Conference on Neural Information Processing Systems Β· 2283 papers
ShapeLLM-Omni: A Native Multimodal LLM for 3D Generation and Understanding
Junliang Ye (Tsinghua University), Jun Zhu (Peking University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityPoint Cloud
π― What it does: Developed ShapeLLM-Omni, a multimodal large language model capable of generating, understanding, and editing text, images, and 3D content within the same model.
ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models
Bosong Huang (Griffith University), Shirui Pan (Griffith University)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime SeriesElectrocardiogram
π― What it does: A post-hoc time series classification model interpretation framework SHAPEX is proposed, which utilizes shape subsequences for segmentation and evaluates their contributions using Shapley values, thereby providing more causal and interpretable explanations.
Shapley-Based Data Valuation for Weighted $k$-Nearest Neighbors
Guangyi Zhang (Shenzhen Technology University), Aristides Gionis (KTH Royal Institute of Technology)
CodeData-Centric LearningTabular
π― What it does: This paper proposes to transform the calculation of Shapley values for weighted k-nearest neighbor models into an unweighted k-nearest neighbor form through data replication, and provides an efficient approximation algorithm.
π― What it does: This paper proposes the maximum correlation coefficient SHGR based on Spearman rank correlation, and provides a differentiable neural network estimator and its cross-encoder architecture for one-time estimation of one-to-one, multivariate to univariate, and complete correlation matrices of two sets of variables; it also verifies its robustness under noise, outliers, and independence assumptions, and applies it to feature selection.
Shift Before You Learn: Enabling Low-Rank Representations in Reinforcement Learning
Bastien Dubail (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
CodeReinforcement LearningTabular
π― What it does: This study investigates the use of shifted successor measures in reinforcement learning for low-rank approximation and provides finite sample upper bounds for sampling error and estimation error.
Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence
Shaopeng Fu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
CodeComputational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper studies prison break attacks on large language models (LLMs) and proposes a method to effectively defend against long-length prison break attacks through short-length adversarial training (AT).
π― What it does: This paper proposes an efficient post-training method called SCFM, which utilizes velocity field self-distillation to quickly distill a large-scale pre-trained flow-matching diffusion model into a high-quality sampler that requires only 3 steps;
ShortListing Model: A Streamlined Simplex Diffusion for Discrete Variable Generation
Yuxuan Song (Tsinghua University), Wei-Ying Ma (Tsinghua University)
CodeGenerationData SynthesisDrug DiscoveryDiffusion modelTextBiomedical Data
π― What it does: A simplified Simplex Diffusion Model (SLM) is proposed for discrete variable generation, transforming the generation process into a process from all candidates to a single category using a stepwise candidate pruning approach.
SIGMA: Refining Large Language Model Reasoning via Sibling-Guided Monte Carlo Augmentation
Yanwei Ren (Beihang University), Liu Liu (Beihang University)
CodeLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: Using the chain reasoning paths generated by MCTS, we improve the reasoning process by reusing the information of discarded sibling nodes to construct a high-quality small dataset.
π― What it does: This paper proposes the SignFlow Bipartite Subgraph Network (SBSN), which achieves edge signing prediction in large-scale signed bipartite graphs through subgraph sampling, directional SignFlow aggregation, and node feature distillation.
SilentStriker: Toward Stealthy Bit-Flip Attacks on Large Language Models
HAOTIAN XU, Cheng Zhuo (Zhejiang University)
CodeGenerationAdversarial AttackTransformerLarge Language ModelText
π― What it does: A covert bit-flipping attack (SilentStriker) for large language models is proposed, which uses a joint loss of key token suppression and perplexity constraints to generate fluent but incorrect text with a small number of bit flips.
Sim-LLM: Optimizing LLM Inference at the Edge through Inter-Task KV Reuse
Ruikun Luo (Huazhong University of Science and Technology), Yun Yang (Swinburne University of Technology)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: In the edge computing scenario, Sim-LLM is proposed to reduce GPU memory usage during LLM inference and improve system throughput by identifying task similarity and reusing the KV cache of previous tasks (especially the upper-level KV).
Simple and Optimal Sublinear Algorithms for Mean Estimation
Beatrice Bertolotti (University of Pavia), Sudarshan Shyam (Aarhus University)
CodeOptimizationTabular
π― What it does: This paper studies the sublinear multivariate mean estimation problem in d-dimensional Euclidean space and proposes three algorithms with optimal sample complexity to estimate the mean of a point set.
SimpleStrat: Diversifying Language Model Generation with Stratification
Justin Wong (University of California Berkeley), Joseph E. Gonzalez (University of California Berkeley)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A training-free sampling method is proposed to enhance the response diversity of large language models in multi-answer tasks through steps such as automatic stratification, heuristic estimation, and probabilistic prompting.
Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning
Chongyu Fan (Michigan State University), Sijia Liu (IBM Research)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: A simple no-reference model-based negative preference optimization framework, SimNPO, is proposed for unlearning in large language models.
π― What it does: This paper presents SimulMEGA, an unsupervised policy learning framework based on a Mixture-of-Experts (MoE) router, designed to automatically determine read/write actions in simultaneous speech translation (SimulST), while being compatible with streaming inference for both speech-to-text and text-to-speech.
Simultaneous Statistical Inference for Off-Policy Evaluation in Reinforcement Learning
Tianpai Luo (Tsinghua University), Weichi Wu (Tsinghua University)
CodeReinforcement LearningBiomedical Data
π― What it does: A global confidence interval inference framework for offline reinforcement learning is proposed, which can simultaneously cover the value function estimates of the entire initial state space.
π― What it does: This paper proposes a natural gradient variational inference method called SING, which efficiently approximates posterior trajectories and learns drift functions in latent stochastic differential equation (SDE) models.
π― What it does: A method for adapting the Pathology Foundation Model (PFM) task on whole slide images (WSI) for gene mutation prediction is proposed, which can perform feature extraction and multi-instance learning (MIL) aggregation simultaneously on a single GPU.
π― What it does: This paper proposes a single-channel adaptive image tokenizer called KARL, which can predict the minimum number of tokens required for each image in a single forward pass to achieve optimal compression.
π― What it does: A single-step inverse operator learning framework is proposed, utilizing frequency domain information and semigroup properties to achieve one-time generation in conditional time series diffusion models, significantly reducing sampling steps.
π― What it does: By adding a lightweight linear branch to a single teacher network and using constrained angular and internal angle diversity loss, multi-angle multi-view knowledge enhancement is generated, thereby improving the effectiveness of knowledge distillation.
Antonio Ferrara (CENTAI), Francesco Bonchi (CENTAI)
CodeOptimizationTabularBenchmark
π― What it does: A unified, scale-adaptive hypothesis testing framework is proposed, transforming fairness assessment into statistical significance testing, accommodating both large and small sample subgroups.
Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks
Ryien Hosseini (University of Chicago), Henry Hoffmann (University of Chicago)
CodeGraph Neural NetworkGraph
π― What it does: The study proposes a GNN enhancement method based on Sketch random features to address long-range dependencies, over-smoothing, and insufficient expressiveness.
SkyLadder: Better and Faster Pretraining via Context Window Scheduling
Tongyao Zhu (National University of Singapore), Min-Yen Kan (National University of Singapore)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper explores the impact of context window size on the pre-training of large language models and proposes the SkyLadder method, which implements dynamic scheduling from short to long windows during the pre-training process through adjustable masking, balancing model performance and training efficiency.
Sloth: scaling laws for LLM skills to predict multi-benchmark performance across families
Felipe Maia Polo (University of Michigan), Mikhail Yurochkin (MBZUAI)
CodeLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: A scale law based on low-dimensional implicit skills, called Sloth, is proposed to predict the performance of different LLM families on multiple benchmarks without the need to train a large number of models from the same family.
SMARTraj$^2$: A Stable Multi-City Adaptive Method for Multi-View Spatio-Temporal Trajectory Representation Learning
Tangwen Qian (Chinese Academy of Sciences), Yongjun Xu (Chinese Academy of Sciences)
CodeRepresentation LearningRecurrent Neural NetworkGraph Neural NetworkTransformerContrastive LearningTime Series
π― What it does: A multi-view spatiotemporal trajectory representation learning method, SMARTrajβ―2, is proposed for achieving stable representation learning in multi-city environments.
Enrique Nueve (University of Colorado Boulder), Bo Waggoner (University of Colorado Boulder)
Code
π― What it does: A novel prediction market based on the Dual Cost Function Market (DCFMM) is proposed - the Smooth Quadratic Prediction Market. The theoretical properties, trader incentives, and behavior under budget constraints and buy-only conditions are analyzed within this framework. Furthermore, an adaptive liquidity extension method is proposed.
π― What it does: The SNAP framework is proposed, utilizing Sparse Test-Time Adaptation (STTA) to significantly reduce inference latency on edge devices while maintaining nearly unchanged model performance.
π― What it does: A backdoor attack framework named SNEAKDOOR is proposed, which implants covert triggers during the data distillation process based on distribution matching, allowing the downstream model to perform normally on regular inputs while being induced to misclassify when the trigger is activated.
Soft Task-Aware Routing of Experts for Equivariant Representation Learning
Jaebyeong Jeon (Yonsei University), Kibok Lee (Yonsei University)
CodeObject DetectionRepresentation LearningMixture of ExpertsContrastive LearningImage
π― What it does: Proposes Soft Task-Aware Routing (STAR) to coordinate invariant and equivariant representation learning, reducing redundant feature learning.
Solver-Free Decision-Focused Learning for Linear Optimization Problems
Senne Berden (KU Leuven), Tias Guns (KU Leuven)
CodeOptimizationTabular
π― What it does: This paper proposes a decision-focused learning method called LAVA, which constructs a convex loss to directly guide the training of prediction models by utilizing the geometric relationship between the optimal solution of linear optimization problems and its adjacent vertices, thereby significantly reducing the computational cost in predict-then-optimize scenarios.
Solver-Informed RL: Grounding Large Language Models for Authentic Optimization Modeling
Yitian Chen (Cardinal Operations), Yinyu Ye (Stanford University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: Proposes the Solver-Informed Reinforcement Learning (SIRL) framework, which utilizes verifiable rewards (by executing the generated code and .lp files through an optimization solver) to train large language models for the automated generation of accurate, executable optimization models, supplemented by instance-enhanced self-consistency data synthesis methods.
π― What it does: This study investigates the impact of optimizers (SGD, RMSProp, Adam, etc.) on the fairness of deep neural network populations, providing theoretical and empirical evidence that adaptive optimizers are more likely to converge to fair optimal solutions.
SONAR: Long-Range Graph Propagation Through Information Waves
Alessandro Trenta (University of Pisa), Davide Bacciu (University of Pisa)
CodeGraph Neural NetworkGraph
π― What it does: A differential equation graph neural network called SONAR based on the graph wave equation is proposed to achieve long-distance information propagation.
π― What it does: We propose a full song generation framework called SongBloom, which alternates between generating semantic sketches and audio details, supporting the generation of complete and structurally coherent songs from lyrics and short audio clips.
SORTeD Rashomon Sets of Sparse Decision Trees: Anytime Enumeration
Elif Arslan (Delft University of Technology), Emir DemiroviΔ (Delft University of Technology)
CodeOptimizationTabular
π― What it does: A framework named SORTD is proposed, which can enumerate the Rashomon set of sparse decision trees in ascending order of target values (i.e., best-first), thus obtaining an effective set at any moment.
SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement
Xiyao Wang (University of Maryland), Lijuan Wang (Microsoft)
CodeOptimizationReinforcement LearningVision Language ModelMultimodality
π― What it does: Significant improvements were achieved in visual language models through self-improvement via reinforcement learning, using a small number of training samples (11k/7.5k).
Rees Chang (University of Illinois at Urbana-Champaign), Ryan P Adams
CodeGenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelTabularPhysics Related
π― What it does: A space group invariant crystal generation model SGEquiDiff is proposed, which can naturally handle space group constraints and output space group invariant likelihoods.
SpaceServe: Spatial Multiplexing of Complementary Encoders and Decoders for Multimodal LLMs
zhicheng li, Huimin Cui (Chinese Academy of Sciences)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
π― What it does: The SpaceServe system is proposed, which separates visual/audio encoders from a shared text decoder in multimodal large language model services. It utilizes fine-grained GPU SM partitioning to achieve parallel spatial multiplexing of encoders and decoders on the same GPU, significantly improving inference throughput and latency performance.
Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
Mateusz Pach (Technical University of Munich), Zeynep Akata (Technical University of Munich)
CodeRepresentation LearningTransformerVision Language ModelAuto EncoderImageMultimodality
π― What it does: This paper addresses the reconstruction of layer activations using Sparse Autoencoders (SAE) for visual language models (such as CLIP) and proposes the MonoSemanticity (MS) score to quantitatively assess the univocality of neurons; it utilizes the obtained univocal neurons for concept insertion and suppression in multimodal large language models (such as LLaVA).
Sparse Gaussian Processes: Structured Approximations and Power-EP Revisited
Thang D Bui, Michalis Titsias
CodeTabular
π― What it does: A sparse Gaussian process approximation with a block diagonal structure is proposed, and this structure is applied to the Power Expectation Propagation framework to improve posterior approximation.
Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning
Yong Liu (National University of Singapore), Yang You (National University of Singapore)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes Sparse-MeZO, a memory-efficient zero-order optimization method for fine-tuning large-scale language models (such as LLaMA-7b/30b), significantly reducing memory requirements;
π― What it does: Sparse acceleration of attention computation for video generation Diffusion Transformers (DiT) is proposed, introducing the training-independent SVG2 framework, which utilizes semantic-aware permutation to accurately identify key tokens and achieves efficient sparse attention through dynamic Topp selection and custom kernels.
π― What it does: This paper proposes SparseDiT, which significantly reduces computational costs and improves inference speed by applying sparsification to the tokens in the Diffusion Transformer.
Sparta Alignment: Collectively Aligning Multiple Language Models through Combat
Yuru Jiang (Zhejiang University), Yulia Tsvetkov (University of Washington)
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: An algorithm named SPARTA ALIGNMENT is proposed, which collectively aligns multiple large language models (LLMs) through competition and adversarial methods. This algorithm overcomes the shortcomings of a single model in generating diversity and assessing bias by allowing multiple models to compete with each other.
π― What it does: This paper proposes embedding the Ring Attractor into the reinforcement learning decision-making process to explicitly encode spatial relationships into the action space, achieving more efficient action selection.
SpatialLM: Training Large Language Models for Structured Indoor Modeling
Yongsen Mao (Manycore Tech Inc), Zihan Zhou (Manycore Tech Inc)
CodeObject DetectionGenerationTransformerLarge Language ModelSupervised Fine-TuningPoint Cloud
π― What it does: Train and fine-tune the large language model SPATIALLM to accept 3D point clouds and generate structured scene descriptions containing walls, doors, windows, and 3D object bounding boxes (output in the form of Python scripts).
π― What it does: This paper proposes an unsupervised domain adaptation framework for person retrieval called STCSP, which can eliminate noise in pseudo-labels and gradually bridge the gap between the source domain and the target domain.
SPAZER: Spatial-Semantic Progressive Reasoning Agent for Zero-shot 3D Visual Grounding
Zhao Jin (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
CodeObject DetectionTransformerVision Language ModelPoint CloudRetrieval-Augmented Generation
π― What it does: This paper presents SPAzer, a zero-shot 3D visual localization agent that achieves the localization of target objects in 3D scenes through spatial-semantic progressive reasoning on a global 3D view rendered from multiple perspectives.
SpecEdge: Scalable Edge-Assisted Serving Framework for Interactive LLMs
Jinwoo Park, Dongsu Han
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: This paper proposes SpecEdge, a distributed framework that splits LLM inference into candidate token drafting on edge GPUs and validation on server GPUs; it achieves low-cost, high-throughput interactive LLM services through proactive edge drafting and pipeline-aware scheduling.
SpecEM: Training-Free LLM Ensembling via Iterative Drafting, Verification, and Online Feedback
Bo Lv (Peng Cheng Laboratory), Ping Luo (Peng Cheng Laboratory)
CodeGenerationTransformerLarge Language ModelText
π― What it does: SpecEM is proposed, a plug-and-play LLM integration framework that achieves paragraph-level semantic collaboration between models through an iterative draft-validate-online feedback mechanism.
SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning
Rui Pan (Princeton University), Ravi Netravali (Princeton University)
CodeComputational EfficiencyLarge Language ModelTextBenchmark
π― What it does: SpecReason is proposed to infer intermediate reasoning steps through a lightweight model and validate them with a base model, significantly accelerating LRM inference.
Spectral Compressive Imaging via Chromaticity-Intensity Decomposition
Xiaodong Wang (Zhejiang University), Xin Yuan (Westlake University)
CodeRestorationCompressionTransformerImage
π― What it does: The paper proposes a reconstruction framework CIDNet for spectral compressed imaging using chromatic-intensity decomposition in a dual-camera CASSI system.
π― What it does: This paper proposes Spectral ConvCNP, which utilizes frequency domain convolution to replace traditional local CNNs, enhancing the global dependency modeling capability of neural process models.
Siavash Ameli (University of California), Michael W. Mahoney (University of California)
CodeGraph
π― What it does: This paper proposes a method called free decompression, which uses the eigenvalues of randomly sampled submatrices to infer the spectral density of large Hermitian matrices that cannot be formed.
Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum
Snir Hordan (Technion Israel Institute of Technology), Nadav Dym (Technion Israel Institute of Technology)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes and analyzes the expressive power of Spectral Enhanced Graph Neural Networks (SGNN), proving that common models (such as EPNN) are still incomplete on graphs with only simple spectra (different eigenvalues);
Spectral Learning for Infinite-Horizon Average-Reward POMDPs
Alessio Russo (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
CodeReinforcement Learning
π― What it does: In the infinite average reward POMDP learning problem, this paper proposes the Mixed Spectral Estimation and Mixed Spectral UCRL algorithms, which can utilize samples from different adaptive belief-based policies to achieve more efficient model estimation and policy learning.
π― What it does: The spectral filter in the Spectral Transform Unit (STU) is transformed into an explicit Linear Dynamical System (LDS) through a provable sparse approximation method, enabling low-order recursive inference for long sequences, maintaining training stability and significantly reducing inference costs.
Speculate Deep and Accurate: Lossless and Training-Free Acceleration for Offloaded LLMs via Substitute Speculative Decoding
Pei-Shuo Wang (National Yang Ming Chiao Tung University), Kai-Chiang Wu (National Yang Ming Chiao Tung University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a training-free, lossless acceleration scheme called SubSpec, which utilizes low-bit quantization alternative layers, shared GPU layers, and KV-Cache to construct a draft model that is highly aligned with the target model, thereby achieving efficient Speculative Decoding in parameter offloading scenarios.
SpEx: A Spectral Approach to Explainable Clustering
Tal Argov (Tel Aviv University), Tal Wagner (Tel Aviv University)
CodeExplainability and InterpretabilityTabular
π― What it does: A general interpretable clustering method SPEX based on spectral graph partitioning is proposed, which can generate axis-aligned decision tree explanations for any reference clustering (regardless of whether it has centroids) or directly construct adjacency graphs from the data.
SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding
Yangxuan Zhou (Zhejiang University), Gang Pan (Zhejiang University)
CodeClassificationAnomaly DetectionSpiking Neural NetworkTime SeriesBiomedical Data
π― What it does: The SPICED framework is proposed for unsupervised continuous EEG decoding, enabling the model to dynamically expand and retain prior knowledge as new individuals continuously emerge.
SpiderSolver: A Geometry-Aware Transformer for Solving PDEs on Complex Geometries
Kai Qi (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
CodeTransformerMeshPhysics Related
π― What it does: We propose SpiderSolver, a Transformer that utilizes a spider web-style tokenization and coarse-fine attention to quickly solve partial differential equations under complex geometric boundaries.
π― What it does: This paper proposes a spike neural network based on attention, SpikeSR, for efficient super-resolution reconstruction of remote sensing images (RSI).
π― What it does: This paper proposes and verifies that Spiking Neural Networks (SNN) are essentially low-pass filters in the frequency domain, leading to the attenuation of high-frequency information; based on this, Max-Former (a Transformer structure) and Max-ResNet (a CNN structure) are designed to recover high-frequency information and enhance performance.
SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing
Sung-Hoon Yoon (Harvard University), Mengyu Wang (Harvard University)
CodeImage TranslationGenerationLarge Language ModelFlow-based ModelRectified FlowImageTextBenchmark
π― What it does: Proposes SplitFlow, which achieves high-quality text-to-image editing by semantically decomposing the target text prompt, calculating independent editing flows, and aggregating them.
Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation
Siwei Wen (Shanghai Artificial Intelligence Laboratory), Weijia Li
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes FakeVLM, a large multimodal model specifically designed for synthetic image authenticity discrimination and anomaly trace explanation, and constructs the FakeClue dataset, which contains 100k multi-category images and fine-grained natural language anomaly descriptions.
π― What it does: The SPOT-Trip framework is proposed to generate a complete itinerary (including an intermediate POI sequence) when users travel from their hometown to unknown areas, explicitly learning users' static preferences (long-term interests) and dynamic preferences (short-term interests).
π― What it does: Proposes the SPOT framework, which utilizes a MILP-based decision tree policy optimization combined with policy iteration and reduced space branch and bound to solve interpretable decision tree policies in Markov decision processes.
Spotlight Attention: Towards Efficient LLM Generation via Non-linear Hashing-based KV Cache Retrieval
Wenhao Li (Xiamen University), Rongrong Ji (Xiamen University)
CodeGenerationRetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Proposes Spotlight Attention, which dynamically selects important key-value pairs from the KV cache using a nonlinear hash function, significantly accelerating the inference of LLMs.
SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning
Peixian MA, Jian Guo (International Digital Economy Academy)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Proposes SQL-R1, a NL2SQL reasoning model trained through reinforcement learning, capable of generating interpretable SQL statements in complex multi-table queries.
Russell Tsuchida (Monash University), Dino Sejdinovic (University of Adelaide)
CodeTabular
π― What it does: This paper is the first to use squared families as prior distributions, constructing the Generalised Squared Families (GSF) and proving its conjugacy for various likelihoods. Utilizing this conjugacy, closed-form posteriors and marginal likelihoods are obtained. Subsequently, the GSF is combined with neural network features to propose the GSF Process (GSFP) model for Bayesian regression in feature space, which can naturally express multimodal uncertainty and achieve excellent performance in few-shot learning tasks.
SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation
Ziqiang Cui (City University of Hong Kong), Chen Ma (City University of Hong Kong)
CodeRecommendation SystemTransformerLarge Language ModelContrastive LearningSequential
π― What it does: A framework called SRA-CL based on semantic retrieval enhanced contrastive learning has been developed to improve sequential recommendation models.
π― What it does: This paper studies a method for colorizing anime facial line art based on a continuous time diffusion model, proposing the SSIMBaD framework, which achieves consistency between forward noise and reverse recovery through a perceptually unified Sigma space transformation, and incorporates lightweight trajectory refinement.
Stab-SGD: Noise-Adaptivity in Smooth Optimization with Stability Ratios
David A. R. Robin (INRIA - ENS Paris PSL Research University), Kevin Scaman (INRIA - ENS Paris)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: Introduced a 'stability ratio' as a local relative noise measure, and based on it, proposed a Stab-SGD algorithm that is unscheduled and can adapt to noise levels.
π― What it does: A stable end-to-end physics-guided neural network (sPHNN) is proposed, which can learn globally asymptotically stable nonlinear dynamics from sparse data.
StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models
Haoxin Yang (South China University of Technology), Shengfeng He (Singapore Management University)
CodeGenerationData SynthesisAnomaly DetectionSafty and PrivacyTransformerMixture of ExpertsDiffusion modelAuto EncoderImage
π― What it does: A global binary watermark is embedded in the generation process of the latent diffusion model (LDM), and a unified copyright protection and tampering localization framework (StableGuard) is constructed through self-supervised joint training.
Brandon R. Feng (North Carolina State University), Brian Reich
CodeTime Series
π― What it does: The STACI method is proposed, which uses variational Bayesian neural networks to approximate non-stationary spatiotemporal Gaussian processes and infers statistically valid confidence intervals through local spatiotemporal consistency, thereby achieving large-scale spatiotemporal interpolation and reliable uncertainty quantification.
Stackelberg Self-Annotation: A Robust Approach to Data-Efficient LLM Alignment
Xu Chu (Peking University), Yujie Jin (Peking University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: A robust alignment framework SGPO based on Stackelberg game is proposed, and a self-labeling, data-efficient SSAPO algorithm is implemented, achieving LLM alignment with very few human labels (about 2K).
π― What it does: The STAIR method is proposed, which approximates phase differences using temporal distance and selects only phase-aligned segments in preference learning, thereby completing multi-stage tasks more efficiently.
π― What it does: A scalable latent space normal flow model STARFlow based on Transformer autoregressive flow has been developed for high-resolution image generation.
State Space Prompting via Gathering and Spreading Spatio-Temporal Information for Video Understanding
Jiahuan Zhou (Wangxuan Institute of Computer Technology Peking University), Gang Hua (Amazon.com)
CodeRecognitionPrompt EngineeringVideo
π― What it does: A Prompt learning framework called State Space Prompting (SSP) is proposed for the pre-trained state space model (VideoMamba), which achieves efficient video understanding by aggregating local spatial information within each frame and propagating global temporal information across frames.
State-Covering Trajectory Stitching for Diffusion Planners
Kyowoon Lee (KAIST), Jaesik Choi (KAIST)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelSequential
π― What it does: The SCoTS method is proposed, which generates diverse and long-term trajectories through reward-free trajectory stitching and expansion on offline datasets to enhance the performance of diffusion planners.
Statistics Caching Test-Time Adaptation for Vision-Language Models
Zenghao Guan (Institute of Information Engineering Chinese Academy of Sciences), Xiaoyan Gu (Institute of Information Engineering Chinese Academy of Sciences)
CodeDomain AdaptationComputational EfficiencyTransformerVision Language ModelImage
π― What it does: A statistical cache-based adaptive method SCA is designed, which utilizes feature statistics instead of raw features to achieve continuous knowledge accumulation, and enhances the performance of VLM on unlabeled test data through soft pseudo-labels and instance-level adaptive fusion.
STEAD: Robust Provably Secure Linguistic Steganography with Diffusion Language Model
Yuang Qi (University of Science and Technology of China), Kejiang Chen (University of Science and Technology of China)
CodeDiffusion modelText
π― What it does: This paper proposes a robust and provably secure language steganography method called STEAD, which can locate robust embedding positions during the sequence generation process and achieve robust extraction.
SteerConf: Steering LLMs for Confidence Elicitation
Ziang Zhou (Hong Kong Polytechnic University), Li Qing
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: The SteerConf framework is proposed, which guides the self-expression confidence of LLMs by adjusting the prompt level, and calibrates confidence using consistency of confidence and answer consistency to improve confidence calibration and error prediction.
Steering Generative Models with Experimental Data for Protein Fitness Optimization
Jason Yang (California Institute of Technology), Yisong Yue (California Institute of Technology)
CodeOptimizationDrug DiscoveryReinforcement LearningDiffusion modelBiomedical Data
π― What it does: This study investigates a generative model-guided framework (SGPO) based on a small number of experimental labels for protein adaptive optimization.
Steering Information Utility in Key-Value Memory for Language Model Post-Training
Chunyuan Deng (Rice University), Hanjie Chen (Rice University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the InfoSteer method, which encourages the model to better utilize pre-trained knowledge during the post-training phase of LLM by performing forward intervention on the key-value structure of FFN and applying entropy regularization.
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes a flexible activation scheduling and backtracking framework (FASB) to dynamically determine whether and how to intervene in the activation of LLM during the inference process.
StegoZip: Enhancing Linguistic Steganography Payload in Practice with Large Language Models
Jun Jiang (University of Science and Technology of China), Kejiang Chen (University of Science and Technology of China)
CodeCompressionSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the StegoZip framework, which utilizes large language models for dynamic semantic redundancy pruning and index compression encoding of secret information, in order to enhance the payload of linguistic steganography while maintaining decodability.
π― What it does: Based on LoRA, StelLA is proposed, which performs three-factor decomposition of USVα΅ and optimizes U and V on the Stiefel manifold, explicitly learning input/output subspaces to achieve parameter-efficient fine-tuning.