Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. Some works conduct layer-level hybrid structures that combine Transformer and Mamba layers, aiming to make full use of both advantages. This paper proposes TransMamba, a novel sequence-level hybrid framework that unifies Transformer and Mamba through shared parameter matrices (QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory Converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for balancing effectiveness and efficiency. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to single and hybrid baselines, and validated the deeper consistency between Transformer and Mamba paradigms at sequence level, offering a scalable solution for next-generation language modeling.

Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Kun Li, Haoyuan Wu, Wujiajia, Zihao Zheng, Chenchen Zhang, Kun Shi, Xue Gong, Qi Yi, Ruibin Xiong, Tingqiang Xu, Yuhao Jiang, Jianfeng Yan, Yuyuan Zeng, Guanghui Xu, Jinbao Xue, Zhijiang xu, Zheng Fang, Shuai LI, Qibin Liu, Xiaoxue Li, Zhuoyu Li, Yangyu Tao, Fei Gao, Cheng Jiang, Bochao Wang, Kai Liu, Jianchen Zhu, Wai Lam, Bo Zhou, Di Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Mixture-of-Experts (MoE) models rely on an external router to assign tokens to experts.This design inherently separates the routing decision from each expert's internal capabilities, leading to suboptimal performance.In this work, we address this limitation with Union-of-Experts (UoE), an MoE variant that performs "expertautonomous routing".The core mechanism of UoE is to pre-designate a minute fraction of neurons within each expert as routing neurons.Experts autonomously select relevant tokens by comparing the activation intensity of these neurons, aligning routing decisions with each expert's functional profile.To prevent the waste of activations from unselected experts, we aggregate all routing neuron outputs and sum them into the final layer output.This aggregation acts as a novel virtual shared expert whose parameters are distributed across the individual experts, and improves overall parameter efficiency.We pre-train UoE models with up to 3B parameters, demonstrating that they outperform traditional MoEs with matched efficiency.Furthermore, our analysis of the routing neurons provides valuable insights into expert-autonomous selection and advances the understanding of MoE routing.

An Wang, Xingwu Sun, Ruobing Xie, Shuaipeng Li, Jiaqi Zhu, Zhen Yang, Pinxue Zhao, Weidong Han, Zhanhui Kang, Di Wang, Naoaki Okazaki, Cheng-zhong Xu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Mixture-of-Experts (MoE) models mostly use a router to assign tokens to specific expert modules, activating only partial parameters and often outperforming dense models. We argue that the separation between the router’s decision-making and the experts’ execution is a critical yet overlooked issue, leading to suboptimal expert selection and learning. To address this, we propose Autonomy-of-Expert (AoE), a novel MoE paradigm in which experts autonomously select themselves to process inputs. AoE is based on the insight that an expert is aware of its own capacity to effectively process a token, an awareness reflected in the scale of its internal activations. In AoE, routers are removed; instead, experts pre-compute internal activations for inputs and are ranked based on their activation norms. Only the top-ranking experts proceed with the forward pass, while the others abort. The overhead of pre-computing activations is reduced through a low-rank weight factorization. This self-evaluating-then-partner-comparing approach ensures improved expert selection and effective learning. We pre-train language models having 700M up to 4B parameters, demonstrating that AoE outperforms traditional MoE models with comparable efficiency.

Low-precision training is considered an effective strategy for reducing both training and downstream inference costs. Previous scaling laws for precision mainly focus on integer quantization, which pay less attention to the constituents in floating-point (FP) quantization, and thus cannot well fit the LLM losses in this scenario. In contrast, while FP quantization training is more commonly implemented in production, it's research has been relatively superficial. In this paper, we thoroughly explore the effects of FP quantization targets, exponent bits, mantissa bits, and the calculation granularity of the scaling factor in FP quantization training performance of LLM models. In addition to an accurate FP quantization unified scaling law, we also provide valuable suggestions for the community: (1) Exponent bits contribute slightly more to the model performance than mantissa bits. We provide the optimal exponent-mantissa bit ratio for different bit numbers, which is available for future reference by hardware manufacturers; (2) We discover the formation of the critical data size in low-precision LLM training. Too much training data exceeding the critical data size will inversely bring in degradation of LLM performance; (3) The optimal FP quantization precision is directly proportional to the computational power, but within a wide computational power range. We estimate that the best cost-performance precision should lie between 4-8 bits.

Recent advances in large vision-language models (LVLMs) have showcased their remarkable capabilities across a wide range of multimodal vision-language tasks. However, these models remain vulnerable to visual adversarial attacks, which can substantially compromise their performance. In this paper, we introduce F3, a novel adversarial purification framework that employs a counterintuitive ''fighting fire with fire'' strategy: intentionally introducing simple perturbations to adversarial examples to mitigate their harmful effects. Specifically, F3 leverages cross-modal attentions derived from randomly perturbed adversary examples as reference targets. By injecting noise into these adversarial examples, F3 effectively refines their attention, resulting in cleaner and more reliable model outputs. Remarkably, this seemingly paradoxical approach of employing noise to counteract adversarial attacks yields impressive purification results. Furthermore, F3 offers several distinct advantages: it is training-free and straightforward to implement, and exhibits significant computational efficiency improvements compared to existing purification methods. These attributes render F3 particularly suitable for large-scale industrial applications where both robust performance and operational efficiency are critical priorities. The code is available at https://github.com/btzyd/F3.

Large vision-language models (LVLMs) have demonstrated exceptional performance on complex multimodal tasks. However, they continue to suffer from significant hallucination issues, including object, attribute, and relational hallucinations. To accurately detect these hallucinations, we investigated the variations in cross-modal attention patterns between hallucination and non-hallucination states. Leveraging these distinctions, we developed a lightweight detector capable of identifying hallucinations. Our proposed method, Detecting Hallucinations by Cross-modal Attention Patterns (DHCP), is straightforward and does not require additional LVLM training or extra LVLM inference steps. Experimental results show that DHCP achieves remarkable performance in hallucination detection. By offering novel insights into the identification and analysis of hallucinations in LVLMs, DHCP contributes to advancing the reliability and trustworthiness of these models. The code is available at https://github.com/btzyd/DHCP.

To optimize large Transformer model training, both efficient parallel computing and advanced data management are indispensable. However, current methods often assume a stable and uniform training workload, neglecting data-induced imbalances-arising from both sampling and packing processes-which can impede training performance. Specifically, data sampling imbalance arises from uneven sequence length distribution of the training data, while data packing imbalance stems from the discrepancy between the linear memory complexity and quadratic time complexity of the attention mechanism. To address these imbalance issues, we develop Hydraulis, which jointly optimizes the parallel strategies and data assignment. For one thing, we introduce large model training with dynamic heterogeneous parallel strategies in response to the sequence length variations within and across training iterations. For another, we devise a two-stage data assignment approach, which strikes a good balance in terms of the training workloads both within and across model replicas. Empirical results demonstrate that Hydraulis outperforms existing systems by 1.32-2.66×. Our source code is available: https://github.com/PKU-DAIR/Hetu.

Despite the great success of large language models (LLMs) in various tasks, they suffer from generating hallucinations. We introduce Truth Forest, a method that enhances truthfulness in LLMs by uncovering hidden truth representations using multi-dimensional orthogonal probes. Specifically, it creates multiple orthogonal bases for modeling truth by incorporating orthogonal constraints into the probes. Moreover, we introduce Random Peek, a systematic technique considering an extended range of positions within the sequence, reducing the gap between discerning and generating truth features in LLMs. By employing this approach, we improved the truthfulness of Llama-2-7B from 40.8% to 74.5% on TruthfulQA. Likewise, significant improvements are observed in fine-tuned models. We conducted a thorough analysis of truth features using probes. Our visualization results show that orthogonal probes capture complementary truth-related features, forming well-defined clusters that reveal the inherent structure of the dataset.

In current deep learning tasks, Adam-style optimizers—such as Adam, Adagrad, RMSprop, Adafactor, and Lion—have been widely used as alternatives to SGD-style optimizers. These optimizers typically update model parameters using the sign of gradients, resulting in more stable convergence curves. The learning rate and the batch size are the most critical hyperparameters for optimizers, which require careful tuning to enable effective convergence. Previous research has shown that the optimal learning rate increases linearly (or follows similar rules) with batch size for SGD-style optimizers. However, this conclusion is not applicable to Adam-style optimizers. In this paper, we elucidate the connection between optimal learning rates and batch sizes for Adam-style optimizers through both theoretical analysis and extensive experiments. First, we raise the scaling law between batch sizes and optimal learning rates in the “sign of gradient” case, in which we prove that the optimal learning rate first rises and then falls as the batch size increases. Moreover, the peak value of the surge will gradually move toward the larger batch size as training progresses. Second, we conduct experiments on various CV and NLP tasks and verify the correctness of the scaling law.

Shulin Liu, Tao Yang, Tianchi Yue, Feng Zhang, Di Wang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.