analiza aiqbench
P$^2$-DPO introduces a novel training paradigm for Large Vision-Language Models that generates its own preference pairs to specifically target perceptual bottlenecks and visual robustness against image degradation. By employing an on-policy construction method and a calibration loss, the approach aligns visual signals with text generation more effectively than baselines relying on costly human feedback. This method demonstrates improved performance on benchmarks and attention region fidelity while reducing the data and cost requirements typically associated with human-in-the-loop optimization.
P\textsuperscript{2}-DPO: Grounding Hallucination in Perceptual Processing via Calibration Direct Preference Optimization
arXiv:2606.03376v1 Announce Type: cross Abstract: Hallucination has recently garnered significant research attention in Large Vision-Language Models (LVLMs). Direct Preference Optimization (DPO) aims to learn directly from the corrected preferences provided by humans, thereby addressing the hallucination issue. Despite its success, this paradigm has yet to specifically target the perceptual bottleneck in attended regions or address insufficient Visual Robustness against image degradation. Furthermore, existing preference pairs are often vision-agnostic and their inherently off-policy nature limits their effectiveness in guiding model learning. To address these challenges, we propose Perceptual Processing Direct Preference Optimization (P\textsuperscript{2}-DPO), a novel training paradigm in which the model generates and learns from its own preference pairs, thereby directly addressing the identified visual bottlenecks while inherently avoiding the issues of vision-agnostic and off-policy data. It introduces: (1) an on-policy preference pairs construction method targeting Focus-and-Enhance perception and Visual Robustness, and (2) a well-designed Calibration Loss to precisely align visual signals with the causal generation of text. Experimental results demonstrate that with a comparable amount of training data and cost, P\textsuperscript{2}-DPO outperforms strong baselines that rely on costly human feedback on benchmarks. Furthermore, evaluations on Attention Region Fidelity (ARF) and image degradation scenarios validate the effectiveness of P\textsuperscript{2}-DPO in addressing perceptual bottleneck in attended regions and improving Visual Robustness against degraded inputs.
Abstract:Hallucination has recently garnered significant research attention in Large Vision-Language Models (LVLMs). Direct Preference Optimization (DPO) aims to learn directly from the corrected preferences provided by humans, thereby addressing the hallucination issue. Despite its success, this paradigm has yet to specifically target the perceptual bottleneck in attended regions or address insufficient Visual Robustness against image degradation. Furthermore, existing preference pairs are often vision-agnostic and their inherently off-policy nature limits their effectiveness in guiding model learning. To address these challenges, we propose Perceptual Processing Direct Preference Optimization (P$^2$-DPO), a novel training paradigm in which the model generates and learns from its own preference pairs, thereby directly addressing the identified visual bottlenecks while inherently avoiding the issues of vision-agnostic and off-policy data. It introduces: (1) an on-policy preference pairs construction method targeting Focus-and-Enhance perception and Visual Robustness, and (2) a well-designed Calibration Loss to precisely align visual signals with the causal generation of text. Experimental results demonstrate that with a comparable amount of training data and cost, P$^2$-DPO outperforms strong baselines that rely on costly human feedback on benchmarks. Furthermore, evaluations on Attention Region Fidelity (ARF) and image degradation scenarios validate the effectiveness of P$^2$-DPO in addressing perceptual bottleneck in attended regions and improving Visual Robustness against degraded inputs.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2606.03376 [cs.CV]
(or arXiv:2606.03376v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2606.03376 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 · 312 cuvinte
Subiecte conexe
P$^2$-DPO: Grounding Hallucination in Perceptual Processing via Calibration Direct Preference Optimization
arXiv · 5 iun. 2026
How Tableau Built Agentic Analytics at Salesforce: A Four-Part Framework
Salesforce Blog · 12 oct. 2026
Natera’s intelligent appointment scheduling with Amazon Bedrock AgentCore
AWS · 26 aug. 2026
FDIR: Harmonizing Fidelity and Human-Machine Preference in Lossy Compression Image Restoration
arXiv · 4 aug. 2026