](https://mlsanity.com/paper/2410.19313/) (2024)
- [FracTrain: Fractionally Squeezing Bit Savings Both Temporally and Spatially for Efficient DNN Training](https://mlsanity.com/paper/2012.13113/) (2020)
- [Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation](https://mlsanity.com/paper/2601.22813/) (2026)
- [Revisiting BFloat16 Training](https://mlsanity.com/paper/2010.06192/) (2020)
- [Unit Scaling: Out-of-the-Box Low-Precision Training](https://mlsanity.com/paper/2303.11257/) (2023)
- [Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention](https://mlsanity.com/paper/2510.04212/) (2025)
- [Stochastic Rounding for LLM Training: Theory and Practice](https://mlsanity.com/paper/2502.20566/) (2025)
- [Speeding up and reducing memory usage for scientific machine learning via mixed precision](https://mlsanity.com/paper/2401.16645/) (2024)
- [Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention](https://mlsanity.com/paper/2607.04422/) (2026)
- [Understanding and Exploiting Weight Update Sparsity for Communication-Efficient Distributed RL](https://mlsanity.com/paper/2602.03839/) (2026)
- [Direct Quantized Training of Language Models with Stochastic Rounding](https://mlsanity.com/paper/2412.04787/) (2024)
- [Understanding Quantization of Optimizer States in LLM Pre-training: Dynamics of State Staleness and Effectiveness of State Resets](https://mlsanity.com/paper/2603.16731/) (2026)
- [ECO: Quantized Training without Full-Precision Master Weights](https://mlsanity.com/paper/2601.22101/) (2026)
- [Pretraining large language models with MXFP4 on Native FP4 Hardware](https://mlsanity.com/paper/2605.09825/) (2026)
- [STQuant: Spatio-Temporal Adaptive Framework for Optimizer Quantization in Large Multimodal Model Training](https://mlsanity.com/paper/2604.06836/) (2026)
- [Model Compression with Exact Budget Constraints via Riemannian Manifolds](https://mlsanity.com/paper/2605.00649/) (2026)
- [dMX: Differentiable Mixed-Precision Assignment for Low-Precision Floating-Point Formats](https://mlsanity.com/paper/2606.04115/) (2026)
- [Quantizing Time-Series Models As Dynamical Systems: Trajectory-Based Quantization Sensitivity Score](https://mlsanity.com/paper/2606.13300/) (2026)
- [FORGE: Fused On-Register Gradient Elimination for Memory-Efficient LLM Training](https://mlsanity.com/paper/2606.22932/) (2026)

## Abstract (arXiv)

Direct low-precision write-back can erase nonzero optimizer proposals. We ask what a high-precision reference trace establishes before a low-precision run. The exact target-code event is auditable coordinatewise on a realized target trajectory; pre-run aggregate projection also assumes the reference remains a useful counterfactual. In a controlled two-layer grid, 55/72 cells have measured and predicted post-initialization crossings: times span 384×, 52/55 are within 15%, and 4/72 differ in category. Matched decoder experiments show stochastic rather than nearest write-back recovers most of the loss gap. A prospective analytic-grid E4M3 audit reuses one fp32 trace across three unseen NeoX-style seeds. It passes absolute-accuracy and skill gates (macro RMSE 0.00858) but fails directional specificity. In a target-outcome-blind comparison, a historical template has lower descriptive RMSE (0.00360) than the predeclared source predictor (0.00438); a post-outcome decomposition assigns 99.65% of variation to common time, while a privileged matched-reference correction reaches 0.00283. Persistent-native Study~1 pairs three seeds across two schedules. Five cells are canonical; a manual sixth lacks canonical process identity, so the registered result remains inconclusive. A retrospective protocol-deviation analysis is negative because the complete constant-mid cohort is disjoint from the recovered cosine-restart cell. Study~2 reports mean full-SR/dead-zone-SR recoveries of 0.9766/0.9777 and a ratio of 1.0012, a policy contrast rather than causal mediation. Simulated-INT3 Study~3 replays six checkpoints and observes a 7.3071-nat (69.71%) validation-loss reduction in one fixed seed. Exact events and write-back effects are auditable, but aggregate forecasts can reflect shared time rather than source-specific transfer.