Beyond In-Distribution Preservation: Recovering Generalization in Quantized VLAs via Vulnerability-Oriented Tuning

📄 arXiv: 2610.05745v1 📥 PDF

作者: Shen Ruan, Wenchang Gao, Jin Wang, Siao Liu, Zhoxizhuoma, Dongchun Ren, Xin Zheng

分类: cs.LG, cs.RO

发布日期: 2026-10-05

🔗 代码/项目: GITHUB


💡 一句话要点

提出PIVOT-Q以解决量化VLA的鲁棒性与泛化问题

🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱九:具身大模型 (Embodied Foundation Models)

关键词: 量化模型 鲁棒性 泛化能力 脆弱性导向 策略蒸馏 环境变化 深度学习 强化学习

📋 核心要点

  1. 现有的后训练量化方法在保持VLA性能方面存在不足,未能有效保证模型的鲁棒性和泛化能力。
  2. 本文提出了PIVOT-Q框架,通过选择性干预脆弱状态来恢复量化模型的泛化能力,利用全精度策略作为教师进行指导。
  3. 实验结果表明,PIVOT-Q在多个VLA骨干和量化方法下均表现出一致的恢复效果,相较于全状态蒸馏,使用的状态级蒸馏预算仅为7.4%。

📝 摘要(中文)

后训练量化已被证明在标准评估条件下能够保持VLA性能,但其对全精度模型的鲁棒性和泛化能力的影响尚未深入研究。本研究系统性地探讨了后量化VLA策略在环境干扰下的鲁棒性和泛化能力。实验证明,尽管量化策略在分布内性能上保持相似,但对细微环境变化却变得脆弱。我们观察到,动作差异集中在小部分状态上,而教师指导对差异的影响则相反:在高差异状态下改善泛化,但在低差异状态下可能降低泛化。因此,提出了基于脆弱性导向的调优框架PIVOT-Q,选择性地修正量化学生在回滚中遇到的脆弱状态,使用冻结的全精度策略作为教师。

🔬 方法详解

问题定义:本研究旨在解决后训练量化VLA策略在环境变化下的鲁棒性和泛化能力不足的问题。现有方法未能有效应对细微环境变化导致的脆弱性,影响了模型的实际应用。

核心思路:论文提出的PIVOT-Q框架通过识别和选择性修正脆弱状态,利用全精度策略作为教师进行指导,从而提高量化模型的泛化能力。这样的设计旨在针对性地解决模型在特定状态下的脆弱性,而非全局蒸馏。

技术框架:PIVOT-Q框架包括几个主要模块:首先,识别脆弱状态,使用折扣累积差异进行评估;其次,应用相位平衡稀疏监督进行干预;最后,使用行为锚点防止不必要的变化。整个流程强调了选择性干预的重要性。

关键创新:PIVOT-Q的关键创新在于其脆弱性导向的调优策略,区别于传统的全状态蒸馏方法,能够更有效地提升量化模型的泛化能力。

关键设计:在设计中,采用了折扣累积差异来识别脆弱状态,使用相位平衡稀疏监督来优化干预过程,并引入行为锚点以确保模型的稳定性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果显示,PIVOT-Q在七个LIBERO-Plus环境变化下均实现了一致的恢复效果,相较于全状态蒸馏,PIVOT-Q在所有设置中均表现出更优的性能,且仅使用了7.4%的状态级蒸馏预算,显示出显著的效率提升。

🎯 应用场景

该研究的潜在应用领域包括自动驾驶、机器人控制和智能制造等需要高鲁棒性和泛化能力的系统。通过提高量化模型的性能,PIVOT-Q能够在实际环境中更好地适应变化,提高系统的可靠性和安全性,具有重要的实际价值和未来影响。

📄 摘要(原文)

Post-training quantization has been shown to preserve VLA performance under standard evaluation conditions, but whether it preserves the full-precision model's robustness and generalization remains underexplored. In this study, we systematically study the robustness and generalization of post-quantized VLA policies under environmental disturbances. Empirical results show that quantized policies can become fragile to subtle environmental variations despite retaining comparable in-distribution performance. We further observe that action discrepancies are concentrated in a small subset of rollout states, while teacher guidance has opposite effects depending on discrepancy: it improves generalization at high-discrepancy states but can degrade it at low-discrepancy states. These findings reveal that effective post-quantization recovery requires selectively intervening on vulnerable states rather than globally distilling the student. We therefore propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), a vulnerability-aware On-Policy Distillation (OPD) framework that selectively corrects vulnerable states encountered during quantized-student rollouts using the frozen full-precision policy as a teacher. PIVOT-Q identifies vulnerable states using discounted accumulated discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to prevent unnecessary changes. Experiments under seven LIBERO-Plus environmental variations demonstrate consistent recovery across multiple VLA backbones and quantization methods. Notably, PIVOT-Q consistently outperforms full-state distillation across all settings while using only 7.4% of its state-level distillation budget. Our code is available at https://github.com/ruanruan-andy/PIVOT-Q.