Kuangpu Guo

Yongcan Yu, Lingxiao He, Jian Liang et al.

Test-time reinforcement learning (TTRL) always adapts models at inference time via pseudo-labeling, leaving it vulnerable to spurious optimization signals from label noise. Through an empirical study, we observe that responses with medium consistency form an ambiguity region and constitute the primary source of reward noise. Crucially, we find that such spurious signals can be even amplified through group-relative advantage estimation. Motivated by these findings, we propose a unified framework, Debiased and Denoised test-time Reinforcement Learning (DDRL), to mitigate spurious signals. Concretely, DDRL first applies a frequency-based sampling strategy to exclude ambiguous samples while maintaining a balanced set of positive and negative examples. It then adopts a debiased advantage estimation with fixed advantages, removing the bias introduced by group-relative policy optimization. Finally, DDRL incorporates a consensus-based off-policy refinement stage, which leverages the rejection-sampled dataset to enable efficient and stable model updates. Experiments on three large language models across multiple mathematical reasoning benchmarks demonstrate that DDRL consistently outperforms existing TTRL baselines. The code will soon be released at https://github.com/yuyongcan/DDRL.

Kuangpu Guo, Qingyan Zheng, Jian Liang et al.

The training-free integration of expert models via model merging has exposed significant security risks, enabling free-riders to combine specialized models without authorization. Recent works propose parameter-level defenses that employ linear parameter transformations to neutralize this threat. In this paper, we systematically analyze such defenses and reveal that their protected task vectors are inherently small in magnitude. Consequently, the protected weights remain overwhelmingly dominated by the pretrained model. Based on this observation, we designate the pretrained model as a static reference anchor and propose the Anchor-Guided Attack (AGA) to circumvent existing safeguards. Specifically, AGA aligns the protected model with this anchor to recover the transformation matrix analytically. Extensive evaluations validate that AGA consistently bypasses both individual and composite defenses under realistic defense-agnostic scenarios. Furthermore, we provide Anchor-Repulsive Fine-tuning (ARF), a defense method to mitigate the anchor dominance leveraged by AGA. Empirical results confirm that ARF effectively defeats the proposed attack. Our code is available at https://github.com/krumpguo/secure-merge-attack.

Kuangpu Guo, Yuhe Ding, Jian Liang et al.

Label skews, characterized by disparities in local label distribution across clients, pose a significant challenge in federated learning. As minority classes suffer from worse accuracy due to overfitting on local imbalanced data, prior methods often incorporate class-balanced learning techniques during local training. Although these methods improve the mean accuracy across all classes, we observe that vacant classes-referring to categories absent from a client's data distribution-remain poorly recognized. Besides, there is still a gap in the accuracy of local models on minority classes compared to the global model. This paper introduces FedVLS, a novel approach to label-skewed federated learning that integrates both vacant-class distillation and logit suppression simultaneously. Specifically, vacant-class distillation leverages knowledge distillation during local training on each client to retain essential information related to vacant classes from the global model. Moreover, logit suppression directly penalizes network logits for non-label classes, effectively addressing misclassifications in minority classes that may be biased toward majority classes. Extensive experiments validate the efficacy of FedVLS, demonstrating superior performance compared to previous state-of-the-art (SOTA) methods across diverse datasets with varying degrees of label skews. Our code is available at https://github.com/krumpguo/FedVLS.

Liang Kuang, Kuangpu Guo, Jian Liang et al.

Federated Learning (FL) allows collaborative machine learning training without sharing private data. Numerous studies have shown that one significant factor affecting the performance of federated learning models is the heterogeneity of data across different clients, especially when the data is sampled from various domains. A recent paper introduces variance-aware dual-level prototype clustering and uses a novel $α$-sparsity prototype loss, which increases intra-class similarity and reduces inter-class similarity. To ensure that the features converge within specific clusters, we introduce an improved algorithm, Federated Prototype Learning with Convergent Clusters, abbreviated as FedPLCC. To increase inter-class distances, we weight each prototype with the size of the cluster it represents. To reduce intra-class distances, considering that prototypes with larger distances might come from different domains, we select only a certain proportion of prototypes for the loss function calculation. Evaluations on the Digit-5, Office-10, and DomainNet datasets show that our method performs better than existing approaches.

Yuguang Zhang, Kuangpu Guo, Zhihe Lu et al.

Federated learning (FL) enables collaborative model training across decentralized clients without sharing local data, but is challenged by heterogeneity in data, computation, and communication. Pretrained vision-language models (VLMs), with their strong generalization and lightweight tuning via prompts, offer a promising solution. However, existing federated prompt-learning methods rely only on text prompts and overlook joint label-domain distribution shifts. In this paper, we propose a personalized FL framework based on dual-prompt learning and cross fusion, termed pFedDC. Specifically, each client maintains both global and local prompts across vision and language modalities: global prompts capture common knowledge shared across the federation, while local prompts encode client-specific semantics and domain characteristics. Meanwhile, a cross-fusion module is designed to adaptively integrate prompts from different levels, enabling the model to generate personalized representations aligned with each client's unique data distribution. Extensive experiments across nine datasets with various types of heterogeneity show that pFedDC consistently outperforms state-of-the-art methods.

Yongcan Yu, Lingxiao He, Shuo Lu et al.

Recent advances in vision-language models (VLMs) reasoning have been largely attributed to the rise of reinforcement Learning (RL), which has shifted the community's focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM Reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and strongly influenced by model capacity, data scale, and data distribution. Contrary to common assumptions, our findings show that SFT plays a crucial role across several scenarios: (1) Effectiveness for weaker models. SFT more reliably elicits reasoning capabilities in smaller or weaker VLMs. (2) Data efficiency. SFT with only 2K achieves comparable or better reasoning performance to RL with 20K. (3) Cross-modal transferability. SFT demonstrates stronger generalization across modalities. Moreover, we identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL. These results challenge the prevailing "RL over SFT" narrative. They highlight that the role of SFT may have been underestimated and support a more balanced post-training pipeline in which SFT and RL function as complementary components.