With the rising need for speech-based interaction models, end-to-end Spoken Language Models (SLMs) have emerged as a promising solution.While these models require comprehensive world knowledge for meaningful and reliable human interactions, existing questionanswering (QA) benchmarks fall short in evaluating SLMs' knowledge understanding due to their inability to support end-to-end speech evaluation and account for varied input audio conditions.To address these limitations, we present VoxEval, a novel SpeechQA benchmark that assesses SLMs' knowledge understanding through pure speech interactions.Our benchmark 1) uniquely maintains speech format for both inputs and outputs, 2) evaluates model robustness across diverse input audio conditions, and 3) pioneers the assessment of complex tasks like mathematical reasoning in spoken format.Systematic evaluation demonstrates that VoxEval presents significant challenges to current SLMs, revealing their sensitivity to varying audio conditions and highlighting the need to enhance reasoning capabilities in future development.We hope this benchmark could guide the advancement of more sophisticated and reliable SLMs. 1

Wenqian Cui, Dianzhi Yu, Xiaoqi Jiao, Ziqiao Meng, Guangyan Zhang, Qichao Wang, Steven Y. Guo, Irwin King. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Proteolysis-targeting chimeras (PROTACs) are a groundbreaking technology for targeted protein degradation, but designing effective linkers that connect two molecular fragments to form a drug-candidate PROTAC molecule remains a key challenge. While diffusion models show promise in molecular generation, current diffusion models for PROTAC linker design are typically trained on small molecule datasets, introducing distribution mismatches in the chemical space between small molecules and target PROTACs. Direct fine-tuning on limited PROTAC datasets often results in overfitting and poor generalization. In this work, we propose DAD-PROTAC, a domain-adapted diffusion model for PROTAC linker design, which addresses this distribution mismatch in chemical space through density ratio estimation to bridge the gap between small-molecule and PROTAC domains. By decomposing the target score estimator into a pre-trained score function and a lightweight score correction term, DAD-PROTAC achieves efficient fine-tuning without full retraining. Experimental results demonstrate its superior ability to generate high-quality PROTAC linkers.

Inspired by the impressive capabilities of GPT-4o, there is growing interest in enabling speech language models (SLMs) to engage in natural, fluid spoken interactions with humans. Recent advancements have led to the development of several SLMs that demonstrate promising results in this area. However, current approaches have yet to fully exploit dual-channel speech data, which inherently captures the structure and dynamics of human conversation. In this work, we systematically explore the use of dual-channel speech data in the context of modern large language models, and introduce a novel generative modeling paradigm—Next-Token-Pair Prediction (NTPP)—to enable speaker-independent dual-channel spoken dialogue learning using decoder-only architectures for the first time. We evaluate our approach on standard benchmarks, and empirical results show that our proposed method, NTPP, significantly improves the conversational abilities of SLMs in terms of turn-taking prediction, response coherence, and naturalness. Moreover, compared to existing methods, NTPP achieves substantially lower inference latency, highlighting its practical efficiency for real-time applications. Demo and code can be found at https://audio-3059.pages.dev.

Parameter-efficient fine-tuning (PEFT) is essential for adapting large foundation models without excessive storage cost. However, current approaches such as LoRA treat each layer’s adaptation independently, overlooking correlations across layers. This independence causes the number of trainable parameters to grow linearly with model depth. We provide theoretical and empirical evidence that skip connections in transformers create smooth gradient propagation across layers. This smoothness leads to weight adaptations that concentrate most of their energy in low-frequency spectral components, especially along the layer dimension. Empirical analysis confirms this effect, showing that most of adaptation energy lies in low frequencies. Building on this insight, we propose CrossSpectra, which parameterizes all attention-weight adaptations $(Q, K, V)$ across layers as a single 3D tensor and represents them with sparse spectral coefficients ($\kappa_1, \kappa_2$). Using $\kappa_{1}$ non-zero coefficients within each layer’s frequency space and truncating to $\kappa_{2}$ frequencies across layers, CrossSpectra requires $\mathcal{O}(\kappa_{1}\kappa_{2})$ parameters instead of LoRA’s $\mathcal{O}(Lrd)$, where $L$ is the number of layers and $r$ the rank. Across natural-language and vision benchmarks, \methodname{} matches or surpasses baseline performance while using fewer parameters than LoRA, achieving only $0.36\%$ of LoRA’s parameter count when fine-tuning LLaMA-7B on instruction-following tasks. These results show that exploiting the \textbf{architectural smoothness of transformers} through spectral analysis yields major efficiency gains in PEFT.

Multi-task learning (MTL) is a widely explored paradigm that enables the simultaneous learning of multiple tasks using a single model. Despite numerous solutions, the key issues of optimization conflict and task imbalance remain under-addressed, limiting performance. Unlike existing optimization-based approaches that typically reweight task losses or gradients to mitigate conflicts or promote progress, we propose a novel approach based on Continual Optimization with Symmetry Teleportation (COST). During MTL optimization, when an optimization conflict arises, we seek an alternative loss-equivalent point on the loss landscape to reduce conflict. Specifically, we utilize a low-rank adapter (LoRA) to facilitate this practical teleportation by designing convergent, loss-invariant objectives. Additionally, we introduce a historical trajectory reuse strategy to continually leverage the benefits of advanced optimizers. Extensive experiments on multiple mainstream datasets demonstrate the effectiveness of our approach. COST is a plug-and-play solution that enhances a wide range of existing MTL methods. When integrated with state-of-the-art methods, COST achieves superior performance.

Nowadays deep models are required to be versatile due to the increasing realistic needs. Multi-task learning (MTL) offers an efficient way for this purpose to learn multiple tasks simultaneously with a single model. However, prior MTL solutions often focus on resolving conflicts and imbalances during optimization, which may not outperform simple linear scalarization strategies~\citep{xin2022current}. Instead of altering the optimization trajectory, this paper leverages mode connectivity to efficiently approach the Pareto front and identify the desired trade-off point. Unlike Pareto Front Learning (PFL), which aims to align with the entire Pareto front, we focus on effectively and efficiently exploring optimal trade-offs. However, three challenges persist: (1) the low-loss path can neither fully traverse trade-offs nor align with user preference due to its randomness, (2) commonly adopted Bézier curves in mode connectivity are ill-suited to navigating the complex loss landscapes of deep models, and (3) poor scalability to large-scale task scenarios. To address these challenges, we adopt non-uniform rational B-Splines (NURBS) to model mode connectivity, allowing for more flexible and precise curve optimization. Additionally, we introduce an order-aware objective to explore task loss trade-offs and employ a task grouping strategy to enhance scalability under massive task scenarios. Extensive experiments on key MTL datasets demonstrate that our proposed method, *EXTRA* (EXplore TRAde-offs), effectively identifies the desired point on the Pareto front and achieves state-of-the-art performance. *EXTRA* is also validated as a plug-and-play solution for mainstream MTL approaches.

Many significant problems involving crystal property prediction from 3D structures have limited labeled data due to expensive and time-consuming physical simulations or lab experiments. To overcome this challenge, we propose a pretrain-finetune framework for the crystal property prediction task named CrysDiff based on diffusion models. In the pre-training phase, CrysDiff learns the latent marginal distribution of crystal structures via the reconstruction task. Subsequently, CrysDiff can be fine-tuned under the guidance of the new sparse labeled data, fitting the conditional distribution of the target property given the crystal structures. To better model the crystal geometry, CrysDiff notably captures the full symmetric properties of the crystals, including the invariance of reflection, rotation, and periodic translation. Extensive experiments demonstrate that CrysDiff can significantly improve the performance of the downstream crystal property prediction task on multiple target properties, outperforming all the SOTA pre-training models for crystals with good margins on the popular JARVIS-DFT dataset.

Contrastive learning, a form of Self-Supervised Learning (SSL), typically consists of an alignment term and a regularization term. The alignment term minimizes the distance between the embeddings of a positive pair, while the regularization term prevents trivial solutions and expresses prior beliefs about the embeddings. As a widely used regularization technique, soft decorrelation has been employed by several non-contrastive SSL methods to avoid trivial solutions. While the decorrelation term is designed to address the issue of dimensional collapse, we find that it fails to achieve this goal theoretically and experimentally. Based on such a finding, we extend the soft decorrelation regularization to minimize the distance between the covariance matrix and an identity matrix. We provide a new perspective on the geometric distance between positive definite matrices to investigate why the soft decorrelation cannot efficiently solve the dimensional collapse. Furthermore, we construct a family of loss functions utilizing the Bregman Matrix Divergence (BMD), with the soft decorrelation representing a specific instance within this family. We prove that a loss function (LogDet) in this family can solve the issue of dimensional collapse. Our novel loss functions based on BMD exhibit superior performance compared to the soft decorrelation and other baseline techniques, as demonstrated by experimental results on graph and image datasets.

Organic reaction prediction is a critical task in drug discovery. Recently, researchers have achieved non-autoregressive reaction prediction by modeling the redistribution of electrons, resulting in state-of-the-art top-1 accuracy, and enabling parallel sampling. However, the current non-autoregressive decoder does not satisfy two essential rules of electron redistribution modeling simultaneously: the electron-counting rule and the symmetry rule. This violation of the physical constraints of chemical reactions impairs model performance. In this work, we propose a new framework called ReactionSink that combines two doubly stochastic self-attention mappings to obtain electron redistribution predictions that follow both constraints. We further extend our solution to a general multi-head attention mechanism with augmented constraints. To achieve this, we apply Sinkhorn's algorithm to iteratively update self-attention mappings, which imposes doubly conservative constraints as additional informative priors on electron redistribution modeling. We theoretically demonstrate that our ReactionSink can simultaneously satisfy both rules, which the current decoder mechanism cannot do. Empirical results show that our approach consistently improves the predictive performance of non-autoregressive models and does not bring an unbearable additional computational cost.

Reaction and retrosynthesis prediction are two fundamental tasks in computational chemistry. In recent years, these two tasks have attracted great attentions from both machine learning and drug discovery communities. Various deep learning approaches have been proposed to tackle these two problems and achieved initial success. In this survey, we conduct a comprehensive investigation on advanced deep learning-based reaction and retrosynthesis prediction models. We first summarize the design mechanism, strengths and weaknesses of the state-of-the-art approaches. Then we further discuss limitations of current solutions and open challenges in the problem itself. Last but not the least, we present some promising directions to facilitate future research. To our best knowledge, this paper is the first comprehensive and systematic survey on unified understanding of reaction and retrosynthesis prediction.

Predicting molecular properties is known to be a difficult few-shot learning problem in drug discovery due to high failure rates of virtual screening. Meta-learning frameworks have been successfully leveraged to tackle this few-shot molecular property prediction problem. Recent work has shown that the generalization of meta- learning can be further improved by task augmentation techniques. However, it is challenging to directly apply current popular task augmentation strategies to molecular modeling due to its highly discrete nature. To this end, we propose a motif-based task augmentation (MTA) technique to address this challenge. MTA is mainly conducting task augmentations by generating new labeled samples through retrieving highly relevant motifs from a pre-defined motif vocabulary as an external memory. Both support samples and query samples are interpolated with their corresponding retrieved motifs in the latent space in a convex manner parameterized by a specific probability distribution. Augmented query samples are classified by measuring their Euclidean distances to augmented prototype clusters. With our novel MTA approach, the above two overfitting issues can be alleviated through establishing connections between different tasks with augmented motifs. Empirical results for popular benchmark datasets demonstrate that our approach can consistently outperform existing baseline methods and can also provide some interpretable relation information between structural motifs and molecules.

Aiming to alleviate data sparsity and cold-start problems of tradi- tional recommender systems, incorporating knowledge graphs (KGs) to supplement auxiliary information has recently gained considerable attention. Via unifying the KG with user-item interactions into a tripartite graph, recent works explore the graph topologies to learn the low-dimensional representations of users and items with rich semantics. These real-world tripartite graphs are usually scale-free, however, the intrinsic hierarchical graph structures of which are underemphasized in existing works, consequently, leading to suboptimal recommendation performance. To address this issue and provide more accurate recommendation, we propose a knowledge-aware recommendation method with Lorentz model of the hyperbolic geometry, namely Lorentzian Knowledge-enhanced Graph convolutional networks for Recommendation (LKGR). LKGR facilitates better modeling of scale-free tripartite graphs after the data unification. Specifically, we employ different information propagation strategies in the hyperbolic space to explicitly encode heterogeneous information from historical interactions and KGs. Additionally, our proposed knowledge-aware attention mechanism enables the model to automatically measure the information contribution, producing the coherent information aggregation in the hyperbolic space. Extensive experiments on three real-world benchmarks demonstrate that LKGR outperforms state-of-the-art methods by 3.6-15.3% of [email protected] on Top-K recommendation.

Graph Neural Networks (GNNs) have achieved remarkable success by extending traditional convolution to learning on non-Euclidean data. The key to the GNNs is adopting the neural message-passing paradigm with two stages: aggregation and update. The current design of GNNs considers the topology information in the aggregation stage. However, in the updating stage, all nodes share the same updating function. The identical updating function treats each node embedding as i.i.d. random variables and thus ignores the implicit relationships between neighborhoods, which limits the capacity of the GNNs. The updating function is usually implemented with a linear transformation followed by a non-linear activation function. To make the updating function topology-aware, we inject the topological information into the non-linear activation function and propose Graph-adaptive Rectified Linear Unit (GReLU), which is a new parametric activation function incorporating the neighborhood information in a novel and efficient way. The parameters of GReLU are obtained from a hyperfunction based on both node features and the corresponding adjacent matrix. To reduce the risk of overfitting and the computational cost, we decompose the hyperfunction as two independent components for nodes and features respectively. We conduct comprehensive experiments to show that our plug-and-play GReLU method is efficient and effective given different GNN backbones and various downstream tasks.

The semi-supervised multi-label classification problem primarily deals with Euclidean data, such as text with a 1D grid of tokens and images with a 2D grid of pixels. However, the non-Euclidean graph-structured data naturally and constantly appears in semi-supervised multi-label learning tasks from various domains like social networks, citation networks, and protein-protein interaction (PPI) networks. Moreover, the existing popular node embedding methods, like Graph Neural Networks (GNN), focus on graphs with simplex labels and tend to neglect label correlations in the multi-label setting, so the easy adaption proves empirically ineffective. Therefore, graph representation learning for the semi-supervised multi-label learning task is crucial and challenging. In this work, we incorporate the idea of label embedding into our proposed model to capture both network topology and higher-order multi-label correlations. The label embedding is generated along with the node embedding based on the topological structure to serve as the prototype center for each class. Moreover, the similarity of the label embedding and node embedding can be used as a confidence vector to guide the label smoothing process, formulating as a margin ranking optimization problem to learn the second-order relations between labels. Extensive experiments on real-world datasets from various domains demonstrate that our model significantly outperforms the state-of-the-art models for node-level tasks.

Dynamic graphs arise in a plethora of practical scenarios such as social networks, communication networks, and financial transaction networks. Given a dynamic graph, it is fundamental and essential to learn a graph representation that is expected not only to preserve structural proximity but also jointly capture the time-evolving patterns. Recently, graph convolutional network (GCN) has been widely explored and used in non-Euclidean application domains. The main success of GCN, especially in handling dependencies and passing messages within nodes, lies in its approximation to Laplacian smoothing. As a matter of fact, this smoothing technique can not only encourage must-link node pairs to get closer but also push cannot-link pairs to shrink together, which potentially cause serious feature shrink or oversmoothing problem, especially when stacking graph convolution in multiple layers or steps. For learning time-evolving patterns, a natural solution is to preserve historical state and combine it with the current interactions to obtain the most recent representation. Then the serious feature shrink or oversmoothing problem could happen when stacking graph convolution explicitly or implicitly according to current prevalent methods, which would make nodes too similar to distinguish each other. To solve this problem in dynamic graph embedding, we analyze the shrinking properties in the node embedding space at first, and then design a simple yet versatile method, which exploits L2 feature normalization constraint to rescale all nodes to hypersphere of a unit ball so that nodes would not shrink together, and yet similar nodes can still get closer. Extensive experiments on four real-world dynamic graph datasets compared with competitive baseline models demonstrate the effectiveness of the proposed method.

A large number of real-world graphs or networks are inherently heterogeneous, involving a diversity of node types and relation types. Heterogeneous graph embedding is to embed rich structural and semantic information of a heterogeneous graph into low-dimensional node representations. Existing models usually define multiple metapaths in a heterogeneous graph to capture the composite relations and guide neighbor selection. However, these models either omit node content features, discard intermediate nodes along the metapath, or only consider one metapath. To address these three limitations, we propose a new model named Metapath Aggregated Graph Neural Network (MAGNN) to boost the final performance. Specifically, MAGNN employs three major components, i.e., the node content transformation to encapsulate input node attributes, the intra-metapath aggregation to incorporate intermediate semantic nodes, and the inter-metapath aggregation to combine messages from multiple metapaths. Extensive experiments on three real-world heterogeneous graph datasets for node classification, node clustering, and link prediction show that MAGNN achieves more accurate prediction results than state-of-the-art baselines.