Cisco-VAIL Fingerprinting Shows Nemotron Models Retain Qwen Lineage Despite US Labels

{"The study targeted known derivation pairs where NVIDIA's Nemotron series incorporated Alibaba's Qwen base weights. Two independent methods—CISCO's Model Provenance Kit inspecting internal artifacts and VAIL's external behavioral probes—both detected persistent relationships after post-training. This demonstrates that publisher nationality and model name do not erase upstream dependencies embedded in learned parameters.","Provenance entanglement arises because most production models start from existing checkpoints rather than random initialization. The research notes that no manifest lists these inherited weights, creating an opaque supply chain analogous to untracked software libraries. If an upstream model later reveals a backdoor or bias, downstream users lack a mechanism to identify affected deployments.","Current US policy treats labeled country of origin as a primary risk signal, yet the data show this proxy fails when lineage crosses borders. Regulators and procurement officers therefore require an AI-equivalent SBOM that records base checkpoints, fine-tuning datasets, and teacher models to close the visibility gap.","Next steps include mandatory lineage disclosure in federal AI acquisitions and development of standardized fingerprinting benchmarks. Without these, organizations will continue to inherit undetected characteristics from restricted jurisdictions while believing their models are domestically sourced."}