Large Language Models
Abstract
Recent advances in large language models (LLMs) have significantly improved the performance of dialog systems, yet current approaches often fail to provide accurate guidance of topic due to their inability to discern user confusion in related concepts. To address this, we introduce the Ask-Good-Question (AGQ) framework, which features an improved Concept-Enhanced Item Response Theory (CEIRT) model to better identify users' knowledge levels. Our contributions include applying the CEIRT model along with LLMs to directly generate guiding questions based on the inspiring text, greatly improving information retrieval efficiency during the question & answer process. Through comparisons with other baseline methods, our approach outperforms by significantly enhencing the users' information retrieval experiences.
Recent advances in large language models (LLMs) have significantly improved the performance of dialog systems, yet current approaches often fail to provide accurate guidance of topic due to their inability to discern user confusion in related concepts. To address this, we introduce the Ask-Good-Question (AGQ) framework, which features an improved Concept-Enhanced Item Response Theory (CEIRT) model to better identify users' knowledge levels. Our contributions include applying the CEIRT model along with LLMs to directly generate guiding questions based on the inspiring text, greatly improving information retrieval efficiency during the question & answer process. Through comparisons with other baseline methods, our approach outperforms by significantly enhencing the users' information retrieval experiences.
Abstract
In recent years, with the rapid development of the depth and breadth of large language models' capabilities, various corresponding evaluation benchmarks have been emerging in increasing numbers. As a quantitative assessment tool for model performance, benchmarks are not only a core means to measure model capabilities but also a key element in guiding the direction of model development and promoting technological innovation. We systematically review the current status and development of large language model benchmarks for the first time, categorizing 283 representative benchmarks into three categories: general capabilities, domain-specific, and target-specific. General capability benchmarks cover aspects such as core linguistics, knowledge, and reasoning; domain-specific benchmarks focus on fields like natural sciences, humanities and social sciences, and engineering technology; target-specific benchmarks pay attention to risks, reliability, agents, etc. We point out that current benchmarks have problems such as inflated scores caused by data contamination, unfair evaluation due to cultural and linguistic biases, and lack of evaluation on process credibility and dynamic environments, and provide a referable design paradigm for future benchmark innovation.
In recent years, with the rapid development of the depth and breadth of large language models' capabilities, various corresponding evaluation benchmarks have been emerging in increasing numbers. As a quantitative assessment tool for model performance, benchmarks are not only a core means to measure model capabilities but also a key element in guiding the direction of model development and promoting technological innovation. We systematically review the current status and development of large language model benchmarks for the first time, categorizing 283 representative benchmarks into three categories: general capabilities, domain-specific, and target-specific. General capability benchmarks cover aspects such as core linguistics, knowledge, and reasoning; domain-specific benchmarks focus on fields like natural sciences, humanities and social sciences, and engineering technology; target-specific benchmarks pay attention to risks, reliability, agents, etc. We point out that current benchmarks have problems such as inflated scores caused by data contamination, unfair evaluation due to cultural and linguistic biases, and lack of evaluation on process credibility and dynamic environments, and provide a referable design paradigm for future benchmark innovation.
Deep Learning
Abstract
Lyapunov functions are fundamental to establishing the stability of Markovian models, yet their construction typically demands substantial creativity and analytical effort. In this paper, we show that deep learning can automate this process by training neural networks to satisfy integral equations derived from first-transition analysis. Beyond stability analysis, our approach can be adapted to solve Poisson's equation and estimate stationary distributions. While neural networks are inherently function approximators on compact domains, it turns out that our approach remains effective when applied to Markov chains on non-compact state spaces. We demonstrate the effectiveness of this methodology through several examples from queueing theory and beyond.
Lyapunov functions are fundamental to establishing the stability of Markovian models, yet their construction typically demands substantial creativity and analytical effort. In this paper, we show that deep learning can automate this process by training neural networks to satisfy integral equations derived from first-transition analysis. Beyond stability analysis, our approach can be adapted to solve Poisson's equation and estimate stationary distributions. While neural networks are inherently function approximators on compact domains, it turns out that our approach remains effective when applied to Markov chains on non-compact state spaces. We demonstrate the effectiveness of this methodology through several examples from queueing theory and beyond.
Abstract
We propose an efficient hybrid least squares/gradient descent method to accelerate DeepONet training. Since the output of DeepONet can be viewed as linear with respect to the last layer parameters of the branch network, these parameters can be optimized using a least squares (LS) solve, and the remaining hidden layer parameters are updated by means of gradient descent form. However, building the LS system for all possible combinations of branch and trunk inputs yields a prohibitively large linear problem that is infeasible to solve directly. To address this issue, our method decomposes the large LS system into two smaller, more manageable subproblems $\unicode{x2014}$ one for the branch network and one for the trunk network $\unicode{x2014}$ and solves them separately. This method is generalized to a broader type of $L^2$ loss with a regularization term for the last layer parameters, including the case of unsupervised learning with physics-informed loss.
We propose an efficient hybrid least squares/gradient descent method to accelerate DeepONet training. Since the output of DeepONet can be viewed as linear with respect to the last layer parameters of the branch network, these parameters can be optimized using a least squares (LS) solve, and the remaining hidden layer parameters are updated by means of gradient descent form. However, building the LS system for all possible combinations of branch and trunk inputs yields a prohibitively large linear problem that is infeasible to solve directly. To address this issue, our method decomposes the large LS system into two smaller, more manageable subproblems $\unicode{x2014}$ one for the branch network and one for the trunk network $\unicode{x2014}$ and solves them separately. This method is generalized to a broader type of $L^2$ loss with a regularization term for the last layer parameters, including the case of unsupervised learning with physics-informed loss.
Multimodal Learning
Abstract
Taxonomic classification in biodiversity research involves organizing biological specimens into structured hierarchies based on evidence, which can come from multiple modalities such as images and genetic information. We investigate whether hyperbolic networks can provide a better embedding space for such hierarchical models. Our method embeds multimodal inputs into a shared hyperbolic space using contrastive and a novel stacked entailment-based objective. Experiments on the BIOSCAN-1M dataset show that hyperbolic embedding achieves competitive performance with Euclidean baselines, and outperforms all other models on unseen species classification using DNA barcodes. However, fine-grained classification and open-world generalization remain challenging. Our framework offers a structure-aware foundation for biodiversity modelling, with potential applications to species discovery, ecological monitoring, and conservation efforts.
Taxonomic classification in biodiversity research involves organizing biological specimens into structured hierarchies based on evidence, which can come from multiple modalities such as images and genetic information. We investigate whether hyperbolic networks can provide a better embedding space for such hierarchical models. Our method embeds multimodal inputs into a shared hyperbolic space using contrastive and a novel stacked entailment-based objective. Experiments on the BIOSCAN-1M dataset show that hyperbolic embedding achieves competitive performance with Euclidean baselines, and outperforms all other models on unseen species classification using DNA barcodes. However, fine-grained classification and open-world generalization remain challenging. Our framework offers a structure-aware foundation for biodiversity modelling, with potential applications to species discovery, ecological monitoring, and conservation efforts.
Abstract
This paper presents a novel generative framework for learning shared latent representations across multimodal data. Many advanced multimodal methods focus on capturing all combinations of modality-specific details across inputs, which can inadvertently obscure the high-level semantic concepts that are shared across modalities. Notably, Multimodal VAEs with low-dimensional latent variables are designed to capture shared representations, enabling various tasks such as joint multimodal synthesis and cross-modal inference. However, multimodal VAEs often struggle to design expressive joint variational posteriors and suffer from low-quality synthesis. In this work, ShaLa addresses these challenges by integrating a novel architectural inference model and a second-stage expressive diffusion prior, which not only facilitates effective inference of shared latent representation but also significantly improves the quality of downstream multimodal synthesis. We validate ShaLa extensively across multiple benchmarks, demonstrating superior coherence and synthesis quality compared to state-of-the-art multimodal VAEs. Furthermore, ShaLa scales to many more modalities while prior multimodal VAEs have fallen short in capturing the increasing complexity of the shared latent space.
This paper presents a novel generative framework for learning shared latent representations across multimodal data. Many advanced multimodal methods focus on capturing all combinations of modality-specific details across inputs, which can inadvertently obscure the high-level semantic concepts that are shared across modalities. Notably, Multimodal VAEs with low-dimensional latent variables are designed to capture shared representations, enabling various tasks such as joint multimodal synthesis and cross-modal inference. However, multimodal VAEs often struggle to design expressive joint variational posteriors and suffer from low-quality synthesis. In this work, ShaLa addresses these challenges by integrating a novel architectural inference model and a second-stage expressive diffusion prior, which not only facilitates effective inference of shared latent representation but also significantly improves the quality of downstream multimodal synthesis. We validate ShaLa extensively across multiple benchmarks, demonstrating superior coherence and synthesis quality compared to state-of-the-art multimodal VAEs. Furthermore, ShaLa scales to many more modalities while prior multimodal VAEs have fallen short in capturing the increasing complexity of the shared latent space.
Diffusion Models
Abstract
Diffusion-based generative models have emerged as highly effective methods for synthesizing high-quality samples. Recent works have focused on analyzing the convergence of their generation process with minimal assumptions, either through reverse SDEs or Probability Flow ODEs. The best known guarantees, without any smoothness assumptions, for the KL divergence so far achieve a linear dependence on the data dimension $d$ and an inverse quadratic dependence on $\varepsilon$. In this work, we present a refined analysis that improves the dependence on $\varepsilon$. We model the generation process as a composition of two steps: a reverse ODE step, followed by a smaller noising step along the forward process. This design leverages the fact that the ODE step enables control in Wasserstein-type error, which can then be converted into a KL divergence bound via noise addition, leading to a better dependence on the discretization step size. We further provide a novel analysis to achieve the linear $d$-dependence for the error due to discretizing this Probability Flow ODE in absence of any smoothness assumptions. We show that $\tilde{O}\left(\tfrac{d\log^{3/2}(\frac{1}{\delta})}{\varepsilon}\right)$ steps suffice to approximate the target distribution corrupted with Gaussian noise of variance $\delta$ within $O(\varepsilon^2)$ in KL divergence, improving upon the previous best result, requiring $\tilde{O}\left(\tfrac{d\log^2(\frac{1}{\delta})}{\varepsilon^2}\right)$ steps.
Diffusion-based generative models have emerged as highly effective methods for synthesizing high-quality samples. Recent works have focused on analyzing the convergence of their generation process with minimal assumptions, either through reverse SDEs or Probability Flow ODEs. The best known guarantees, without any smoothness assumptions, for the KL divergence so far achieve a linear dependence on the data dimension $d$ and an inverse quadratic dependence on $\varepsilon$. In this work, we present a refined analysis that improves the dependence on $\varepsilon$. We model the generation process as a composition of two steps: a reverse ODE step, followed by a smaller noising step along the forward process. This design leverages the fact that the ODE step enables control in Wasserstein-type error, which can then be converted into a KL divergence bound via noise addition, leading to a better dependence on the discretization step size. We further provide a novel analysis to achieve the linear $d$-dependence for the error due to discretizing this Probability Flow ODE in absence of any smoothness assumptions. We show that $\tilde{O}\left(\tfrac{d\log^{3/2}(\frac{1}{\delta})}{\varepsilon}\right)$ steps suffice to approximate the target distribution corrupted with Gaussian noise of variance $\delta$ within $O(\varepsilon^2)$ in KL divergence, improving upon the previous best result, requiring $\tilde{O}\left(\tfrac{d\log^2(\frac{1}{\delta})}{\varepsilon^2}\right)$ steps.
Abstract
Diffusion models typically inject isotropic Gaussian noise, disregarding structure in the data. Motivated by the way quantum squeezed states redistribute uncertainty according to the Heisenberg uncertainty principle, we introduce Squeezed Diffusion Models (SDM), which scale noise anisotropically along the principal component of the training distribution. As squeezing enhances the signal-to-noise ratio in physics, we hypothesize that scaling noise in a data-dependent manner can better assist diffusion models in learning important data features. We study two configurations: (i) a Heisenberg diffusion model that compensates the scaling on the principal axis with inverse scaling on orthogonal directions and (ii) a standard SDM variant that scales only the principal axis. Counterintuitively, on CIFAR-10/100 and CelebA-64, mild antisqueezing - i.e. increasing variance on the principal axis - consistently improves FID by up to 15% and shifts the precision-recall frontier toward higher recall. Our results demonstrate that simple, data-aware noise shaping can deliver robust generative gains without architectural changes.
Diffusion models typically inject isotropic Gaussian noise, disregarding structure in the data. Motivated by the way quantum squeezed states redistribute uncertainty according to the Heisenberg uncertainty principle, we introduce Squeezed Diffusion Models (SDM), which scale noise anisotropically along the principal component of the training distribution. As squeezing enhances the signal-to-noise ratio in physics, we hypothesize that scaling noise in a data-dependent manner can better assist diffusion models in learning important data features. We study two configurations: (i) a Heisenberg diffusion model that compensates the scaling on the principal axis with inverse scaling on orthogonal directions and (ii) a standard SDM variant that scales only the principal axis. Counterintuitively, on CIFAR-10/100 and CelebA-64, mild antisqueezing - i.e. increasing variance on the principal axis - consistently improves FID by up to 15% and shifts the precision-recall frontier toward higher recall. Our results demonstrate that simple, data-aware noise shaping can deliver robust generative gains without architectural changes.
Deep Learning Optimization
Abstract
While traditional Deep Learning (DL) optimization methods treat all training samples equally, Distributionally Robust Optimization (DRO) adaptively assigns importance weights to different samples. However, a significant gap exists between DRO and current DL practices. Modern DL optimizers require adaptivity and the ability to handle stochastic gradients, as these methods demonstrate superior performance. Additionally, for practical applications, a method should allow weight assignment not only to individual samples, but also to groups of objects (for example, all samples of the same class). This paper aims to bridge this gap by introducing ALSO $\unicode{x2013}$ Adaptive Loss Scaling Optimizer $\unicode{x2013}$ an adaptive algorithm for a modified DRO objective that can handle weight assignment to sample groups. We prove the convergence of our proposed algorithm for non-convex objectives, which is the typical case for DL models. Empirical evaluation across diverse Deep Learning tasks, from Tabular DL to Split Learning tasks, demonstrates that ALSO outperforms both traditional optimizers and existing DRO methods.
While traditional Deep Learning (DL) optimization methods treat all training samples equally, Distributionally Robust Optimization (DRO) adaptively assigns importance weights to different samples. However, a significant gap exists between DRO and current DL practices. Modern DL optimizers require adaptivity and the ability to handle stochastic gradients, as these methods demonstrate superior performance. Additionally, for practical applications, a method should allow weight assignment not only to individual samples, but also to groups of objects (for example, all samples of the same class). This paper aims to bridge this gap by introducing ALSO $\unicode{x2013}$ Adaptive Loss Scaling Optimizer $\unicode{x2013}$ an adaptive algorithm for a modified DRO objective that can handle weight assignment to sample groups. We prove the convergence of our proposed algorithm for non-convex objectives, which is the typical case for DL models. Empirical evaluation across diverse Deep Learning tasks, from Tabular DL to Split Learning tasks, demonstrates that ALSO outperforms both traditional optimizers and existing DRO methods.
Mixture of Experts
Abstract
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 $\href{https://github.com/dongbw18/MaxScore.git}{MaxScore}$.
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 $\href{https://github.com/dongbw18/MaxScore.git}{MaxScore}$.
Abstract
Emerging expert-specialized Mixture-of-Experts (MoE) architectures, such as DeepSeek-MoE, deliver strong model quality through fine-grained expert segmentation and large top-k routing. However, their scalability is limited by substantial activation memory overhead and costly all-to-all communication. Furthermore, current MoE training systems - primarily optimized for NVIDIA GPUs - perform suboptimally on non-NVIDIA platforms, leaving significant computational potential untapped. In this work, we present X-MoE, a novel MoE training system designed to deliver scalable training performance for next-generation MoE architectures. X-MoE achieves this via several novel techniques, including efficient padding-free MoE training with cross-platform kernels, redundancy-bypassing dispatch, and hybrid parallelism with sequence-sharded MoE blocks. Our evaluation on the Frontier supercomputer, powered by AMD MI250X GPUs, shows that X-MoE scales DeepSeek-style MoEs up to 545 billion parameters across 1024 GPUs - 10x larger than the largest trainable model with existing methods under the same hardware budget, while maintaining high training throughput. The source code of X-MoE is available at https://github.com/Supercomputing-System-AI-Lab/X-MoE.
Emerging expert-specialized Mixture-of-Experts (MoE) architectures, such as DeepSeek-MoE, deliver strong model quality through fine-grained expert segmentation and large top-k routing. However, their scalability is limited by substantial activation memory overhead and costly all-to-all communication. Furthermore, current MoE training systems - primarily optimized for NVIDIA GPUs - perform suboptimally on non-NVIDIA platforms, leaving significant computational potential untapped. In this work, we present X-MoE, a novel MoE training system designed to deliver scalable training performance for next-generation MoE architectures. X-MoE achieves this via several novel techniques, including efficient padding-free MoE training with cross-platform kernels, redundancy-bypassing dispatch, and hybrid parallelism with sequence-sharded MoE blocks. Our evaluation on the Frontier supercomputer, powered by AMD MI250X GPUs, shows that X-MoE scales DeepSeek-style MoEs up to 545 billion parameters across 1024 GPUs - 10x larger than the largest trainable model with existing methods under the same hardware budget, while maintaining high training throughput. The source code of X-MoE is available at https://github.com/Supercomputing-System-AI-Lab/X-MoE.
Interests not found
We did not find any papers that match the below interests.
Try other terms also consider if the content exists in arxiv.org.
- Deep Learning Architectures
- Deep Learning Models
You can edit or add more interests any time.
Help us improve your experience!
This project is on its early stages your feedback can be pivotal on the future of the project.
Let us know what you think about this week's papers and suggestions!
Give Feedback