No Free Efficiency: Revisiting the Trade-off Between Training Efficiency and Model Vulnerability
Efficient training recipes make vision and language models easier to attack and less robust.
The paper examines whether selective data sampling, efficient pre-training, and simplified reinforcement-learning pipelines reduce robustness while cutting cost. Across vision and language models, efficiency-oriented variants showed sharper loss geometry, changed representations, and greater susceptibility to adversarial and privacy attacks. Models trained with simplified zero-RL recipes also showed more catastrophic forgetting and overconfidence than models trained with conventional alignment pipelines. The authors argue for training objectives that jointly consider performance, cost, and security.
- Cross-domain study of efficiency versus model vulnerability
- Efficient variants exhibit sharper loss geometry
- Adversarial and privacy attack susceptibility increases
- Simplified RL raises forgetting and overconfidence
From the original article
1 min read at arxiv.orgTraining 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.
By Yiyong Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.