Auditing Invisible Weight Updates with Reference Traces

Abstract:Master weights and stochastic rounding bypass invisible stored-weight updates but do not locate lost direct-storage proposals or parameters worth protecting. We ask what one matched high-precision pilot trace can audit before a low-precision run. The exact target-code event is distinct from projected-code and archived symmetric-spacing aggregate proxies; neither proxy is a pathwise certificate. In the principal controlled two-layer grid, the archived symmetric-spacing proxy tracks nondegenerate crossings spanning 384x in time with median predicted/measured ratio 1.000 and 94.5% within 15%. In analytic-grid GPT-2-124M/OpenWebText runs with common initial target codes, changing only post-optimizer write-back from round-to-nearest to stochastic rounding recovers 89.2-92.2% of the loss gap through 10,000 steps. The same policy contrast recovers 92.9-93.6% at 3,000 E5M2 steps in one fixed 162.2M-parameter decoder configuration, but its visibility trajectories are not informative enough to establish diagnostic transfer. At exactly 10% of the eligible-2-D blanket fp32-master payload, mean recoveries across three fixed seeds are 71.4% for clipped signed leverage, 69.4% for absolute proposal contribution, and 2.1% for freeze-only; the latter comparison does not establish leverage superiority. In a post-selection fixed-target check, one seed-2031 absolute-proposal-contribution mask, reused unchanged across fixed target seeds 2035-2037, recovers 67.8-69.0% (mean 68.5%) at the same exact 10% eligible-2-D fp32-master payload and 1000-step horizon; target-specific masks recover 67.1-70.6% (mean 68.7%). A reference trace therefore supports separate aggregate-visibility and first-order-importance analyses, with failed transfer and superiority tests delimiting both.

Machine Learning (cs.LG)

arXiv:2607.09800 [cs.LG]

(or arXiv:2607.09800v2 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2607.09800 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.