Long Chain-of-Thought (CoT) reasoning enhances large reasoning models' performance but suffers from severe inefficiencies, as models often overthink simple problems or underthink complex ones. Current sequence-level optimizations, like length penalties, are too coarse-grained to distinguish core logic from verbose language, precluding the necessary token-level control for efficient reasoning CoT. To overcome these limitations, we introduce Time-Frequency token Advantage Clipping (TFAC), a novel training framework designed to build efficient large reasoning models via token-level interventions. Specifically, TFAC functions along two dimensions: 1) The Frequency Dimension: It discourages inefficient loops and encourages deeper exploration by dynamically reducing the advantage scores of high-entropy tokens that are repeatedly generated within a single reasoning path. 2) The Time Dimension: It reduces excessive overthinking of the system by establishing a historical baseline for the occurrence count of each critical token in previously successful trajectories, and clipping the advantages of tokens that exceed this baseline during training. Crucially, to preserve the model's exploratory capabilities on novel problems, this suppression mechanism is automatically disabled when no historical record of success is available. Experiments conducted on the Deepseek-Distill-32B and Qwen3-8B models show that TFAC outperforms leading baseline methods, improving performance by 2.3 and 3.1 percentage points, respectively, while simultaneously reducing inference costs by 35% and 28% in scenarios where correct answers are generated. These results validate the significant efficacy of TFAC in training large reasoning models that are both powerful and highly efficient.

The safety mechanisms of large language models (LLMs) exhibit notable fragility, as even fine-tuning on datasets without harmful content may still undermine their safety capabilities.Meanwhile, existing safety alignment methods predominantly rely on the fine-tuning process, which inadvertently leads to the increased complexity and computational resources required.To address these issues, we introduce LSSF, a novel safety re-alignment framework with Low-Rank Safety Subspace Fusison.Our proposed method exploits the low-rank characteristics of safety information in LLMs by constructing a low-rank projection matrix to extract the principal components of safety vectors.Notably, this projection matrix represents the low-rank safety subspace of the LLMs, which we have observed to remain stable during fine-tuning process and is isolated from the model's general capabilities.These principal components are used to effectively restore safety alignment when combined with fine-tuned LLMs through linear arithmetic.Additionally, to account for the varying encoding densities of safety information across different layers of LLMs, we propose a novel metric called safety singular value entropy.This metric quantifies the encoding density and allows for the dynamic computation of the safety-critical rank for each safety vector.Extensive experiments demonstrate that our proposed post-hoc alignment method can effectively restore the safety alignment of finetuned models with minimal impact on their performance in downstream tasks.

Continual learning for large language models (LLMs) demands a precise balance between $\textbf{plasticity}$ - the ability to absorb new tasks - and $\textbf{stability}$ - the preservation of previously learned knowledge. Conventional rehearsal methods, which replay stored examples, are limited by long-term data inaccessibility; earlier pseudo-rehearsal methods require additional generation modules, while self-synthesis approaches often generate samples that poorly align with real tasks, suffer from unstable outputs, and ignore task relationships. We present $\textbf{\textit{Self-Evolving Pseudo-Rehearsal for Catastrophic Forgetting with Task Similarity}}(\textbf{SERS})$, a lightweight framework that 1) decouples pseudo-input synthesis from label creation, using semantic masking and template guidance to produce diverse, task-relevant prompts without extra modules; 2) applies label self-evolution, blending base-model priors with fine-tuned outputs to prevent over-specialization; and 3) introduces a dynamic regularizer driven by the Wasserstein distance between task distributions, automatically relaxing or strengthening constraints in proportion to task similarity. Experiments across diverse tasks on different LLMs show that our SERS reduces forgetting by over 2\% points against strong pseudo-rehearsal baselines, by ensuring efficient data utilization and wisely transferring knowledge. The code will be released at https://github.com/JerryWangJun/LLM_CL_SERS/.

Deep clustering is a crucial task in machine learning and data mining that focuses on acquiring feature representations conducive to clustering. Previous research relies on self-supervised representation learning for general feature representations, such features may not be optimally suited for downstream clustering tasks. In this article, we introduce MICCF, a framework designed to bridge this gap and enhance clustering performance. MICCF enhances feature representations by combining mutual information constraints at different levels and employs an auxiliary alignment mutual information module for learning clustering-oriented features. To be specific, we propose a dual mutual information constraints module, incorporating minimal mutual information constraints at the feature level and maximal mutual information constraints at the instance level. This reduction in feature redundancy encourages the neural network to extract more discriminative features, while maximization ensures more unbiased and robust representations. To obtain clustering-oriented representations, the auxiliary alignment mutual information module utilizes pseudo-labels to maximize mutual information through a multi-classifier network, aligning features with the clustering task. The main network and the auxiliary module work in synergy to jointly optimize feature representations that are well-suited for the clustering task. We validate the effectiveness of our method through extensive experiments on six benchmark datasets. The results indicate that our method performs well in most scenarios, particularly on fine-grained datasets, where our approach effectively distinguishes subtle differences between closely related categories. Notably, our approach achieved a remarkable accuracy of 96.4% on the ImageNet-10 dataset, surpassing other comparison methods. The code is available at https://github.com/Li-Hyn/MICCF.git .

Deep clustering is a fundamental task in machine learning and data mining that aims at learning clustering-oriented feature representations. In previous studies, most of deep clustering methods follow the idea of self-supervised representation learning by maximizing the consistency of all similar instance pairs while ignoring the effect of feature redundancy on clustering performance. In this paper, to address the above issue, we design a dual mutual information constrained clustering method named DMICC which is based on deep contrastive clustering architecture, in which the dual mutual information constraints are particularly employed with solid theoretical guarantees and experimental validations. Specifically, at the feature level, we reduce the redundancy among features by minimizing the mutual information across all the dimensionalities to encourage the neural network to extract more discriminative features. At the instance level, we maximize the mutual information of the similar instance pairs to obtain more unbiased and robust representations. The dual mutual information constraints happen simultaneously and thus complement each other to jointly optimize better features that are suitable for the clustering task. We also prove that our adopted mutual information constraints are superior in feature extraction, and the proposed dual mutual information constraints are clearly bounded and thus solvable. Extensive experiments on five benchmark datasets show that our proposed approach outperforms most other clustering algorithms. The code is available at https://github.com/Li-Hyn/DMICC.