Jiazhen Huang

Jiazhen Huang, Xiao Chen, Xiao Luo et al.

On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes into dense token-level supervision. Existing methods usually assume trusted PI, such as reference answers or successful traces. We ask whether PI can instead come from an experience-derived skill bank, where retrieved skills are compact and reusable but may also be irrelevant or misleading. We propose Skill-Conditioned Gated Self-Distillation (SGSD), which formulates skill-based SD as teacher hypothesis validation rather than unconditional imitation. SGSD retrieves skill-mistake pairs, constructs a multi-teacher pool, and lets all skill-conditioned teachers score the same plain-prompt student rollout. The verifier validates each teacher's polarity: supporting a success or suppressing a failure gives positive supervision, while the opposite stance is reversed. A robust gated objective then distills informative teacher-student disagreements while suppressing uncertain or extreme signals. Experiments on multiple mathematical reasoning benchmarks show that SGSD consistently improves over GRPO and remains competitive with answer-conditioned OPSD under a weaker PI assumption. For example, on Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and OPSD by 1.7% on average on AIME24, AIME25, and HMMT25. Our code is available at https://github.com/walawalagoose/SGSD.

Xiao Chen, Zhongjing Du, Jiazhen Huang et al.

Test-Time Adaptation (TTA) enhances model robustness to out-of-distribution (OOD) data by updating the model online during inference, yet existing methods lack theoretical insights into the fundamental causes of performance degradation under domain shifts. Recently, Neural Collapse (NC) has been proposed as an emergent geometric property of deep neural networks (DNNs), providing valuable insights for TTA. In this work, we extend NC to the sample-wise level and discover a novel phenomenon termed Sample-wise Alignment Collapse (NC3+), demonstrating that a sample's feature embedding, obtained by a trained model, aligns closely with the corresponding classifier weight. Building on NC3+, we identify that the performance degradation stems from sample-wise misalignment in adaptation which exacerbates under larger distribution shifts. This indicates the necessity of realigning the feature embeddings with their corresponding classifier weights. However, the misalignment makes pseudo-labels unreliable under domain shifts. To address this challenge, we propose NCTTA, a novel feature-classifier alignment method with hybrid targets to mitigate the impact of unreliable pseudo-labels, which blends geometric proximity with predictive confidence. Extensive experiments demonstrate the effectiveness of NCTTA in enhancing robustness to domain shifts. For example, NCTTA outperforms Tent by 14.52% on ImageNet-C.

Xiao Chen, Jiazhen Huang, Qinting Jiang et al.

Test-time adaptation (TTA) has emerged as a critical technique for enhancing the generalization capability of vision-language models (VLMs) during inference. However, existing approaches often incur substantial computational costs and exhibit poor scalability, primarily due to sample-wise adaptation granularity and reliance on costly auxiliary designs such as data augmentation. To address these limitations, we introduce SAIL (Small Aid, Big Leap), a novel adapter-based TTA framework that leverages a lightweight, learnable AdaptNet to enable efficient and scalable model adaptation. As SAIL's core, a frozen pre-trained VLM collaborates with AdaptNet through a confidence-based interpolation weight, generating robust predictions during inference. These predictions serve as self-supervised targets to align AdaptNet's outputs through efficient batch-wise processing, dramatically reducing computational costs without modifying the VLM or requiring memory caches. To mitigate catastrophic forgetting during continual adaptation, we propose a gradient-aware reset strategy driven by a gradient drift indicator (GDI), which dynamically detects domain transitions and strategically resets AdaptNet for stable adaptation. Extensive experiments across diverse benchmarks on two scenarios demonstrate that SAIL achieves state-of-the-art performance while maintaining low computational costs. These results highlight SAIL's effectiveness, efficiency and scalability for real-world deployment. The code will be released upon acceptance.