arXiv subjects

Shimin Shan

Publications and source records attributed to Shimin Shan.

6 recordsLinked to original sources

arXiv2026

LISA: Linear-Indexed Sparse Attention for Efficient Long-Context Reasoning

Recent advances in long chain-of-thought reasoning models such as DeepSeek-R1 have led to increasingly longer inference context lengths under the test-time scaling paradigm. However, the O(n^2) computational complexity of standard self-attention causes inference costs to grow sharply with long sequences, limiting the deployment of long-CoT reasoning in production settings. To address this, we propose LISA (Linear-Indexed Sparse Attention), a plug-and-play attention replacement module that requires no pretraining from scratch. LISA integrates two lightweight components in parallel within the original model: (1) a Linear Attention module that provides long-range memory with O(n) time complexity; (2) a Lightning Indexer that selects the top-M important tokens from the full context to feed into a Sparse Self-Attention. The two branches are fused via a gating mechanism, reducing inference complexity from O(n^2) to O(nM) (M << n) for generating n tokens. We design a two-stage training pipeline: Stage 1 initializes the model by integrating the linear attention to capture long-range dependencies, complemented by a sliding-window attention mechanism that is optimized via knowledge distillation to approximate the full self-attention distribution of a frozen teacher model. In Stage 2, we further introduce the Indexer to replace the static sliding-window mechanism, enabling dynamic token selection from broader contexts. The Indexer is trained using a novel per-head KL divergence loss, which aligns its selection behavior with the attention patterns of the teacher model. Experiments on DeepSeek-distilled-Qwen models demonstrate that LISA achieves a 50% inference speedup under 16K-token context, while improving average performance by 5.6% on reasoning benchmarks including AIME and MATH-500.

cs.AI↗

arXiv2024

How to Train Text Summarization Model with Weak Supervisions

Currently, machine learning techniques have seen significant success across various applications. Most of these techniques rely on supervision from human-generated labels or a mixture of noisy and imprecise labels from multiple sources. However, for certain complex tasks, even noisy or inexact labels are unavailable due to the intricacy of the objectives. To tackle this issue, we propose a method that breaks down the complex objective into simpler tasks and generates supervision signals for each one. We then integrate these supervision signals into a manageable form, resulting in a straightforward learning procedure. As a case study, we demonstrate a system used for topic-based summarization. This system leverages rich supervision signals to promote both summarization and topic relevance. Remarkably, we can train the model end-to-end without any labels. Experimental results indicate that our approach performs exceptionally well on the CNN and DailyMail datasets.

cs.CL↗

arXiv2024

Accelerated Markov Chain Monte Carlo Using Adaptive Weighting Scheme

Gibbs sampling is one of the most commonly used Markov Chain Monte Carlo (MCMC) algorithms due to its simplicity and efficiency. It cycles through the latent variables, sampling each one from its distribution conditional on the current values of all the other variables. Conventional Gibbs sampling is based on the systematic scan (with a deterministic order of variables). In contrast, in recent years, Gibbs sampling with random scan has shown its advantage in some scenarios. However, almost all the analyses of Gibbs sampling with the random scan are based on uniform selection of variables. In this paper, we focus on a random scan Gibbs sampling method that selects each latent variable non-uniformly. Firstly, we show that this non-uniform scan Gibbs sampling leaves the target posterior distribution invariant. Then we explore how to determine the selection probability for latent variables. In particular, we construct an objective as a function of the selection probability and solve the constrained optimization problem. We further derive an analytic solution of the selection probability, which can be estimated easily. Our algorithm relies on the simple intuition that choosing the variable updates according to their marginal probabilities enhances the mixing time of the Markov chain. Finally, we validate the effectiveness of the proposed Gibbs sampler by conducting a set of experiments on real-world applications.

cs.LG↗

arXiv2024

BioMamba: Domain-Adaptive Biomedical Language Models

Background. Biomedical language models should improve performance on biomedical text while retaining general-language-modeling fluency. For Mamba-based models, this trade-off has not been systematically studied across biomedical literature and clinical text. Methods. We developed BioMamba, a family of biomedical Mamba2 models at five scales obtained by continued pretraining of released public Mamba2 checkpoints on a balanced 80%/10%/10% mixture of PubMed abstracts, the Colossal Clean Crawled Corpus (C4), and Wikipedia. The contribution is the adaptation recipe and the accompanying open-weight checkpoints. Results. Across five scales, BioMamba consistently lowered PubMed perplexity, improved Wikipedia-style held-out perplexity by 1.46-4.72 PPL, and left C4 perplexity essentially unchanged. On six out-of-domain multiple-choice benchmarks, BioMamba stayed within +/-3 percentage points of Mamba2 with no systematic regression. After supervised fine-tuning, BioMamba+SFT matched or exceeded Mamba2+SFT on MIMIC-IV note completion and discharge summary generation at every evaluated scale, and improved PubMedQA at every scale. The strongest model (BioMamba-2.7B) reached a PubMed perplexity of 5.28 and accuracies of 90.24% and 73.00% on BioASQ and PubMedQA, respectively. Conclusions. A balanced domain-adaptive continued pretraining recipe strengthens Mamba2 language models on biomedical literature and clinical text while preserving general-language-modeling fluency.

cs.CL↗

arXiv2024

AntibodyFlow: Normalizing Flow Model for Designing Antibody Complementarity-Determining Regions

Therapeutic antibodies have been extensively studied in drug discovery and development in the past decades. Antibodies are specialized protective proteins that bind to antigens in a lock-to-key manner. The binding strength/affinity between an antibody and a specific antigen is heavily determined by the complementarity-determining regions (CDRs) on the antibodies. Existing machine learning methods cast in silico development of CDRs as either sequence or 3D graph (with a single chain) generation tasks and have achieved initial success. However, with CDR loops having specific geometry shapes, learning the 3D geometric structures of CDRs remains a challenge. To address this issue, we propose AntibodyFlow, a 3D flow model to design antibody CDR loops. Specifically, AntibodyFlow first constructs the distance matrix, then predicts amino acids conditioned on the distance matrix. Also, AntibodyFlow conducts constraint learning and constrained generation to ensure valid 3D structures. Experimental results indicate that AntibodyFlow outperforms the best baseline consistently with up to 16.0% relative improvement in validity rate and 24.3% relative reduction in geometric graph level error (root mean square deviation, RMSD).

cs.LG↗

arXiv2024

PGA-SciRE: Harnessing LLM on Data Augmentation for Enhancing Scientific Relation Extraction

Relation Extraction (RE) aims at recognizing the relation between pairs of entities mentioned in a text. Advances in LLMs have had a tremendous impact on NLP. In this work, we propose a textual data augmentation framework called PGA for improving the performance of models for RE in the scientific domain. The framework introduces two ways of data augmentation, utilizing a LLM to obtain pseudo-samples with the same sentence meaning but with different representations and forms by paraphrasing the original training set samples. As well as instructing LLM to generate sentences that implicitly contain information about the corresponding labels based on the relation and entity of the original training set samples. These two kinds of pseudo-samples participate in the training of the RE model together with the original dataset, respectively. The PGA framework in the experiment improves the F1 scores of the three mainstream models for RE within the scientific domain. Also, using a LLM to obtain samples can effectively reduce the cost of manually labeling data.

cs.CL↗