arXiv subjects

Zhan Su

Publications and source records attributed to Zhan Su.

36 records · Page 2Linked to original sources

arXiv2025

Tensorized Clustered LoRA Merging for Multi-Task Interference

Despite the success of the monolithic dense paradigm of large language models (LLMs), the LoRA adapters offer an efficient solution by fine-tuning small task-specific modules and merging them with the base model. However, in multi-task settings, merging LoRA adapters trained on heterogeneous sources frequently causes \textit{task interference}, degrading downstream performance. To address this, we propose a tensorized clustered LoRA (TC-LoRA) library targeting to address the task interference at the \textit{text-level} and \textit{parameter-level}. At the \textit{text-level}, we cluster the training samples in the embedding space to capture input-format similarities, then train a specialized LoRA adapter for each cluster. At the \textit{parameter-level}, we introduce a joint Canonical Polyadic (CP) decomposition that disentangles task-specific and shared factors across LoRA adapters. This joint factorization preserves essential knowledge while reducing cross-task interference. Extensive experiments on out-of-domain zero-shot and skill-composition tasks-including reasoning, question answering, and coding. Compared to strong SVD-based baselines, TC-LoRA achieves +1.4\% accuracy on Phi-3 and +2.3\% on Mistral-7B (+2.3\%), demonstrating the effectiveness of TC-LoRA in LLM adaptation.

cs.LG↗

arXiv2025

Smooth Reading: Bridging the Gap of Recurrent LLM to Self-Attention LLM on Long-Context Tasks

Recently, recurrent large language models (Recurrent LLMs) with linear computational complexity have re-emerged as efficient alternatives to self-attention-based LLMs (Self-Attention LLMs), which have quadratic complexity. However, Recurrent LLMs often underperform on long-context tasks due to their limited fixed-size memory. Previous research has primarily focused on enhancing the memory capacity of Recurrent LLMs through architectural innovations, but these approaches have not yet enabled Recurrent LLMs to match the performance of Self-Attention LLMs on long-context tasks. We argue that this limitation arises because processing the entire context at once is not well-suited for Recurrent LLMs. In this paper, we propose Smooth Reading, a chunk-wise inference method inspired by human reading strategies. Smooth Reading processes context in chunks and iteratively summarizes the contextual information, thereby reducing memory demands and making the approach more compatible with Recurrent LLMs. Our experimental results show that this method substantially narrows the performance gap between Recurrent and Self-Attention LLMs on long-context tasks, while preserving the efficiency advantages of Recurrent LLMs. Our Smooth Reading boosts SWA-3B-4k (a Recurrent LLM) from 5.68% lower to 3.61% higher performance than Self-Attention LLMs on LongBench. Besides, our method maintains the high efficiency, training 3x faster and inferring 2x faster at 64k context compared to Self-Attention LLMs. To our knowledge, this is the first work to achieve comparable performance using Recurrent LLMs compared with Self-Attention LLMs on long-context tasks. We hope our method will inspire future research in this area. To facilitate further progress, we will release code and dataset.

cs.CL↗

arXiv2025

Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts

Merging parameter-efficient task experts has recently gained growing attention as a way to build modular architectures that can be rapidly adapted on the fly for specific downstream tasks, without requiring additional fine-tuning. Typically, LoRA serves as the foundational building block of such parameter-efficient modular architectures, leveraging low-rank weight structures to reduce the number of trainable parameters. In this paper, we study the properties of sparse adapters, which train only a subset of weights in the base neural network, as potential building blocks of modular architectures. First, we propose a simple method for training highly effective sparse adapters, which is conceptually simpler than existing methods in the literature and surprisingly outperforms both LoRA and full fine-tuning in our setting. Next, we investigate the merging properties of these sparse adapters by merging adapters for up to 20 natural language processing tasks, thus scaling beyond what is usually studied in the literature. Our findings demonstrate that sparse adapters yield superior in-distribution performance post-merging compared to LoRA or full model merging. Achieving strong held-out performance remains a challenge for all methods considered.

cs.LG↗

arXiv2025

FusionAudio-1.2M: Towards Fine-grained Audio Captioning with Multimodal Contextual Fusion

High-quality, large-scale audio captioning is crucial for advancing audio understanding, yet current automated methods often generate captions that lack fine-grained detail and contextual accuracy, primarily due to their reliance on limited unimodal or superficial multimodal information. Drawing inspiration from human auditory perception, which adeptly integrates cross-modal cues and performs sophisticated auditory scene analysis, we introduce a novel two-stage automated pipeline. This pipeline first employs specialized pretrained models to extract diverse contextual cues (e.g., speech, music, general sounds, and visual information from associated video). A large language model (LLM) then synthesizes these rich, multimodal inputs to generate detailed and context-aware audio captions. Key contributions of this work include: (1) the proposed scalable method for fine-grained audio caption generation; (2) FusionAudio, a new large-scale dataset comprising 1.2 million such detailed captions, combined with 6 million QA pairs; and (3) enhanced audio models developed using FusionAudio, specifically a CLAP-based audio encoder with superior audio-text alignment and instruction following. This paper paves the way for more nuanced and accurate automated understanding of complex audio environments. Code and data can be found in https://github.com/satsuki2486441738/FusionAudio.

cs.SD↗

arXiv2025

Light-driven lattice metastability for enhanced superconductivity in FeSe/SrTiO3

Driven quantum materials with on demand properties controlled by external stimuli are critical for emergent quantum technology. In optically tunable superconducting heterostructures, the lattice responses at the buried interface may hold the key to the light susceptibility but is very challenging to detect. In this work, a nondestructive synchrotron-based X-ray scattering phase-retrieval technique is implemented in monolayer-FeSe/SrTiO3 heterostructures to capture the three-dimensional interfacial atomic displacements in-situ as the interface superconductivity is actively manipulated by light. It is found that the interlayer sliding between FeSe and SrTiO3 can drastically alter how the lattice responds to the light. In domains with selected stacking configurations, the interface transforms the very weak photoexcitation in SrTiO3 into significant Fe-atom displacements in FeSe and generate metastable interfacial structures that can lead to a persistent superconductivity enhancement. These findings demonstrate an effective strategy for achieving greatly amplified light-lattice coupling for efficient quantum phase manipulations at designed interfaces.

cond-mat.mtrl-sci↗

arXiv2025

SIP-IFVM: A time-evolving coronal model with an extended magnetic field decomposition strategy

Time-evolving magnetohydrodynamic (MHD) coronal modeling, driven by a series of time-dependent photospheric magnetograms, represents a new generation of coronal simulations. This approach offers greater realism compared to traditional coronal models constrained by a static magnetogram. However, its practical application is seriously limited by low computational efficiency and poor numerical stability. Therefore, we propose an extended magnetic field decomposition strategy and implement it in the implicit MHD model to develop a coronal model that is both efficient and numerically stable enough for simulating the long-term evolutions of the global corona. The traditional decomposition strategies split the magnetic field into a time-invariant potential field and a time-dependent component $\mathbf{B}_1$. It works well for quasi-steady-state coronal simulations where $\left|\mathbf{B}_1\right|$ is typically small. However, as the inner-boundary magnetic field evolves, $\left|\mathbf{B}_1\right|$ can grow significantly larger and its discretization errors often lead to nonphysical negative thermal pressure, ultimately causing the code to crash. In this paper, we mitigate such undesired situations by introducing a temporally piecewise-constant variable to accommodate part of the non-potential field and remain $\left|\mathbf{B}_1\right|$ consistently small throughout the simulations. We incorporate this novel magnetic field decomposition strategy into our implicit MHD coronal model and apply it to simulate the evolution of coronal structures within 0.1 AU over two solar-maximum Carrington rotations. The results show that this coronal model effectively captures observations and performs more than 80 times faster than real time using only 192 CPU cores, making it well-suited for practical applications in simulating the time-evolving corona.

astro-ph.SR↗

arXiv2024

Mixture of Latent Experts Using Tensor Products

In multi-task learning, the conventional approach involves training a model on multiple tasks simultaneously. However, the training signals from different tasks can interfere with one another, potentially leading to \textit{negative transfer}. To mitigate this, we investigate if modular language models can facilitate positive transfer and systematic generalization. Specifically, we propose a novel modular language model (\texttt{TensorPoly}), that balances parameter efficiency with nuanced routing methods. For \textit{modules}, we reparameterize Low-Rank Adaptation (\texttt{LoRA}) by employing an entangled tensor through the use of tensor product operations and name the resulting approach \texttt{TLoRA}. For \textit{routing function}, we tailor two innovative routing functions according to the granularity: \texttt{TensorPoly-I} which directs to each rank within the entangled tensor while \texttt{TensorPoly-II} offers a finer-grained routing approach targeting each order of the entangled tensor. The experimental results from the multi-task T0-benchmark demonstrate that: 1) all modular LMs surpass the corresponding dense approaches, highlighting the potential of modular language models to mitigate negative inference in multi-task learning and deliver superior outcomes. 2) \texttt{TensorPoly-I} achieves higher parameter efficiency in adaptation and outperforms other modular LMs, which shows the potential of our approach in multi-task transfer learning.

cs.LG↗

arXiv2024

Aligning Query Representation with Rewritten Query and Relevance Judgments in Conversational Search

Conversational search supports multi-turn user-system interactions to solve complex information needs. Different from the traditional single-turn ad-hoc search, conversational search encounters a more challenging problem of context-dependent query understanding with the lengthy and long-tail conversational history context. While conversational query rewriting methods leverage explicit rewritten queries to train a rewriting model to transform the context-dependent query into a stand-stone search query, this is usually done without considering the quality of search results. Conversational dense retrieval methods use fine-tuning to improve a pre-trained ad-hoc query encoder, but they are limited by the conversational search data available for training. In this paper, we leverage both rewritten queries and relevance judgments in the conversational search data to train a better query representation model. The key idea is to align the query representation with those of rewritten queries and relevant documents. The proposed model -- Query Representation Alignment Conversational Dense Retriever, QRACDR, is tested on eight datasets, including various settings in conversational search and ad-hoc search. The results demonstrate the strong performance of QRACDR compared with state-of-the-art methods, and confirm the effectiveness of representation alignment.

cs.IR↗

arXiv2024

History-Aware Conversational Dense Retrieval

Conversational search facilitates complex information retrieval by enabling multi-turn interactions between users and the system. Supporting such interactions requires a comprehensive understanding of the conversational inputs to formulate a good search query based on historical information. In particular, the search query should include the relevant information from the previous conversation turns. However, current approaches for conversational dense retrieval primarily rely on fine-tuning a pre-trained ad-hoc retriever using the whole conversational search session, which can be lengthy and noisy. Moreover, existing approaches are limited by the amount of manual supervision signals in the existing datasets. To address the aforementioned issues, we propose a History-Aware Conversational Dense Retrieval (HAConvDR) system, which incorporates two ideas: context-denoised query reformulation and automatic mining of supervision signals based on the actual impact of historical turns. Experiments on two public conversational search datasets demonstrate the improved history modeling capability of HAConvDR, in particular for long conversations with topic shifts.

cs.IR↗

arXiv2024

Towards Modular LLMs by Building and Reusing a Library of LoRAs

The growing number of parameter-efficient adaptations of a base large language model (LLM) calls for studying whether we can reuse such trained adapters to improve performance for new tasks. We study how to best build a library of adapters given multi-task data and devise techniques for both zero-shot and supervised task generalization through routing in such library. We benchmark existing approaches to build this library and introduce model-based clustering, MBC, a method that groups tasks based on the similarity of their adapter parameters, indirectly optimizing for transfer across the multi-task dataset. To re-use the library, we present a novel zero-shot routing mechanism, Arrow, which enables dynamic selection of the most relevant adapters for new inputs without the need for retraining. We experiment with several LLMs, such as Phi-2 and Mistral, on a wide array of held-out tasks, verifying that MBC-based adapters and Arrow routing lead to superior generalization to new tasks. We make steps towards creating modular, adaptable LLMs that can match or outperform traditional joint training.

cs.LG↗

arXiv2024

Language Modeling Using Tensor Trains

We propose a novel tensor network language model based on the simplest tensor network (i.e., tensor trains), called `Tensor Train Language Model' (TTLM). TTLM represents sentences in an exponential space constructed by the tensor product of words, but computing the probabilities of sentences in a low-dimensional fashion. We demonstrate that the architectures of Second-order RNNs, Recurrent Arithmetic Circuits (RACs), and Multiplicative Integration RNNs are, essentially, special cases of TTLM. Experimental evaluations on real language modeling tasks show that the proposed variants of TTLM (i.e., TTLM-Large and TTLM-Tiny) outperform the vanilla Recurrent Neural Networks (RNNs) with low-scale of hidden units. (The code is available at https://github.com/shuishen112/tensortrainlm.)

cs.CL↗

arXiv2023

Multi-Head Adapter Routing for Cross-Task Generalization

Parameter-efficient fine-tuning (PEFT) for cross-task generalization consists in pre-training adapters on a multi-task training set before few-shot adaptation to test tasks. Polytropon [Ponti et al., 2023] ($\texttt{Poly}$) jointly learns an inventory of adapters and a routing function that selects a (variable-size) subset of adapters for each task during both pre-training and few-shot adaptation. In this paper, we investigate the role that adapter routing plays in its success and design new variants based on our findings. First, we build on the intuition that finer-grained routing provides more expressivity. Hence, we propose $\texttt{MHR}$ (Multi-Head Routing) which combines subsets of adapter parameters and outperforms $\texttt{Poly}$ under a comparable parameter budget; by only fine-tuning the routing function and not the adapters ($\texttt{MHR}$-$z$) we achieve competitive performance with extreme parameter efficiency. Second, we find that $\texttt{Poly}$/$\texttt{MHR}$ performance is a result of better multi-task optimization, rather than modular inductive biases that facilitate adapter recombination and local adaptation, as previously hypothesized. In fact, we find that $\texttt{MHR}$ exhibits high gradient alignment between training tasks. We find that routing is most beneficial during multi-task pre-training rather than during few-shot adaptation and propose $\texttt{MHR}$-$μ$, which discards routing and fine-tunes the average of the pre-trained adapters on each downstream tasks. This establishes $\texttt{MHR}$-$μ$ as an effective method for single-adapter fine-tuning. We also show that $\texttt{MHR}$-$μ$ can be used as an effective zero-shot transfer method by training the average of the pre-trained adapters for a few additional steps on the multi-task training set: this yields gains up to 3% on absolute accuracy w.r.t. the baselines.

cs.AI↗

arXiv2019

A Generalized Language Model in Tensor Space

In the literature, tensors have been effectively used for capturing the context information in language models. However, the existing methods usually adopt relatively-low order tensors, which have limited expressive power in modeling language. Developing a higher-order tensor representation is challenging, in terms of deriving an effective solution and showing its generality. In this paper, we propose a language model named Tensor Space Language Model (TSLM), by utilizing tensor networks and tensor decomposition. In TSLM, we build a high-dimensional semantic space constructed by the tensor product of word vectors. Theoretically, we prove that such tensor representation is a generalization of the n-gram language model. We further show that this high-order tensor representation can be decomposed to a recursive calculation of conditional probability for language modeling. The experimental results on Penn Tree Bank (PTB) dataset and WikiText benchmark demonstrate the effectiveness of TSLM.

cs.CL↗

arXiv2018

A Quantum Many-body Wave Function Inspired Language Modeling Approach

The recently proposed quantum language model (QLM) aimed at a principled approach to modeling term dependency by applying the quantum probability theory. The latest development for a more effective QLM has adopted word embeddings as a kind of global dependency information and integrated the quantum-inspired idea in a neural network architecture. While these quantum-inspired LMs are theoretically more general and also practically effective, they have two major limitations. First, they have not taken into account the interaction among words with multiple meanings, which is common and important in understanding natural language text. Second, the integration of the quantum-inspired LM with the neural network was mainly for effective training of parameters, yet lacking a theoretical foundation accounting for such integration. To address these two issues, in this paper, we propose a Quantum Many-body Wave Function (QMWF) inspired language modeling approach. The QMWF inspired LM can adopt the tensor product to model the aforesaid interaction among words. It also enables us to reveal the inherent necessity of using Convolutional Neural Network (CNN) in QMWF language modeling. Furthermore, our approach delivers a simple algorithm to represent and match text/sentence pairs. Systematic evaluation shows the effectiveness of the proposed QMWF-LM algorithm, in comparison with the state of the art quantum-inspired LMs and a couple of CNN-based methods, on three typical Question Answering (QA) datasets.

cs.CL↗

arXiv2010

Tomography of the two-electron spin qubits in semiconductors

We propose an approach to reconstruct two-electron spin qubit states in semiconductor quantum dots by employing tomographic techniques. This procedure exploits the combination of fast gate operations on electron spins trapped in dots and dynamical nuclear polarization of the underlying Ga and As nuclei. The presented method can be an important tool in solid state quantum computation for complete characterization of qubit states and theirs correlations.

quant-ph↗

arXiv2010

A Bayesian Method for Detecting and Characterizing Allelic Heterogeneity and Boosting Signals in Genome-Wide Association Studies

The standard paradigm for the analysis of genome-wide association studies involves carrying out association tests at both typed and imputed SNPs. These methods will not be optimal for detecting the signal of association at SNPs that are not currently known or in regions where allelic heterogeneity occurs. We propose a novel association test, complementary to the SNP-based approaches, that attempts to extract further signals of association by explicitly modeling and estimating both unknown SNPs and allelic heterogeneity at a locus. At each site we estimate the genealogy of the case-control sample by taking advantage of the HapMap haplotypes across the genome. Allelic heterogeneity is modeled by allowing more than one mutation on the branches of the genealogy. Our use of Bayesian methods allows us to assess directly the evidence for a causative SNP not well correlated with known SNPs and for allelic heterogeneity at each locus. Using simulated data and real data from the WTCCC project, we show that our method (i) produces a significant boost in signal and accurately identifies the form of the allelic heterogeneity in regions where it is known to exist, (ii) can suggest new signals that are not found by testing typed or imputed SNPs and (iii) can provide more accurate estimates of effect sizes in regions of association.

stat.ME↗

arXiv2010

A Graphene Quantum Dot with a Single Electron Transistor as Integrated Charge Sensor

We have developed an etching process to fabricate a quantum dot and a nearby single electron transistor as a charge detector in a single layer graphene. The high charge sensitivity of the detector is used to probe Coulomb diamonds as well as excited spectrum in the dot, even in the regime where the current through the quantum dot is too small to be measured by conventional transport means. The graphene based quantum dot and integrated charge sensor serve as an essential building block to form a solid-state qubit in a nuclear-spin-free quantum world.

cond-mat.mes-hall↗

arXiv2008

Random Sierpinski network with scale-free small-world and modular structure

In this paper, we define a stochastic Sierpinski gasket, on the basis of which we construct a network called random Sierpinski network (RSN). We investigate analytically or numerically the statistical characteristics of RSN. The obtained results reveal that the properties of RSN is particularly rich, it is simultaneously scale-free, small-world, uncorrelated, modular, and maximal planar. All obtained analytical predictions are successfully contrasted with extensive numerical simulations. Our network representation method could be applied to study the complexity of some real systems in biological and information fields.

cond-mat.stat-mech↗