analiza aiqbench
This research identifies a critical trade-off where strategies designed to reduce the computational cost of training foundation models inadvertently increase their susceptibility to adversarial attacks, privacy breaches, and catastrophic forgetting. By analyzing internal model geometry and functional representations across vision and language domains, the study demonstrates that efficiency optimizations often compromise robustness and security rather than offering a "free lunch." The authors advocate for a paradigm shift toward multi-objective training pipelines that jointly optimize performance, cost, and safety to prevent these unintended vulnerabilities.
No Free Efficiency: Revisiting the Trade-off Between Training Efficiency and Model Vulnerability
arXiv:2609.33898v1 Announce Type: new Abstract: Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.
Abstract:Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.
Machine Learning (cs.LG); Cryptography and Security (cs.CR)
arXiv:2609.33898 [cs.LG]
(or arXiv:2609.33898v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.33898 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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.
Text integral sub licența deschisă a publicației · 299 cuvinte
Subiecte conexe
Dropless MoE Training in JAX with Primus-Turbo
AMD · 10 iun. 2026
Deconfusing Corrigibility
AI Alignment Forum · 17 oct. 2026
Fable is SOTA at CIFAR Speedrun (& specification gaming)
LessWrong · 20 iul. 2026
[AINews] Reflection Beam - 501B-A23B American Open Model
Latent Space · 6 oct. 2026