Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT

2026-05-09

AI Summary

Unconstrained fine-tuning often induces catastrophic forgetting in flow-matching vision-language-action (VLA) models, degrading their foundational capabilities. This work proposes Conservative Supervised Fine-Tuning (ConSFT), inspired by trust-region methods in reinforcement learning, which dynamically scales the learning signal according to model confidence to enable sparse yet precise parameter updates. ConSFT effectively balances task-specific adaptation with the preservation of pre-existing abilities—without requiring historical data, architectural modifications, or explicit regularization. Evaluated on the LIBERO and RoboTwin benchmarks, ConSFT improves average foundational capability retention by over 20% compared to standard fine-tuning, matching the performance of methods that rely on experience replay, while also mitigating spatial overfitting in real-world robotic deployments.

Abstract

Unconstrained fine-tuning of flow-matching Vision-Language-Action (VLA) models drives dense parameter overwrites, degrading pre-trained capabilities. We present Conservative Supervised Fine-Tuning (ConSFT), an optimization objective that adapts to target distributions while mitigating catastrophic forgetting, requiring zero prior data or architectural overhead. By dynamically scaling learning signals based on model confidence, ConSFT suppresses excessive gradients from low-confidence samples to prevent disproportionate parameter updates, thereby bounding the intrinsic parameter disruption risk. Inspired by reinforcement learning's trust-region clipping, this formulation establishes a progressive learning dynamic to secure target convergence and prior capability retention, maintaining sparse parameter updates without relying on the parallel reference networks required by explicit regularization. We evaluate ConSFT on the LIBERO and RoboTwin benchmarks across state-of-the-art flow-matching VLAs ($π_0$, $π_{0.5}$, and GR00T-N1.6-3B). The method outperforms vanilla SFT in capability retention by an average absolute margin of over 20\%, matching the efficacy of data-heavy Experience Replay in a prior-data-free regime. Real-world robotic deployments confirm that ConSFT precludes spatial overfitting during downstream adaptation, preserving pre-trained physical skills while acquiring sequential target tasks.

Problem

• flow-matching
• Vision-Language-Action
• catastrophic forgetting
• fine-tuning
• parameter disruption

Innovation

• Conservative SFT
• Flow-Matching VLA
• Catastrophic Forgetting
• Trust-Region Optimization
• Parameter Update Sparsity