π― What it does: A unified scale-invariant graph-language model (SFGL) is proposed, which simultaneously completes graph structure generation and text embedding in semi-supervised text classification.
Scaling Diffusion Language Models via Adaptation from Autoregressive Models
Shansan Gong (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelText
π― What it does: Adapt existing autoregressive large models (such as GPT-2 and LLaMA-2) using techniques like attention mask annealing and output displacement to train diffusion language models DiffuGPT and DiffuLLaMA with 127M, 355M, and 7B parameters.
Scaling Large Language Model-based Multi-Agent Collaboration
Chen Qian (Tsinghua University), Maosong Sun (Tsinghua University)
CodeTransformerLarge Language ModelAgentic AIText
π― What it does: This paper proposes the MACNET framework, which organizes multiple agents through a directed acyclic graph (DAG) to collaboratively reason and complete complex tasks, achieving cooperation among over a thousand agents.
π― What it does: In multi-task offline reinforcement learning, a single large-scale model JOWA is trained, utilizing a world model of image observations and joint optimization of Q-values to achieve a general agent for multiple Atari games.
Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Aohan Zeng (Tsinghua University), Jie Tang (Tsinghua University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
π― What it does: This paper proposes a pre-training method for speech language models (SpeechLM) that utilizes synthesized interleaved speech-text data to massively expand the text corpus.
Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study
Shawn Tan (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Laboratory), Yikang Shen (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Laboratory)
CodeRetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextSequential
π― What it does: Proposed and implemented an attention mechanism based on the stick-breaking process, replacing the traditional softmax + RoPE;
Shen Nie (Renmin University of China), Chongxuan Li (Renmin University of China)
CodeGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
π― What it does: This study investigates the scalability and effectiveness of Masked Diffusion Models (MDM) in text generation and language understanding, and proposes an unsupervised Classifier-Free Guidance (CFG) to enhance conditional reasoning performance.
Schur's Positive-Definite Network: Deep Learning in the SPD cone with structure
Can Pouliquen (ENS de Lyon), Titouan Vayer (ENS de Lyon)
CodeTabular
π― What it does: A neural network module named SpodNet is proposed, which ensures that the output matrix is strictly symmetric positive definite (SPD) during training and can embed additional structural constraints (such as element-wise sparsity), achieving joint learning of SPD and sparsity.
ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery
Ziru Chen (Ohio State University), Huan Sun (Ohio State University)
CodeLarge Language ModelTextBenchmark
π― What it does: A benchmark called ScienceAgentBench is proposed to evaluate the automation capabilities of language models in data-driven scientific discovery.
SciLitLLM: How to Adapt LLMs for Scientific Literature Understanding
Sihang Li (University of Science and Technology of China), Hengxing Cai (DP Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A specialized LLM for scientific literature understanding, SciLitLLM, has been constructed, and a complete training pipeline combining Continuous Pre-Training (CPT) and Supervised Fine-Tuning (SFT) has been proposed.
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By generating untrustworthy samples through self-supervision and performing preference fine-tuning based on this, the fidelity of large language models in conditional text generation tasks is improved.
π― What it does: This paper proposes Score Forgetting Distillation (SFD), a data-free, score-based machine unlearning method for forgetting specified categories or concepts in diffusion models and achieving first-order sampling.
Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning
Hai Zhang (Tongji University), Lanqing Li (Zhejiang Lab)
CodeMeta LearningReinforcement LearningTabular
π― What it does: This paper studies the issue of task representation drift in Context-Based Offline Meta Reinforcement Learning (COMRL) and proposes a theoretical framework for achieving monotonic performance improvement by controlling the updates of the context encoder.
π― What it does: A scalable and decoupled low-rank adaptation method, SD-LoRA, is proposed for class-incremental learning, which is completely sample-free and can be optimized end-to-end.
SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
Han Shen (Rensselaer Polytechnic Institute), Tianyi Chen (Rensselaer Polytechnic Institute)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The SEAL framework is proposed, which uses a two-layer optimization to learn a data selector that automatically filters safe and high-quality samples during the fine-tuning process to enhance the model's safety and performance.
Searching for Optimal Solutions with LLMs via Bayesian Optimization
Dhruv Agarwal (University of Massachusetts Amherst), Rashmi Gangadharaiah (Amazon Web Services AI Labs)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: A search framework called BOPRO has been designed and implemented, which combines Bayesian optimization with large language models to automatically adjust search strategies in uncertain environments.
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: For full parameter or PEFT fine-tuning of large language models, a zero-order optimizer HiZOO utilizing diagonal Hessian information is proposed, along with theoretical analysis and experimental validation.
π― What it does: Two second-order optimal methods, LEN (Lazy Extra Newton) and LEN-restart, are proposed for solving min-max optimization problems that are convex-concave and strongly convex-strongly concave, utilizing 'lazy Hessian' to reuse old Hessians, significantly reducing computational costs.
Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models
Γngela LΓ³pez-Cardona (TelefΓ³nica Scientific Research), Ioannis Arapakis (TelefΓ³nica Scientific Research)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes the GazeReward framework, which integrates implicit feedback generated by eye tracking (ET) into the reward model (RM) to enhance the human alignment effects of large language models (LLMs).
Selective Aggregation for Low-Rank Adaptation in Federated Learning
Pengxin Guo (University of Hong Kong), Liangqiong Qu (University of Hong Kong)
CodeFederated LearningTransformerLarge Language ModelText
π― What it does: Addressing the low-rank adaptation problem of LoRA in federated learning, the roles of matrices A and B were analyzed from both theoretical and experimental perspectives. It was found that matrix A learns general knowledge while matrix B learns client-specific knowledge. Based on this, FedSA-LoRA was proposed and extended to rsLoRA and VeRA, forming FedSA-rsLoRA and FedSA-VeRA.
π― What it does: This paper proposes the PASLE framework for Online Testing Time Adaptation (OTTA) on unlabeled test data, which classifies samples into deterministic (one-hot) and uncertain (candidate set) categories through selective label enhancement, and dynamically refines the candidate labels of uncertain samples during the training process.
Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models
Ce Zhang (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
CodeGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: A training-free self-correcting decoding method DeGF is proposed, which utilizes visual feedback from a text-to-image generation model to recursively correct hallucinations during the decoding process of LVLM-generated responses.
π― What it does: A self-improving robust preference optimization (SRPO) framework is proposed for training language models from human preference data.
Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models
Fushuo Huo (Hong Kong Polytechnic University), Peilin Zhao (Tencent)
CodeGenerationComputational EfficiencyTransformerVision Language ModelMultimodality
π― What it does: This paper proposes a Self-Introspective Decoding (SID) mechanism that utilizes self-assessment of the importance of visual tokens through a pre-trained large visual-language model. It retains only the least important visual tokens in the early decoding layers to generate visual-textual associations and suppresses hallucinations through contrastive decoding.
Self-Normalized Resets for Plasticity in Continual Learning
Vivek Farias, Adam Daniel Jozefiak
CodeOptimizationText
π― What it does: This paper proposes an adaptive threshold-based self-normalizing reset algorithm SNR to address the plasticity loss problem in continual learning.
CodeData SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposes AUTOIF, a method for instruction-following data synthesis without manual annotation by automatically generating verifiable instructions and corresponding execution code;
π― What it does: A self-supervised diffusion model DELID is proposed, which learns electron-perceptive molecular representations through molecular substructure decomposition and electronic layer information retrieved from databases, thereby improving molecular property prediction performance without the need for expensive quantum mechanical calculations.
π― What it does: A fully self-supervised MRI denoising framework called Di-Fusion is proposed, which utilizes post-diffusion steps and adaptive sampling to achieve denoising and iterative refinement of dMRI.
Self-Updatable Large Language Models by Integrating Context into Model Parameters
Yu Wang (University of California San Diego), Julian McAuley (University of California San Diego)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: A self-updating large language model called SELF-PARAM is proposed, which achieves this without additional parameters by minimizing KL divergence. It can directly embed contextual knowledge into model parameters, enabling rapid memory and long-term retention.
π― What it does: A Selective Knowledge Distillation (SelKD) framework is proposed, allowing the student model to learn knowledge from specified sub-tasks of the teacher model, and achieving flexible selection and matching of knowledge through Inverse Optimal Transport (IOT).
π― What it does: This paper proposes Semantic-Aware Representation Learning (SARL), which utilizes sparse activation and object prototype learning to capture semantic similarity across tasks, guiding representation learning to reduce interference and enhance lifelong learning effectiveness.
Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors
Weixuan Wang (University of Edinburgh), Wei Peng (RMIT University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A dynamic activation intervention method based on input semantic adaptation (SADI) is proposed, which adjusts the internal activations of LLMs in real-time during the inference phase to achieve behavior alignment.
Semi-Parametric Retrieval via Binary Bag-of-Tokens Index
Jiawei Zhou (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
CodeRetrievalContrastive LearningTextBenchmark
π― What it does: A semi-parametric retrieval framework SIDR is proposed, which supports both parameterized indexing based on neural embeddings and non-parameterized indexing based on Bag-of-Tokens;
π― What it does: A semi-supervised CLIP adaptation method called SEMICLIP is proposed, which utilizes unlabeled images through semantic concept mining and trapezoidal consistency.
π― What it does: This paper presents PreWorld, a semi-supervised visual-centric 3D occupancy world model, which achieves unified modeling of 3D occupancy prediction, 4D prediction, and motion planning through a two-stage training process.
Sensitivity Verification for Additive Decision Tree Ensembles
Arhaan Ahmad (Indian Institute of Technology Bombay), S. Akshay (Indian Institute of Technology Bombay)
CodeClassificationOptimizationTabularBenchmark
π― What it does: This paper proposes a sensitivity validation framework for additive decision tree ensemble models and implements a tool called SENSPB based on pseudo-Boolean constraints to test the sensitivity of a given feature set.
Sensitivity-Constrained Fourier Neural Operators for Forward and Inverse Problems in Parametric Differential Equations
Abdolmehdi Behroozi (Penn State University), Daniel Kifer (Penn State University)
CodePhysics RelatedOrdinary Differential Equation
π― What it does: A new sensitivity-constrained Fourier neural operator (SC-FNO) is proposed to address forward and inverse problems in parameter differential equations, particularly the challenges in sensitivity computation and concept drift.
SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction
Lu Dai (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
CodeRetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: An automatic evaluation metric based on Semantic Perplexity (SePer) and its difference (ΞSePer) is proposed to measure the actual knowledge gain from retrieval increments in Retrieval-Augmented Generation (RAG). It estimates the belief distribution changes of large language models (LLMs) through sampling and semantic clustering.
π― What it does: Proposes the Seq-VCR regularization method, which combines pauses (tokens) to enhance the diversity of intermediate representations in Transformers, addressing the representation collapse problem in multi-step reasoning tasks.
π― What it does: A sampler named Sequential Controlled Langevin Diffusion (SCLD) has been developed, which combines continuous-time SMC with diffusion sampling to achieve an end-to-end trainable sampling method.
Varun Mulchandani (North Carolina State University), Jung-Eun Kim (North Carolina State University)
CodeDomain AdaptationData-Centric LearningImage
π― What it does: A training data pruning method is proposed that does not require domain knowledge or labeled information. By removing a small number of hard-to-learn samples that have a strong impact on model learning, the method weakens the model's dependence on spurious correlations and enhances the model's robustness in out-of-distribution scenarios.
SFESS: Score Function Estimators for $k$-Subset Sampling
Klas Wijk (KTH Royal Institute of Technology), Hossein Azizpour (KTH Royal Institute of Technology)
CodeScore-based ModelImage
π― What it does: A score function-based k-subset sampling estimator (SFESS) is proposed, achieving unbiased gradients and the generation of k-hot samples.
π― What it does: This paper proposes an implicit surface representation method based on Line Segment Field (LSF), referred to as SALS, and implements the corresponding neural network and point cloud reconstruction pipeline.
π― What it does: A Shapley-Guided Utility Learning (SGUL) framework is proposed to evaluate the value of neighboring nodes during the graph reasoning phase in the absence of test labels.
π― What it does: This paper studies the behavior of Sharpness-Aware Minimization (SAM) in the later stages of training, finding that using SAM for just a few rounds in the later training phase can achieve the same generalization performance and flatter minima as the full SAM, while early use of SAM has limited impact on the final results.
Sharpness-Aware Minimization: General Analysis and Improved Rates
Dimitris Oikonomou (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
CodeOptimizationImage
π― What it does: A unified Sharpness-Aware Minimization (Unified SAM) algorithm is proposed, combining the update rules of traditional SAM and USAM, and providing convergence analysis for arbitrary sampling (including importance sampling and Ο-nice sampling);
Shedding Light on Time Series Classification using Interpretability Gated Networks
Yunshi Wen (Rensselaer Polytechnic Institute), Anak Agung Julius (Rensselaer Polytechnic Institute)
CodeClassificationExplainability and InterpretabilityMixture of ExpertsTime SeriesBiomedical Data
π― What it does: A hybrid model called InterpGN is proposed, which combines interpretable experts and deep neural networks to construct logical predicates using shapelets, achieving interpretable time series classification.
ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents
Haiyang SHEN, Yun Ma (Peking University)
CodeTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: A large-scale real API benchmark called SHORTCUTSBENCH was created, and the performance of 10 LLM-driven API agents was evaluated in terms of API selection, parameter filling, and input requests.
Shot2Story: A New Benchmark for Comprehensive Understanding of Multi-shot Videos
Mingfei Han (Bytedance Inc.), Heng Wang (Data61, CSIRO)
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
π― What it does: Proposed the Shot2Story benchmark, providing 42,958 multi-shot short videos with single-shot visuals and narrative subtitles, long video summaries, and question-answer pairs. Designed single-shot subtitles, video summaries, and multi-shot question-answer tasks, and conducted baseline experiments.
Sedrick Keh (Toyota Research Institute), Achal Dave (Toyota Research Institute)
CodeTransformerVision Language ModelImageTextMultimodality
π― What it does: The study investigates the impact of introducing image data early in the pre-training phase on VLM performance, comparing two-stage and one-stage pre-training strategies.
Show-o: One Single Transformer to Unify Multimodal Understanding and Generation
Jinheng Xie (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextMultimodality
π― What it does: A single Transformer model, Show-o, has been constructed to simultaneously achieve multimodal understanding (such as visual question answering and image captioning) and multimodal generation (such as text-to-image, text-guided filling, and video generation).
π― What it does: A SIM framework based on Surface Vision Transformer (SiT) is proposed, utilizing self-supervised video surface masking autoencoder (vsMAE) pre-training, and aligning through tri-modal CLIP to map the brain electroencephalogram (fMRI), video, and audio representations of 3-second movie clips into a shared space, achieving cross-individual and cross-scene multimodal retrieval and video reconstruction.
π― What it does: The SiMHand framework is proposed, which pre-trains 3D gesture estimation using a large-scale dataset of hand images from the wild, leveraging contrastive learning on similar hand pairs mined from different videos.
SIMPL: Scalable and hassle-free optimisation of neural representations from behaviour
Tom George, Claudia Clopath (Imperial College London)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTime Series
π― What it does: The SIMPL algorithm is proposed, which utilizes behavioral initial conditions to iteratively optimize latent variables and tuning curves in neural networks.
Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation
Mufei Li (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
CodeGenerationRetrievalGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: This paper proposes and implements a knowledge graph-based retrieval-augmented generation framework called SubgraphRAG, which first retrieves variable-sized and flexibly structured knowledge subgraphs through efficient subgraph retrieval, and then uses a non-fine-tuned large language model (LLM) to reason over the subgraphs, providing answers along with interpretable explanations.
π― What it does: This paper proposes SGF, a self-supervised world model that does not use RNNs, Transformers, discrete representations, or image reconstruction. It learns temporally consistent representations through frame/action stacking and data augmentation, achieving efficient training on the Atari 100k benchmark.
SimpleTM: A Simple Baseline for Multivariate Time Series Forecasting
Hui Chen (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison)
CodeTransformerTime Series
π― What it does: A lightweight multivariate time series forecasting model called SimpleTM is proposed, which uses Stationary Wavelet Transform (SWT) for multi-scale decomposition of time series and incorporates the geometric algebra vector product into the attention mechanism to capture high-order correlations between channels.
π― What it does: A simplified deep temporal difference learning method called PQN is proposed, eliminating the target network and large replay buffer, achieving stability through regularization (LayerNorm + β2).
π― What it does: By simulating the training dynamics, we reverse-engineer the training dataβgiven the initial and final parameters of a deep network, we optimize a synthetic dataset composed of random noise so that the final parameters of the network trained on this synthetic dataset are as similar as possible to the final parameters of the real network, thereby obtaining images similar to the original training data.
SimulPL: Aligning Human Preferences in Simultaneous Machine Translation
Donglei Yu (University of Chinese Academy of Sciences), Chengqing Zong (Institute of Automation, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: The SimulPL framework is proposed, which enhances model performance by aligning five major preferences of humans in synchronous machine translation (translation quality, monotonicity, key information, conciseness, and latency);
π― What it does: A framework named SINGER is proposed, which learns the solution operator of high-dimensional partial differential equations (PDEs) using subnetwork parameters driven by graph neural network-based stochastic differential equations, thereby directly obtaining approximate analytical solutions to PDEs without the need for spatial discretization.
π― What it does: This paper proposes a technique called TeKAP, which generates various synthetic teacher knowledge by injecting random noise into the feature maps and logits of a single teacher model, thereby enhancing the generalization ability of the student model using multiple perspectives without the need to train multiple teachers.
SiReRAG: Indexing Similar and Related Information for Multihop Reasoning
Nan Zhang (Pennsylvania State University), Chien-Sheng Wu (Pennsylvania State University)
CodeRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: A dual-perspective RAG indexing method called SIRERAG is proposed, which combines similarity and relevance to enhance the retrieval performance of multi-hop reasoning question answering.
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
π― What it does: A complete process for generating high-quality vector graphics (TikZ) from hand-drawn sketches is proposed, including the SKETI kZ dataset, data augmentation methods, and a dedicated image-language model IMGTI kZ.
CodeRecommendation SystemComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningSequential
π― What it does: In this work, the authors first evaluate the role of large language models (LLMs) in sequential recommendation, finding that most intermediate layers are redundant; they then propose a hierarchical knowledge distillation-based SLMREC model that utilizes a small LLM to achieve performance on par with large LLM Rec, significantly reducing model size and inference/training costs.
Small Models are LLM Knowledge Triggers for Medical Tabular Prediction
Jiahuan Yan (Zhejiang University), Jian Wu (University of Illinois Urbana-Champaign)
CodeClassificationOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularBiomedical Data
π― What it does: An unsupervised self-prompting cycle named SERSAL is proposed, utilizing collaborative learning between small models and large language models (LLMs) to enhance the performance of LLMs in numerical table prediction tasks (such as medical diagnosis).
π― What it does: This paper proposes SmartPretrain, a general, model-agnostic, and dataset-agnostic self-supervised pre-training framework that combines contrastive learning and reconstruction learning to enhance motion prediction performance in autonomous driving.
SMI-Editor: Edit-based SMILES Language Model with Fragment-level Supervision
Kangjie Zheng (Peking University), Ming Zhang (Peking University)
CodeDrug DiscoveryTransformerTextGraph
π― What it does: A SMILES language model based on editing, called SMI-EDITOR, is designed to randomly drop molecular substructures and allow the model to recover the original SMILES through editing operations, thus addressing the rapid saturation and substructure semantic loss issues of traditional MLMs.
SMT: Fine-Tuning Large Language Models with Sparse Matrices
Haoze He (Carnegie Mellon University), Heather Miller (University of California, Berkeley)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: For fine-tuning large language models, a Sparse Matrix Tuning (SMT) method is proposed that only updates sparse submatrices, significantly reducing memory and computational overhead;
SOAP: Improving and Stabilizing Shampoo using Adam for Language Modeling
Nikhil Vyas (Harvard University), Sham M. Kakade
CodeOptimizationTransformerLarge Language ModelText
π― What it does: An optimizer named SOAP is proposed, which combines the feature vector space of the Shampoo preprocessor with the adaptive updates of AdamW;
SoftMatcha: A Soft and Fast Pattern Matcher for Billion-Scale Corpus Searches
Hiroyuki Deguchi (NAIST), Sho Yokoi (NINJAL)
CodeRetrievalComputational EfficiencyText
π― What it does: A soft pattern matching algorithm that combines word vectors and inverted indexing is proposed, capable of achieving second-level retrieval on a billion-level corpus;
π― What it does: A PDE solving framework based on constraint learning, SCL, is proposed and implemented, capable of uniformly handling unsupervised, supervised, and scenarios with prior knowledge (structure, measurements, known solutions).
Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model
Longrong Yang (Zhejiang University), Xi Li (Kuaishou Technology)
CodeMixture of ExpertsVision Language ModelMultimodality
π― What it does: Identifying and eliminating gradient conflicts in Mixture-of-Experts through token-level gradient analysis, thereby enhancing the performance of large visual language models.
π― What it does: A large-scale end-to-end synthetic song detection dataset, SONICS, has been constructed, and an efficient SpecTTTra model for capturing long-term dependencies has been proposed, evaluating its performance in distinguishing between artificial and real songs.
SonicSim: A customizable simulation platform for speech processing in moving sound source scenarios
Kai Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
CodeRestorationGenerationData SynthesisAudio
π― What it does: A customizable mobile sound source simulation tool called SonicSim based on Habitat-sim has been developed, and it has been used to generate a large-scale mobile sound source speech separation/enhancement dataset called SonicSet.
SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization
Hong Qian (East China Normal University), Yang Yu (Nanjing University)
CodeOptimizationTabularBenchmark
π― What it does: The SOO-Bench benchmark suite is proposed, providing customizable narrow distribution offline datasets and stability-optimality metrics to systematically evaluate the stability and optimality of offline black-box optimization algorithms under different data distributions.
SOREL: A Stochastic Algorithm for Spectral Risks Minimization
Yuze Ge (Fudan University), Rujun Jiang (Fudan University)
CodeOptimizationTabular
π― What it does: The SOREL algorithm is proposed for achieving stochastic gradient optimization in the spectral risk minimization problem and provides convergence guarantees.
SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels
Xiangyu Dong (Chinese University of Hong Kong), Sibo Wang (Chinese University of Hong Kong)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: The research focuses on node anomaly detection (NAD) with very few labels and proposes a multi-space graph neural network framework called SpaceGNN, aimed at better utilizing graph structural information and improving detection performance in scenarios with limited labels.
SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training
Tianjin Huang (University of Exeter), Shiwei Liu (University of Oxford)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study addresses and mitigates gradient and loss spikes during the training of large language models, proposing an optimizer called SPAM that features momentum reset and spike-aware clipping, along with a low-memory implementation of sparse momentum.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
π― What it does: A SPAR self-play + tree search self-improvement framework is proposed, which enhances the instruction-following ability of LLMs through self-generated and self-evaluated preference pairs by the actor and evaluator.
CodeDomain AdaptationExplainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringAuto EncoderImage
π― What it does: Train PatchSAE on CLIP ViT to extract interpretable visual concepts and achieve spatial localization, and then analyze the relationship between concepts and task categories during the prompt-based adaptation process using this model.
SPARTUN3D: Situated Spatial Understanding of 3D World in Large Language Model
Yue Zhang (Michigan State University), Lifu Huang (University of California Davis)
CodeData SynthesisRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityPoint Cloud
π― What it does: A scalable LLM-based coordinate 3D dataset, Spartun3D, has been constructed, and based on this, improvements have been made to the 3D visual language model to achieve situated spatial understanding.
π― What it does: Spatial-Mamba is proposed in visual tasks, capturing spatial dependencies in images by introducing structure-aware state fusion in the state space.
CodeHyperparameter SearchDrug DiscoveryNeural Architecture SearchConvolutional Neural NetworkTime SeriesBiomedical Data
π― What it does: Evaluated whether foundational models in specialized fields (genomics, satellite imaging, time series) can outperform traditional supervised learning, and proposed two automated workflows for systematic comparison.
Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement
Haijin Zeng (Harvard University), Yong Xu (Harbin Institute of Technology)
CodeRestorationCompressionDiffusion modelImage
π― What it does: A prediction-based mixed decoupled subspace diffusion framework (PSR-SCI) is proposed, which first uses a lightweight predictor to generate a coarse multispectral image, then decomposes it into low-dimensional abundance maps and spectral coefficients through a reversible spectral embedding module, and subsequently reconstructs high-frequency details in the low-dimensional subspace using a pre-trained RGB diffusion model, ultimately restoring the complete multispectral image.
Spectral-Refiner: Accurate Fine-Tuning of Spatiotemporal Fourier Neural Operator for Turbulent Flows
Shuhao Cao (University of Missouri Kansas City), Yuanzhe Xi (Emory University)
CodeOptimizationComputational EfficiencyTime SeriesPhysics Related
π― What it does: A new spatiotemporal Fourier neural operator (ST-FNO) and a hybrid training paradigm have been designed to accurately solve the turbulent Navier-Stokes equations.
π― What it does: A graph neural network that combines spectral filtering with Riemannian geometryβCUSPβis proposed for node classification and link prediction tasks.
Speech Robust Bench: A Robustness Benchmark For Speech Recognition
Muhammad A Shah, Nicolas Kourtellis (Telefonica Research)
CodeRecognitionAdversarial AttackBenchmarkAudio
π― What it does: This paper presents the Speech Robust Bench (SRB), a comprehensive robustness benchmark covering 114 types of speech recognition challenge scenarios.
SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking
Xingrun Xing (University of Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation)
CodeSpiking Neural NetworkLarge Language ModelText
π― What it does: Designed and implemented a scalable spiking neural network large language model, SpikeLLM, with parameters ranging from 7 to 70 billion, and proposed GIF neurons and the Optimal Brain Spiking (OBSpiking) framework to enhance spiking encoding efficiency and trainability.
Sports-Traj: A Unified Trajectory Generation Model for Multi-Agent Movement in Sports
Yi Xu (Northeastern University), Yun Fu (Northeastern University)
CodeGenerationTransformerAuto EncoderTime Series
π― What it does: A unified trajectory generation model called UniTraj is proposed, capable of handling various tasks such as trajectory prediction, completion, and spatiotemporal recovery in one go.
SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models
Haotian Xia (University of California, Irvine), Hanjie Chen (Rice University)
CodeTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityBenchmark
π― What it does: The SPORTU sports understanding benchmark is proposed, which includes two subsets: text and slow-motion videos, to evaluate the capabilities of multimodal large language models in rule reasoning, strategy understanding, and video perception.
Spread Preference Annotation: Direct Preference Judgment for Efficient LLM Alignment
Dongyoung Kim (Korea Advanced Institute of Science and Technology), Jaehyung Kim (Yonsei University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By utilizing a minimal amount of manually labeled preference data, new preference samples are generated through self-supervision and self-correction, enhancing the consistency between LLMs and human preferences.
π― What it does: This paper proposes a spreading OOD detection task on graph structures, constructs an evaluation benchmark based on epidemic propagation models, and introduces an energy distribution-based aggregation detector named EDBD.
Spurious Forgetting in Continual Learning of Language Models
Junhao Zheng (South China University of Technology), Qianli Ma (South China University of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper studies the phenomenon of 'pseudo-forgetting' in large language models during continual learning, revealing that performance degradation primarily stems from task alignment failure rather than knowledge loss, and proposes a strategy of freezing lower-level parameters (Freeze) to alleviate this issue.
SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget
Zihao Wang (Peking University), Shaoduo Gan (Peking University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a 2D compression algorithm called SQUEEZEATTENTION, which is based on hierarchical importance dynamic allocation of KV-cache. During the inference process, the influence of each attention layer is measured using cosine similarity, and the cache budget is reallocated according to importance.
π― What it does: Proposed and implemented the SSLAM framework, which significantly improves the model's performance in multi-source audio scenarios by incorporating audio mixing (partial mixing) and source retention loss in self-supervised audio pre-training.
π― What it does: The Orthogonal Low-rank Embedding (OLE) method is migrated to Self-Supervised Learning (SSL), proposing the SSOLE framework to achieve both positive sample alignment and negative sample separation.