![[2602.21565] Training-free Composition of Pre-trained GFlowNets for Multi-Objective Generation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21565] Training-free Composition of Pre-trained GFlowNets for Multi-Objective Generation
Summary
This article presents a novel approach to using pre-trained GFlowNets for multi-objective generation without the need for additional training, enhancing efficiency in diverse applications.
Why It Matters
The proposed training-free mixing policy allows for rapid adaptation of GFlowNets to various objectives, significantly reducing computational overhead and making it applicable to real-world scenarios where multiple conflicting objectives are common. This advancement could streamline processes in scientific discovery and other fields requiring diverse solution exploration.
Key Takeaways
- Introduces a training-free method for composing pre-trained GFlowNets.
- Enables quick adaptation to multiple objectives without retraining.
- Demonstrates comparable performance to traditional methods requiring additional training.
- Supports a range of reward combinations, enhancing flexibility.
- Proves recovery of target distribution for linear scalarization.
Computer Science > Machine Learning arXiv:2602.21565 (cs) [Submitted on 25 Feb 2026] Title:Training-free Composition of Pre-trained GFlowNets for Multi-Objective Generation Authors:Seokwon Yoon, Youngbin Choi, Seunghyuk Cho, Seungbeom Lee, MoonJeong Park, Dongwoo Kim View a PDF of the paper titled Training-free Composition of Pre-trained GFlowNets for Multi-Objective Generation, by Seokwon Yoon and 5 other authors View PDF HTML (experimental) Abstract:Generative Flow Networks (GFlowNets) learn to sample diverse candidates in proportion to a reward function, making them well-suited for scientific discovery, where exploring multiple promising solutions is crucial. Further extending GFlowNets to multi-objective settings has attracted growing interest since real-world applications often involve multiple, conflicting objectives. However, existing approaches require additional training for each set of objectives, limiting their applicability and incurring substantial computational overhead. We propose a training-free mixing policy that composes pre-trained GFlowNets at inference time, enabling rapid adaptation without finetuning or retraining. Importantly, our framework is flexible, capable of handling diverse reward combinations ranging from linear scalarization to complex non-linear logical operators, which are often handled separately in previous literature. We prove that our method exactly recovers the target distribution for linear scalarization and quantify the approximati...
