Tengyu Pan

Zhenyu Li, Yike Zhang, Tengyu Pan et al. · tsinghua

Empowering LLMs with the ability to precisely understand long contexts is crucial for many downstream applications. However, handling long contexts with conventional transformer architecture requires substantial training and inference resources. Existing context condensing methods cannot accurately understand the full context, as there is a considerable amount of information loss in the condensing process. To address these issues, we present FocusLLM, a framework designed to extend the fixed context length of any decoder-only LLM, allowing the model to focus on relevant information from very long sequences. FocusLLM first divides long text input into chunks based on the model's original context length. It then employs the dynamic condensing process to distill crucial information from each chunk. Ultimately, through the novel parallel decoding mechanism, FocusLLM can integrate the extracted information into its local context. FocusLLM stands out for great training efficiency and versatility: trained with an 8K input length and with much less training cost than previous methods, FocusLLM exhibits superior performance across downstream tasks and maintains strong language modeling ability when handling extensive long texts, even up to 400K tokens. Our code is available at https://github.com/leezythu/FocusLLM.

Weilin Zhao, Tengyu Pan, Xu Han et al. · tsinghua

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12$\times$ speedup over the state-of-the-art speculative sampling method EAGLE-2. Code available at https://github.com/thunlp/FR-Spec.

Bowen Dong, Yilong Fan, Yutao Sun et al.

Routing networks in sparsely activated mixture-of-experts (MoE) dynamically allocate input tokens to top-k experts through differentiable sparse transformations, enabling scalable model capacity while preserving computational efficiency. Traditional MoE networks impose an expert capacity constraint to ensure GPU-friendly computation. However, this leads to token dropping when capacity is saturated and results in low hardware efficiency due to padding in underutilized experts. Removing the capacity constraint, in turn, compromises load balancing and computational efficiency. To address these issues, we propose Maximum Score Routing ($\mathbf{MaxScore}$), a novel MoE routing paradigm that models routing as a minimum-cost maximum-flow problem and integrates a SoftTopk operator. MaxScore resolves the fundamental limitations of iterative rerouting and optimal transport formulations, achieving lower training losses and higher evaluation scores at equivalent FLOPs compared to both constrained and unconstrained baselines. Implementation details and experimental configurations can be obtained from $\href{https://github.com/dongbw18/MaxScore.git}{MaxScore}$.

Yutao Sun, Zhenyu Li, Yike Zhang et al.

Transformer-based architectures have become the prevailing backbone of large language models. However, the quadratic time and memory complexity of self-attention remains a fundamental obstacle to efficient long-context modeling. To address this limitation, recent research has introduced two principal categories of efficient attention mechanisms. Linear attention methods achieve linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics, thereby enabling scalable inference with reduced computational overhead. Sparse attention techniques, in contrast, limit attention computation to selected subsets of tokens based on fixed patterns, block-wise routing, or clustering strategies, enhancing efficiency while preserving contextual coverage. This survey provides a systematic and comprehensive overview of these developments, integrating both algorithmic innovations and hardware-level considerations. In addition, we analyze the incorporation of efficient attention into largescale pre-trained language models, including both architectures built entirely on efficient attention and hybrid designs that combine local and global components. By aligning theoretical foundations with practical deployment strategies, this work aims to serve as a foundational reference for advancing the design of scalable and efficient language models.

Tengyu Pan, Zhichao Duan, Zhenyu Li et al.

Text embedding models are essential for various natural language processing tasks, enabling the effective encoding of semantic information into dense vector representations. These models are typically optimized using triplets of (query, positive, negative) data pairs for contrastive learning, where the negative samples play a critical role in enhancing the model's ability to discern subtle semantic distinctions. In this work, we introduce a Multi-Granularity Hard-negative (MGH) synthesis framework that leverages large language models (LLMs) to generate diverse negative samples with varying levels of similarity with the query. This approach facilitates a coarse-to-fine curriculum learning strategy during supervised training, allowing the embedding model to progressively learn more nuanced semantic representations. Meanwhile, we propose an Anchor Token Aware (ATA) pooling method that assigns higher weights to anchor tokens based on aggregation patterns observed in LLMs, improving text embedding accuracy without increasing model complexity. Comprehensive experiments on the MTEB benchmark demonstrate that our methods achieve state-of-the-art performance, surpassing existing synthesis strategies both with synthetic data and when combined with public retrieval datasets.

Ziyu Wang, Tengyu Pan, Zhenyu Li et al.

fMRI (functional Magnetic Resonance Imaging) visual decoding involves decoding the original image from brain signals elicited by visual stimuli. This often relies on manually labeled ROIs (Regions of Interest) to select brain voxels. However, these ROIs can contain redundant information and noise, reducing decoding performance. Additionally, the lack of automated ROI labeling methods hinders the practical application of fMRI visual decoding technology, especially for new subjects. This work presents TROI (Trainable Region of Interest), a novel two-stage, data-driven ROI labeling method for cross-subject fMRI decoding tasks, particularly when subject samples are limited. TROI leverages labeled ROIs in the dataset to pretrain an image decoding backbone on a cross-subject dataset, enabling efficient optimization of the input layer for new subjects without retraining the entire model from scratch. In the first stage, we introduce a voxel selection method that combines sparse mask training and low-pass filtering to quickly generate the voxel mask and determine input layer dimensions. In the second stage, we apply a learning rate rewinding strategy to fine-tune the input layer for downstream tasks. Experimental results on the same small sample dataset as the baseline method for brain visual retrieval and reconstruction tasks show that our voxel selection method surpasses the state-of-the-art method MindEye2 with an annotated ROI mask.

Zhichao Duan, Tengyu Pan, Zhenyu Li et al.

Document-level relation extraction (DocRE) is the process of identifying and extracting relations between entities that span multiple sentences within a document. Due to its realistic settings, DocRE has garnered increasing research attention in recent years. Previous research has mostly focused on developing sophisticated encoding models to better capture the intricate patterns between entity pairs. While these advancements are undoubtedly crucial, an even more foundational challenge lies in the data itself. The complexity inherent in DocRE makes the labeling process prone to errors, compounded by the extreme sparsity of positive relation samples, which is driven by both the limited availability of positive instances and the broad diversity of positive relation types. These factors can lead to biased optimization processes, further complicating the task of accurate relation extraction. Recognizing these challenges, we have developed a robust framework called \textit{\textbf{COMM}} to better solve DocRE. \textit{\textbf{COMM}} operates by initially employing an instance-aware reasoning method to dynamically capture pertinent information of entity pairs within the document and extract relational features. Following this, \textit{\textbf{COMM}} takes into account the distribution of relations and the difficulty of samples to dynamically adjust the margins between prediction logits and the decision threshold, a process we call Concentrated Margin Maximization. In this way, \textit{\textbf{COMM}} not only enhances the extraction of relevant relational features but also boosts DocRE performance by addressing the specific challenges posed by the data. Extensive experiments and analysis demonstrate the versatility and effectiveness of \textit{\textbf{COMM}}, especially its robustness when trained on low-quality data (achieves \textgreater 10\% performance gains).