Improving LLM-Based Recommenders with Conservative Generative Flow Networks

Xuan Yu, Feng Niu, Rui Zhu, Yudong Zhang, Xu Wang, Yang Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:150907-150921, 2026.

Abstract

Generative Flow Networks (GFlowNets) have recently been used to improve diversity and mitigate popularity bias in LLM-based recommender systems, yet most objectives are developed under online-style assumptions. In offline LLM-based recommendation, learning is constrained to a fixed logged dataset, yielding partial support over token transitions on the dataset-induced token-prefix DAG. Naively applying Sub-Trajectory Balance (SubTB) becomes non-identifiable and can arbitrarily allocate probability mass to unsupported regions. We formalize this failure and identify three sources of non-identifiability that induce distributional shift between the dataset-implied policy and the learned policy: (i) flow overestimation, (ii) forward mass leakage, and (iii) backward compensation. To address it, we propose CFlower, which introduces a conservative SubTB objective that explicitly penalizes unsupported forward flow mass, and combines it with dataset-constrained policy learning with on-policy sampling on the dataset-induced DAG for efficient training under offline constraints. Experiments on three Amazon recommendation datasets show that CFlower improves distributional matching and delivers a stronger accuracy–exposure trade-off than prior GFlowNet and SFT baselines, while serving as a more reliable reference policy for downstream RL fine-tuning.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-yu26am, title = {Improving {LLM}-Based Recommenders with Conservative Generative Flow Networks}, author = {Yu, Xuan and Niu, Feng and Zhu, Rui and Zhang, Yudong and Wang, Xu and Wang, Yang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {150907--150921}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/yu26am/yu26am.pdf}, url = {https://proceedings.mlr.press/v306/yu26am.html}, abstract = {Generative Flow Networks (GFlowNets) have recently been used to improve diversity and mitigate popularity bias in LLM-based recommender systems, yet most objectives are developed under online-style assumptions. In offline LLM-based recommendation, learning is constrained to a fixed logged dataset, yielding partial support over token transitions on the dataset-induced token-prefix DAG. Naively applying Sub-Trajectory Balance (SubTB) becomes non-identifiable and can arbitrarily allocate probability mass to unsupported regions. We formalize this failure and identify three sources of non-identifiability that induce distributional shift between the dataset-implied policy and the learned policy: (i) flow overestimation, (ii) forward mass leakage, and (iii) backward compensation. To address it, we propose CFlower, which introduces a conservative SubTB objective that explicitly penalizes unsupported forward flow mass, and combines it with dataset-constrained policy learning with on-policy sampling on the dataset-induced DAG for efficient training under offline constraints. Experiments on three Amazon recommendation datasets show that CFlower improves distributional matching and delivers a stronger accuracy–exposure trade-off than prior GFlowNet and SFT baselines, while serving as a more reliable reference policy for downstream RL fine-tuning.} }
Endnote
%0 Conference Paper %T Improving LLM-Based Recommenders with Conservative Generative Flow Networks %A Xuan Yu %A Feng Niu %A Rui Zhu %A Yudong Zhang %A Xu Wang %A Yang Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-yu26am %I PMLR %P 150907--150921 %U https://proceedings.mlr.press/v306/yu26am.html %V 306 %X Generative Flow Networks (GFlowNets) have recently been used to improve diversity and mitigate popularity bias in LLM-based recommender systems, yet most objectives are developed under online-style assumptions. In offline LLM-based recommendation, learning is constrained to a fixed logged dataset, yielding partial support over token transitions on the dataset-induced token-prefix DAG. Naively applying Sub-Trajectory Balance (SubTB) becomes non-identifiable and can arbitrarily allocate probability mass to unsupported regions. We formalize this failure and identify three sources of non-identifiability that induce distributional shift between the dataset-implied policy and the learned policy: (i) flow overestimation, (ii) forward mass leakage, and (iii) backward compensation. To address it, we propose CFlower, which introduces a conservative SubTB objective that explicitly penalizes unsupported forward flow mass, and combines it with dataset-constrained policy learning with on-policy sampling on the dataset-induced DAG for efficient training under offline constraints. Experiments on three Amazon recommendation datasets show that CFlower improves distributional matching and delivers a stronger accuracy–exposure trade-off than prior GFlowNet and SFT baselines, while serving as a more reliable reference policy for downstream RL fine-tuning.
APA
Yu, X., Niu, F., Zhu, R., Zhang, Y., Wang, X. & Wang, Y.. (2026). Improving LLM-Based Recommenders with Conservative Generative Flow Networks. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:150907-150921 Available from https://proceedings.mlr.press/v306/yu26am.html.

Related Material