Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards, Xuehui Yu★, Fucheng Cai, Meiyi Wang, Xiaopeng Fan, Harold Soh★, International Conference on Machine Learning (ICML), 2026
Links:
Inference-time guidance can easily push your sampling process off the data manifold. This work answers the question: how do we harness large, complex pretrained generative priors to satisfy multiple constraints at inference time without drifting off-manifold (i.e., avoiding hallucinated generation)?
The paper introduces Conflict-Aware Additive (CAR) Guidance, a plug-and-play module that detects and rectifies this off-manifold drift on the fly. CAR Guidance is validated across pixel-space image editing, robot planning, and 3D point-cloud robot manipulation.
🌟 A key insight is that, in compositional-reward settings, the approximation error grows sharply with both the gradient misalignment between guidance channels (\(1 - \cos\phi\), where \(\phi\) is the average angular divergence) and the number of reward functions \(G\).
Resources
You can find our paper here. Check out our repository here on GitHub.
Citation
Please consider citing our paper if you build upon our results and ideas.
Xuehui Yu★, Fucheng Cai, Meiyi Wang, Xiaopeng Fan, Harold Soh★, “Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards”, International Conference on Machine Learning (ICML), 2026
@inproceedings{yu2026conflict, title={Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards}, author={Yu, Xuehui and Cai, Fucheng and Wang, Meiyi and Fan, Xiaopeng and Soh, Harold}, booktitle={International Conference on Machine Learning (ICML)}, year={2026} }
Contact
If you have questions or comments, please contact Xuehui.
Acknowledgements
This research is supported by the RIE2025 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) (Grant No. I2501E0041), administered by A*STAR, as well as supported by Schaeffler (Singapore) PTE. LTD. and NTU Singapore through Schaeffler-NTU Corporate Lab: Intelligent Mechatronics Hub. —
