What Limits Bidirectional Model’s Generative Capabilities? A Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning

Zuchao Li, Yonghua Hei, Qiwei Li, Lefei Zhang, Ping Wang, Hai Zhao, Baoyuan Qi, Liu Guoming
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:34564-34577, 2025.

Abstract

Large Language Models (LLMs) excel in generation tasks, yet their causal attention mechanisms limit performance in embedding tasks. While bidirectional modeling may enhance embeddings, naively fine-tuning unidirectional models bidirectionally severely degrades generative performance. To investigate this trade-off, we analyze attention weights as dependence indicators and find that bidirectional fine-tuning increases subsequent dependence, impairing unidirectional generation. Through systematic Transformer module evaluations, we discover the FFN layer is least affected by such dependence. Leveraging this discovery, we propose UBMoE-LLM, a novel Uni-Bi-directional Mixture-of-Experts LLM, which integrates the original unidirectional FFN with a bidirectionally fine-tuned FFN via unsupervised contrastive learning. This MoE-based approach enhances embedding performance while preserving robust generation. Extensive experiments across diverse datasets and model scales validate our attention dependence metric and demonstrate UBMoE-LLM’s superior generative quality and reduced hallucination. Code is available at: https://github.com/heiyonghua/ubmoe_llm.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-li25y, title = {What Limits Bidirectional Model’s Generative Capabilities? {A} Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning}, author = {Li, Zuchao and Hei, Yonghua and Li, Qiwei and Zhang, Lefei and Wang, Ping and Zhao, Hai and Qi, Baoyuan and Guoming, Liu}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {34564--34577}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/li25y/li25y.pdf}, url = {https://proceedings.mlr.press/v267/li25y.html}, abstract = {Large Language Models (LLMs) excel in generation tasks, yet their causal attention mechanisms limit performance in embedding tasks. While bidirectional modeling may enhance embeddings, naively fine-tuning unidirectional models bidirectionally severely degrades generative performance. To investigate this trade-off, we analyze attention weights as dependence indicators and find that bidirectional fine-tuning increases subsequent dependence, impairing unidirectional generation. Through systematic Transformer module evaluations, we discover the FFN layer is least affected by such dependence. Leveraging this discovery, we propose UBMoE-LLM, a novel Uni-Bi-directional Mixture-of-Experts LLM, which integrates the original unidirectional FFN with a bidirectionally fine-tuned FFN via unsupervised contrastive learning. This MoE-based approach enhances embedding performance while preserving robust generation. Extensive experiments across diverse datasets and model scales validate our attention dependence metric and demonstrate UBMoE-LLM’s superior generative quality and reduced hallucination. Code is available at: https://github.com/heiyonghua/ubmoe_llm.} }
Endnote
%0 Conference Paper %T What Limits Bidirectional Model’s Generative Capabilities? A Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning %A Zuchao Li %A Yonghua Hei %A Qiwei Li %A Lefei Zhang %A Ping Wang %A Hai Zhao %A Baoyuan Qi %A Liu Guoming %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-li25y %I PMLR %P 34564--34577 %U https://proceedings.mlr.press/v267/li25y.html %V 267 %X Large Language Models (LLMs) excel in generation tasks, yet their causal attention mechanisms limit performance in embedding tasks. While bidirectional modeling may enhance embeddings, naively fine-tuning unidirectional models bidirectionally severely degrades generative performance. To investigate this trade-off, we analyze attention weights as dependence indicators and find that bidirectional fine-tuning increases subsequent dependence, impairing unidirectional generation. Through systematic Transformer module evaluations, we discover the FFN layer is least affected by such dependence. Leveraging this discovery, we propose UBMoE-LLM, a novel Uni-Bi-directional Mixture-of-Experts LLM, which integrates the original unidirectional FFN with a bidirectionally fine-tuned FFN via unsupervised contrastive learning. This MoE-based approach enhances embedding performance while preserving robust generation. Extensive experiments across diverse datasets and model scales validate our attention dependence metric and demonstrate UBMoE-LLM’s superior generative quality and reduced hallucination. Code is available at: https://github.com/heiyonghua/ubmoe_llm.
APA
Li, Z., Hei, Y., Li, Q., Zhang, L., Wang, P., Zhao, H., Qi, B. & Guoming, L.. (2025). What Limits Bidirectional Model’s Generative Capabilities? A Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:34564-34577 Available from https://proceedings.mlr.press/v267/li25y.html.

Related Material