arXiv：cs.LG（机器学习，全量分类）· Shuyang Jiang, Fucheng Deng, Yuchuan Luo, Zhenyu Wu·· 5 小时前AI 评分51
梯度冲突能否预测理解与生成的权衡？统一多模态模型冲突指标有效性的受控审计
Does Gradient Conflict Predict the Understanding--Generation Trade-off? A Controlled Audit of Conflict-Metric Validity in Unified Multimodal Models
AI 导读
arXiv 论文（arXiv:2609.38465）在受控测试床 GRIDUMM 中审计统一多模态模型（UMM）训练中梯度冲突指标对理解-生成权衡的预测有效性。
正文
Abstract:Unified multimodal models (UMMs) are increasingly designed around gradient conflict between understanding and generation objectives. The premise that reducing these metrics improves the downstream understanding-generation trade-off has never been tested directly. We audit it in a controlled testbed, GRIDUMM, which mirrors key structural ingredients of UMM training while making the ground-truth trade-off exactly computable. Across 63 configurations and 372 measured checkpoints, no directional conflict metric reaches an absolute Spearman correlation of 0.3 with a confidence interval excluding zero for conflict measured during training against the eventual trade-off. A dose-response intervention that monotonically suppresses conflict leaves the trade-off flat, separating correlation from causation. The norm ratio is a generation-failure detector and becomes null among configurations that master generation. Functional interference measures outperform directional conflict metrics, while training loss tracks the trade-off strongly. Our results do not show that conflict is useless; they show that its validity as a diagnostic target must be established, not assumed, and we release the audit protocol as a reusable standard.
| Comments: | 23 pages, 10 figures, 5 tables. Code will be made publicly available upon acceptance |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV) |
| Cite as: | arXiv:2609.38465 [cs.LG] |
| (or arXiv:2609.38465v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.38465 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Shuyang Jiang [view email]
[v1]
Tue, 29 Sep 2026 19:57:47 UTC (2,589 KB)
来源：arXiv：cs.LG（机器学习，全量分类） · arxiv.org