MOON Improves Multitask Learning with OrthoNormalized Gradient Updates.

Shiji Zhou, Kunlin Lyu, Lei Zhang, Ruodong Wang, Yifan Sun· August 13, 2026 View original

Key takeaways

  • MOON is a new multi-objective optimization method for multi-task learning.
  • It manipulates gradients under spectral-nuclear norm geometry, suitable for matrix parameters.
  • MOON improves optimization efficiency and final performance in multi-task models.
  • The method is particularly beneficial for modern architectures like Transformers.

Who benefits

AI ResearchSoftware DevelopmentData ScienceNatural Language ProcessingComputer Vision

Summary

This paper introduces MOON (Multi-Objective OrthoNormalized Updates), a novel approach for multi-task learning that addresses limitations of Euclidean gradient manipulation in multi-objective optimization. MOON performs gradient manipulation under spectral-nuclear norm geometry, leading to more efficient optimization and improved performance in modern architectures like Transformers.

Current multi-objective optimization (MOO) methods in multi-task learning often treat model parameters as flat vectors, performing gradient manipulations in Euclidean space. This approach overlooks the inherent matrix structure of parameters in modern neural networks, such as Transformers, potentially hindering optimization efficiency. Researchers propose MOON (Multi-Objective OrthoNormalized Updates) to overcome this. MOON manipulates gradients under spectral-nuclear norm geometry, which is more appropriate for matrix-valued parameters, and then uses orthonormalized gradients for updates. This method aims to find the steepest descent direction more accurately in matrix space. The theoretical analysis shows MOON's convergence for non-convex objectives, and empirical results across various benchmarks confirm its ability to improve both optimization efficiency and the final performance of multi-task models. The code is publicly available.

Why it matters

For professionals developing multi-task AI models, MOON offers a more effective optimization strategy that can lead to better performance and faster convergence, especially with complex architectures like Transformers.

How to implement this in your domain

  1. 1Review the MOON framework and its theoretical underpinnings for multi-task learning.
  2. 2Integrate the MOON algorithm into your existing multi-task model training pipelines.
  3. 3Benchmark MOON's performance against current multi-objective optimization methods on your specific tasks.
  4. 4Analyze the impact of MOON on convergence speed and final model accuracy for various multi-task scenarios.

Original post by Shiji Zhou, Kunlin Lyu, Lei Zhang, Ruodong Wang, Yifan Sun

"arXiv:2608.11749v1 Announce Type: new Abstract: Multi-objective optimization (MOO) has demonstrated significant success in multi-task learning by mitigating task conflicts through gradient manipulation. However, most existing methods flatten model parameters into vectors and perf…"

View on X

Originally posted by Shiji Zhou, Kunlin Lyu, Lei Zhang, Ruodong Wang, Yifan Sun on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses