analiza aiqbench
This paper addresses a critical instability in Direct Preference Optimization (DPO) known as the "squeezing effect," where probability mass collapses away from both preferred and unrelated responses during training. The authors introduce Gradient-Gated DPO, a method that selectively attenuates updates to low-probability rejected responses to prevent destructive suppression and maintain healthier model dynamics. By stabilizing preference optimization across multiple architectures and objectives, this work offers a robust mechanism for improving the reliability of language model alignment without altering the core objective function.
Gradient-Gated DPO: Stabilizing Preference Optimization in Language Models
arXiv:2605.02626v2 Announce Type: replace Abstract: Direct Preference Optimization (DPO) improves relative preference by increasing the margin between chosen and rejected responses, but this objective does not specify how probability mass should be redistributed to achieve that margin. Consequently, preference separation can improve even as probability mass collapses away from both preferred and unrelated responses, a phenomenon associated with the squeezing effect. We show that this pathology is localized: continued negative updates become destructive when rejected responses enter low-probability valleys, where further suppression induces harmful redistribution through the coupled softmax geometry. We introduce Gradient-Gated Preference Optimization (Gate-DPO), which uses the current probability geometry of each rejected response to selectively attenuate its contribution, limiting excessive rejected suppression while delaying preference-loss saturation. Gate-DPO is objective-agnostic and composes with existing preference methods. Across four architectures, two preference datasets, and multiple objectives, gating consistently reduces squeezing and improves chosen-response likelihood, while exhibiting robust reductions in rejected-response collapse across architectures and objectives. Mass-dynamics analysis further shows that large preference margins can conceal substantial suppression of unrelated responses, while gating achieves healthier redistribution. Together, our results show that preference margin alone is insufficient to characterize optimization health and that selectively controlling probability dynamics provides a simple, general mechanism for stabilizing preference optimization.
Abstract:Direct Preference Optimization (DPO) improves relative preference by increasing the margin between chosen and rejected responses, but this objective does not specify how probability mass should be redistributed to achieve that margin. Consequently, preference separation can improve even as probability mass collapses away from both preferred and unrelated responses, a phenomenon associated with the squeezing effect. We show that this pathology is localized: continued negative updates become destructive when rejected responses enter low-probability valleys, where further suppression induces harmful redistribution through the coupled softmax geometry. We introduce Gradient-Gated Preference Optimization (Gate-DPO), which uses the current probability geometry of each rejected response to selectively attenuate its contribution, limiting excessive rejected suppression while delaying preference-loss saturation. Gate-DPO is objective-agnostic and composes with existing preference methods. Across four architectures, two preference datasets, and multiple objectives, gating consistently reduces squeezing and improves chosen-response likelihood, while exhibiting robust reductions in rejected-response collapse across architectures and objectives. Mass-dynamics analysis further shows that large preference margins can conceal substantial suppression of unrelated responses, while gating achieves healthier redistribution. Together, our results show that preference margin alone is insufficient to characterize optimization health and that selectively controlling probability dynamics provides a simple, general mechanism for stabilizing preference optimization.
Machine Learning (cs.LG)
arXiv:2605.02626 [cs.LG]
(or arXiv:2605.02626v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2605.02626 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 292 cuvinte
Subiecte conexe
Introducing Instella-MoE: A State-of-the-Art Fully Open Mixture-of-Experts Language Model
AMD · 24 iul. 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
When Partial Credit Opposes Completion
Snorkel AI Blog · 15 oct. 2026