The attention mechanism is a fundamental component of the Transformer model, contributing to interactions among distinct tokens. In general, the attention scores are determined simply by the key-query products. However, this work’s occasional trial (combining DAPE and NoPE) of including additional MLPs on attention scores without position encoding indicates that the classical key-query multiplication may limit the performance of Transformers. In this work, we conceptualize attention as a feature map and apply the convolution operator (for neighboring attention scores across different heads) to mimic the processing methods in computer vision. Specifically, the main contribution of this paper is identifying and interpreting the Transformer length extrapolation problem as a result of the limited expressiveness of the naive query and key dot product, and we successfully translate the length extrapolation issue into a well-understood feature map processing problem, which is called Convolutional Data-Adaptive Position Encoding (CDAPE).The novel insight, which can be adapted to various attention-related models, reveals that the current Transformer architecture has the potential for further evolution. Extensive experiments demonstrate that treating attention as a feature map and applying convolution as a processing method significantly enhances Transformer performance.

The capacity of Large Language Models (LLMs) to comprehend and reason over long contexts is pivotal for advancements in diverse fields. Yet, they still stuggle with capturing long-distance dependencies within sequences to deeply understand semantics. To address this issue, we introduce Query-aware Inference for LLMs (Q-LLM), a system designed to process extensive sequences akin to human cognition. By focusing on memory data relevant to a given query, Q-LLM can accurately capture pertinent information within a fixed window size and provide precise answers to queries. It doesn’t require extra training and can be seamlessly integrated with any LLMs. Q-LLM using LLaMA3 (QuickLLaMA) can read Harry Potter within 30s and accurately answer the questions. On widely recognized benchmarks, Q-LLM improved by 7.17% compared to the current state-of-the-art on LLaMA3, and by 3.26% on Mistral on the \infty-bench. In the Needle-in-a-Haystack and BABILong task, Q-LLM improved upon the current SOTA by 7.0% and 6.1%. Our code is in https://github.com/dvlab-research/Q-LLM.

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research.Traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity.To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation.Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity.This ensures more reliable and effective training.We constructed a large-scale 1.2M logits dataset and trained a series of science-focused models.Experimental results demonstrate that our method achieves significant improvements, with accuracy gains of 18% on Mawps and 22.7% on TabMWP.Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%.Codes are available at https://github.com/dvlab- research/Logits-Based-Finetuning.

Jing Xiong, Jianghan Shen, Fanghua Ye, Chaofan Tao, Zhongwei Wan, Jianqiao Lu, Xun Wu, Chuanyang Zheng, Zhijiang Guo, Min Yang, Lingpeng Kong, Ngai Wong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Chuanyang Zheng, Yihang Gao, Guoxuan Chen, Han Shi, Jing Xiong, Xiaozhe Ren, Chao Huang, Zhenguo Li, Yu Li. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We present a new family of mobile hybrid vision networks, called iFormer, with a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network, \textit{i.e.}, ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks. Notably, iFormer achieves an impressive Top-1 accuracy of 80.4% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13, surpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios. Code and models are available at: https://github.com/ChuanyangZheng/iFormer.

Extrapolating ultra-long contexts (text length $>$128K) remains a major challenge for large language models (LLMs), as most training-free extrapolation methods are not only severely limited by memory bottlenecks, but also suffer from the attention sink, which restricts their scalability and effectiveness in practice. In this work, we propose ParallelComp, a parallel long-context compression method that effectively overcomes the memory bottleneck, enabling 8B-parameter LLMs to extrapolate from 8K to 128K tokens on a single A100 80GB GPU in a training-free setting. ParallelComp splits the input into chunks, dynamically evicting redundant chunks and irrelevant tokens, supported by a parallel KV cache eviction mechanism. Importantly, we present a systematic theoretical and empirical analysis of attention biases in parallel attention—including the attention sink, recency bias, and middle bias—and reveal that these biases exhibit distinctive patterns under ultra-long context settings. We further design a KV cache eviction technique to mitigate this phenomenon. Experimental results show that ParallelComp enables an 8B model (trained on 8K context) to achieve 91.17% of GPT-4’s performance under ultra-long contexts, outperforming closed-source models such as Claude-2 and Kimi-Chat. We achieve a 1.76x improvement in chunk throughput, thereby achieving a 23.50x acceleration in the prefill stage with negligible performance loss and pave the way for scalable and robust ultra-long contexts extrapolation in LLMs. We release the code at https://github.com/menik1126/ParallelComp.

The attention mechanism is a core component of the Transformer architecture. Various methods have been developed to compute attention scores, including multi-head attention (MHA), multi-query attention, group-query attention and so on. We further analyze the MHA and observe that its performance improves as the number of attention heads increases, provided the hidden size per head remains sufficiently large. Therefore, increasing both the head count and hidden size per head with minimal parameter overhead can lead to significant performance gains at a low cost. Motivated by this insight, we introduce Simulated Attention Score (SAS), which **maintains a compact model size while simulating a larger number of attention heads and hidden feature dimension per head.** This is achieved by projecting a low-dimensional head representation into a higher-dimensional space, effectively increasing attention capacity without increasing parameter count. Beyond the head representations, we further extend the simulation approach to feature dimension of the key and query embeddings, enhancing expressiveness by mimicking the behavior of a larger model while preserving the original model size. **To control the parameter cost, we also propose Parameter-Efficient Attention Aggregation (PEAA).** Comprehensive experiments on a variety of datasets and tasks demonstrate the effectiveness of the proposed SAS method, achieving significant improvements over different attention variants.

Vision-language (VL) models have achieved unprece-dented success recently, in which the connection module is the key to bridge the modality gap. Nevertheless, the abun-dant visual clues are not sufficiently exploited in most existing methods. On the vision side, most existing approaches only use the last feature of the vision tower, without using the low-level features. On the language side, most existing meth-ods only introduce shallow vision-language interactions. In this paper, we present a vision-inspired vision-language con-nection module, dubbed as VIVL, which efficiently exploits the vision cue for VL models. To take advantage of the lower-level information from the vision tower, a feature pyramid extractor (FPE) is introduced to combine features from differ-ent intermediate layers, which enriches the visual cue with negligible parameters and computation overhead. To en-hance VL interactions, we propose deep vision-conditioned prompts (DVCP) that allows deep interactions of vision and language features efficiently. Our VIVL exceeds the previous state-of-the-art method by 18.1 CIDEr when training from scratch on the COCO caption task, which greatly improves the data efficiency. When used as a plug-in module, VIVL consistently improves the performance for various backbones and VL frameworks, delivering new state-of-the-art results on multiple benchmarks, e.g., NoCaps and VQAv2.

Despite the success of large language models (LLMs), the task of theorem proving still remains one of the hardest reasoning tasks that is far from being fully solved. Prior methods using language models have demonstrated promising results, but they still struggle to prove even middle school level theorems. One common limitation of these methods is that they assume a fixed theorem library during the whole theorem proving process. However, as we all know, creating new useful theorems or even new theories is not only helpful but crucial and necessary for advancing mathematics and proving harder and deeper results. In this work, we present LEGO-Prover, which employs a growing skill library containing verified lemmas as skills to augment the capability of LLMs used in theorem proving. By constructing the proof modularly, LEGO-Prover enables LLMs to utilize existing skills retrieved from the library and to create new skills during the proving process. These skills are further evolved (by prompting an LLM) to enrich the library on another scale. Modular and reusable skills are constantly added to the library to enable tackling increasingly intricate mathematical problems. Moreover, the learned library further bridges the gap between human proofs and formal proofs by making it easier to impute missing steps. LEGO-Prover advances the state-of-the-art pass rate on miniF2F-valid (48.0\% to 57.0\%) and miniF2F-test (45.5\% to 50.0\%). During the proving process, LEGO-Prover also generates over 20,000 skills (theorems/lemmas) and adds them to the growing library. Our ablation study indicates that these newly added skills are indeed helpful for proving theorems, resulting in a 4.9\% improvement in success rate

Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning. Dual Queries first query LLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question's knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5\% to 94.2\%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundaries of in-context learning and opens up new avenues for addressing complex reasoning challenges.

Positional encoding plays a crucial role in transformers, significantly impact- ing model performance and length generalization. Prior research has introduced absolute positional encoding (APE) and relative positional encoding (RPE) to distinguish token positions in given sequences. However, both APE and RPE remain fixed after model training regardless of input data, limiting their adaptability and flexibility. Hence, we expect that the desired positional encoding should be data-adaptive and can be dynamically adjusted with the given attention. In this paper, we propose a Data-Adaptive Positional Encoding (DAPE) method, which dynamically and semantically adjusts based on input context and learned fixed priors. Experimental validation on real-world datasets (Arxiv, Books3, and CHE) demonstrates that DAPE enhances model performances in terms of trained length and length generalization, where the improvements are statistically significant. The model visualization suggests that our model can keep both local and anti-local information. Finally, we successfully train the model on sequence length 128 and achieve better performance at evaluation sequence length 8192, compared with other static positional encoding methods, revealing the benefit of the adaptive positional encoding method.

Single-cell RNA sequencing has achieved massive success in biological research fields. Discovering novel cell types from single-cell transcriptomics has been demonstrated to be essential in the field of biomedicine, yet is time-consuming and needs prior knowledge. With the unprecedented boom in cell atlases, auto-annotation tools have become more prevalent due to their speed, accuracy and user-friendly features. However, existing tools have mostly focused on general cell-type annotation and have not adequately addressed the challenge of discovering novel rare cell types. In this work, we introduce scNovel, a powerful deep learning-based neural network that specifically focuses on novel rare cell discovery. By testing our model on diverse datasets with different scales, protocols and degrees of imbalance, we demonstrate that scNovel significantly outperforms previous state-of-the-art novel cell detection models, reaching the most AUROC performance(the only one method whose averaged AUROC results are above 94%, up to 16.26% more comparing to the second-best method). We validate scNovel's performance on a million-scale dataset to illustrate the scalability of scNovel further. Applying scNovel on a clinical COVID-19 dataset, three potential novel subtypes of Macrophages are identified, where the COVID-related differential genes are also detected to have consistent expression patterns through deeper analysis. We believe that our proposed pipeline will be an important tool for high-throughput clinical data in a wide range of applications.

Jing Xiong, Jianhao Shen, Ye Yuan, Haiming Wang, Yichun Yin, Zhengying Liu, Lin Li, Zhijiang Guo, Qingxing Cao, Yinya Huang, Chuanyang Zheng, Xiaodan Liang, Ming Zhang, Qun Liu. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Vision Transformers (ViTs) yield impressive performance across various vision tasks. However, heavy computation and memory footprint make them inaccessible for edge devices. Previous works apply importance criteria determined independently by each individual component to prune ViTs. Considering that heterogeneous components in ViTs play distinct roles, these approaches lead to suboptimal performance. In this paper, we introduce joint importance, which integrates essential structural-aware interactions between components for the first time, to perform collaborative pruning. Based on the theoretical analysis, we construct a Taylor-based approximation to evaluate the joint importance. This guides pruning toward a more balanced reduction across all components. To further reduce the algorithm complexity, we incorporate the interactions into the optimization function under some mild assumptions. Moreover, the proposed method can be seamlessly applied to various tasks including object detection. Extensive experiments demonstrate the effectiveness of our method. Notably, the proposed approach outperforms the existing state-of-the-art approaches on ImageNet, increasing accuracy by 0.7% over the DeiT-Base baseline while saving 50% FLOPs. On COCO, we are the first to show that 70% FLOPs of FasterRCNN with ViT backbone can be removed with only 0.3% mAP drop. The code is available at https://github.com/hikvision-research/SAViT.