Conference on Neural Information Processing Systems Β· 2283 papers
RoMa: A Robust Model Watermarking Scheme for Protecting IP in Diffusion Models
Yingsha Xie (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: A robust model watermarking scheme named RoMa is proposed, which enhances the persistence of the watermark against fine-tuning by embedding triggers in the diffusion model and introducing path-specific smoothness.
RoME: Domain-Robust Mixture-of-Experts for MILP Solution Prediction across Domains
Tianle Pu (National University of Defense Technology), Changjun Fan (National University of Defense Technology)
CodeOptimizationGraph Neural NetworkMixture of ExpertsGraph
π― What it does: The RoME framework is proposed, utilizing Mixture-of-Experts and distributionally robust optimization to achieve training and inference of cross-domain MILP solvers.
π― What it does: A public furniture synthesis benchmark dataset, RoomBench, has been constructed, and a parameter-shared dual U-Net network, RoomEditor, has been proposed to achieve high-fidelity and geometrically consistent indoor furniture insertion.
Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping
Justin Lazarow (Apple), Afshin Dehghan (Apple)
CodeObject DetectionPose EstimationTransformerSimultaneous Localization and MappingOptical FlowImage
π― What it does: Without relying on explicit point clouds or depth maps, this paper utilizes 3D bounding boxes as primitives to achieve 3D object detection, camera localization, and global semantic map construction from unordered RGB (or RGB-D) image sets in indoor spaces.
Root Cause Analysis of Outliers with Missing Structural Knowledge
William Roy Orchard (University of Cambridge), Dominik Janzing (Amazon)
CodeAnomaly DetectionTabular
π― What it does: This paper proposes two root cause analysis methods for single-sample anomaly situations that do not require estimating conditional distributions or complete structural models, avoiding the statistical infeasibility of traditional methods when there is a lack of a large number of posterior samples.
π― What it does: By treating offline black-box optimization as a distribution translation task, a probabilistic bridge model is proposed, utilizing synthetic Gaussian process functions for pre-training, generating low-value to high-value design transfer paths, ultimately producing better candidate designs.
Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection
Zheng Zhan (Microsoft), yelong shen
CodeMixture of ExpertsText
π― What it does: Proposes the Routing Mamba (RoM) framework, applying sparse experts to the projection layer of state space models like Mamba to achieve efficient scaling.
RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards
Jingnan Zheng (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeSafty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: A security monitoring model based on active inference, RSafe, has been designed and implemented. It can perform step-by-step reasoning on inputs within the user-specified security policy range and provide interpretable security judgments.
rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset
Yifei Liu (University of Science and Technology of China), Mao Yang (Microsoft Research Asia)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A large-scale, high-quality competitive programming dataset rStar-Coder has been constructed, containing 418K problems and 580K long reasoning answers, and solutions are validated through reliable test case generation and mutual verification mechanisms.
Chenlong Zhang (Chinese Academy of Sciences), Yubo Chen (Chinese Academy of Sciences)
CodeOptimizationLarge Language ModelReinforcement LearningText
π― What it does: Proposes the RULE framework, which utilizes reinforcement learning for selective forgetting in large language models, focusing on learning rejection strategies to delineate the boundaries of forgetting and retention, thereby avoiding unnatural or crashing responses from the model when faced with forgetting queries.
S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning
Hanqing Zeng (Meta AI), Benyu Zhang (Meta AI)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
π― What it does: A parameter-efficient LLM fine-tuning framework named S'MoRE is proposed and implemented, integrating the low-rank efficiency of LoRA with the plasticity of MoE.
π― What it does: This paper studies the training dynamics of fully connected feedforward neural networks under continuous time gradient flow, proving a dichotomy: the gradient flow either converges to a critical point or approaches a generalized critical value as the parameters go to infinity. It further shows that when the loss is below a certain threshold, it will necessarily converge to the optimal value. For polynomial objective functions, the authors prove that in the case of sufficiently large networks and datasets, the loss cannot reach zero (there is no global minimum), but can get arbitrarily close to zero, leading to the inference that the gradient flow must diverge. Numerical experiments validate this divergence phenomenon and extend it to more practical tasks such as PDE solving and MNIST image classification.
Safe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models
Chenhang Cui (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeGenerationAdversarial AttackTransformerLarge Language ModelAgentic AIVision Language ModelDiffusion modelImageMultimodality
π― What it does: This study investigates an attack method that uses seemingly safe images and prompts to induce large visual language models (LVLM) to generate unsafe content; it proposes a Safe Avalanche Agent (SSA) framework that utilizes the model's general reasoning capabilities and the safe avalanche effect for staged attacks;
Safe-Sora: Safe Text-to-Video Generation via Graphical Watermarking
Zihan Su (Tsinghua University), Fei Yu
CodeGenerationData SynthesisSafty and PrivacyVideoText
π― What it does: Directly embed visible graphic watermarks in the text-to-video generation framework to achieve copyright protection for generated videos.
SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism
Beitao Chen (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China), Lianli Gao (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelVision Language ModelMultimodality
π― What it does: This paper proposes an untrained SafePTR framework to address multimodal jailbreak attacks in multimodal large language models (MLLMs) by pruning harmful tokens at vulnerable layers and restoring beneficial features in subsequent layers to enhance security.
SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification
ZhengLin Lai, Jianqiang Li (Shenzhen University)
CodeSafty and PrivacyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: For the Mixture-of-Experts (MoE) architecture of LLMs, a systematic analysis and quantification of its 'positional vulnerability' in safety alignment is conducted, and the SAFEX framework is proposed to identify, locate, and verify safety-critical experts; through expert-level masking and LoRA fine-tuning, the model's rejection rate of harmful prompts is validated and improved.
π― What it does: Designed an FP4 micro-scale attention implementation of SageAttention3 for inference acceleration, and proposed a trainable 8-bit attention for training acceleration in SageBwd.
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningTextAudio
π― What it does: This paper presents SALMONN-omni, a single LLM capable of full-duplex independent processing of speech input and output without audio stream injection.
SALoM: Structure Aware Temporal Graph Networks with Long-Short Memory Updater
Hanwen Liu (Zhejiang University), Mingli Song (Zhejiang University)
CodeRecurrent Neural NetworkGraph Neural NetworkMixture of ExpertsGraphTime SeriesOrdinary Differential Equation
π― What it does: The SALoM framework is proposed, utilizing continuous time memory modules and Long Short-Term Memory Updater (LSMU), combining co-occurrence encoding and information bottleneck fusion to achieve unified modeling of long-term and short-term temporal dependencies as well as structural information in dynamic graphs.
SAM2Flow: Interactive Optical Flow Estimation with Dual Memory for in vivo Microcirculation Analysis
Luojie Huang (Johns Hopkins University), Nicholas J. Durr (Johns Hopkins University)
CodeRecurrent Neural NetworkTransformerOptical FlowVideoBiomedical Data
π― What it does: An interactive optical flow estimation model, SAM2Flow, is proposed to extract microvascular blood flow information from OBM videos, allowing users to specify regions of interest (ROI) through point prompts.
SAMA: Towards Multi-Turn Referential Grounded Video Chat with Large Language Models
Ye Sun (Fudan University), Yu-Gang Jiang (Fudan University)
CodeObject DetectionSegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningVideoTextBenchmark
π― What it does: Developed a multi-turn video dialogue system SAMA, providing unified video localization understanding, a video grounding dataset, model architecture, and benchmarks;
Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs
Yaniv Nikankin (Technion Israel Institute of Technology), Yonatan Belinkov (Technion Israel Institute of Technology)
CodeRepresentation LearningTransformerVision Language ModelTextMultimodality
π― What it does: By comparing the computational subgraphs (circuits) of visual and text tasks, this study investigates the performance gap between visual and text in visual language models under the same task, and proposes a back-patching method during testing to improve the accuracy of visual tasks.
Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N Sampling in Early Decoding
Yiming Wang (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: The paper proposes an ST-BoN decoding method that allows large language models to perform inference by self-estimating and truncating the least optimal samples in advance, without needing to fully generate N samples or relying on a reward model.
π― What it does: This study investigates the potential semantic structure of SAM2 in few-shot segmentation and proposes the SANSA framework, which transforms SAM2 from visual matching to semantic matching through a lightweight adapter. It utilizes its built-in Memory Attention mechanism to achieve k-shot segmentation tasks with a single inference for various prompts.
π― What it does: The SCENT framework is proposed, combining template-based synthesis with GFlowNet for synthesizable and low-cost molecular generation, addressing the three major challenges of synthesis cost, library scale, and small fragment sets.
Scalable Best-of-N Selection for Large Language Models via Self-Certainty
Zhewei Kang, Dawn Song
CodeTransformerLarge Language ModelTextChain-of-Thought
π― What it does: Designed and evaluated a self-certainty metric based on the probability distribution of LLM itself for Best-of-N selection, and proposed a self-certainty weighting method based on Borda voting.
Yingru Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
CodeTransformerReinforcement Learning
π― What it does: The Ensemble++ framework is proposed, which achieves scalable approximate Thompson Sampling using shared factors and random linear combinations;
π― What it does: This paper studies a scalable knowledge graph embedding algorithm called SEPAL, which generates vector features suitable for downstream machine learning tasks on large-scale knowledge graphs.
Scalable In-context Ranking with Generative Models
Nilesh Gupta (University of Texas at Austin), Felix X. Yu
CodeRetrievalOptimizationTransformerLarge Language ModelContrastive LearningText
π― What it does: This paper proposes BlockRank, an efficient In-Context Ranking method for large language models, which utilizes structured sparse attention and contrastive auxiliary attention loss to achieve relevance ranking of candidate lists in a single forward pass.
Scalable Signature Kernel Computations via Local Neumann Series Expansions
Matthew Tamayo-Rios (ETH Zurich), Rima Alaifari (RWTH Aachen University)
CodeOptimizationComputational EfficiencyTime SeriesSequentialFinance Related
π― What it does: A scalable signature kernel computation method based on local Neumann series expansion (PowerSig) is proposed, capable of handling high-dimensional long sequences with millions of points on a single GPU;
Scalable Valuation of Human Feedback through Provably Robust Model Alignment
Masahiro Fujisawa (University of Osaka), Michael A Osborne
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
π― What it does: This paper proposes HΓΆlder-DPO, a variant of Direct Preference Optimization (DPO) based on HΓΆlder divergence, which achieves provable redescending robustness under label noise and automatically estimates and locates mislabeling.
Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching
Zhong Li (Leiden University), Matthijs van Leeuwen (Leiden University)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyFlow-based ModelTabularBenchmark
π― What it does: A time-conditioned contraction matching (TCCM) algorithm based on flow matching is proposed for semi-supervised tabular data anomaly detection.
Florian Redhardt (ETH Zurich), Simon Schug (Princeton University)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: The study achieves generalization over a family of composable tasks by expanding the data volume and model scale (mainly multi-layer perceptrons) and demonstrates that under a training distribution that sufficiently covers the task space, the network can perform well on new combinatorial tasks.
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: A Caco framework was constructed to automatically generate verifiable, scalable, and diverse reasoning data for training large language models through code-assisted chain reasoning.
π― What it does: Conduct a scalable, data-driven probabilistic robustness analysis of semantic segmentation networks (SSN), utilizing Conformal Inference and surrogate models to construct observable starsets and determine the robust/non-robust/unknown status of pixels.
π― What it does: This paper extends Maximum Update Parameterization (Β΅P) from standard Transformers to diffusion Transformers and validates its feasibility on various mainstream diffusion models.
Scaling Epidemic Inference on Contact Networks: Theory and Algorithms
Guanghui Min (University of Virginia), Chen Chen (University of Virginia)
CodeGraph
π― What it does: This paper proposes a fast infection probability inference method based on residual propagation, called RAPID, for large-scale contact networks under the SIR model.
Chenghao Xiao (DAMO Academy), Yu Rong (Alibaba Group)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningVideoTextMultimodalityAudio
π― What it does: This paper proposes a language-centered full-modal embedding framework LCO-EMB, which fine-tunes a multimodal large language model (MLLM) using contrastive learning with only text to enhance cross-modal representation quality, and explores the Generative-Representation Scaling Law (GRSL).
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityChain-of-Thought
π― What it does: A complete framework is proposed to enable visual language models (VLM) to perform deep reasoning on long videos, covering a large-scale long video reasoning dataset, two-stage training (chain-of-thought supervised fine-tuning + reinforcement learning), and multi-modal reinforcement learning parallel techniques for long videos.
Scaling Unlocks Broader Generation and Deeper Functional Understanding of Proteins
Aadyot Bhatnagar (Profluent Bio), Ali Madani (Profluent Bio)
CodeGenerationProtein Structure PredictionTransformerLarge Language ModelMixture of ExpertsBiomedical Data
π― What it does: Trained and evaluated the sparse generative protein language model ProGen3, with parameters ranging from 112M to 46B. The study systematically explored the impact of model size on protein generation, diversity, laboratory expression rates, and functional predictions, and validated the expressibility of the generated proteins in the laboratory.
π― What it does: This paper proposes LrcSSM, a nonlinear state space model that transforms liquid resistor-capacitor networks (LRC) into a parallel computable form;
π― What it does: A framework called RPG based on cyclic diffusion is proposed, which can generate complete neural network models with hundreds of millions of parameters on a single GPU.
ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection
Tao Yin (Chongqing University), Meng Yan (Chongqing University)
CodeAnomaly DetectionGraph Neural NetworkContrastive LearningTime Series
π― What it does: This paper proposes the ScatterAD method, which utilizes a spatiotemporal scatter mechanism for anomaly detection in industrial IoT multivariate time series data.
scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration
Jianle Sun (Shanghai Artificial Intelligence Laboratory), Peng Ye (Guangzhou Laboratory)
CodeOptimizationData-Centric LearningAuto EncoderGenerative Adversarial NetworkMultimodalityBiomedical Data
π― What it does: A framework for unpaired single-cell multi-omics data integration named scMRDR is proposed, which achieves the decoupling of shared and specific subspaces through Ξ²-VAE, and incorporates isometric regularization, adversarial alignment, and masked reconstruction loss.
SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs
Jinhong Deng (University of Electronic Science and Technology of China), Yang He (Agency for Science, Technology and Research)
CodeComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
π― What it does: A visual token pruning method named SCOPE is proposed, which preserves semantic integrity by considering both the significance of tokens and their semantic coverage.
scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery
Yiming Gao (Texas A and M University), Eric P. Xing (Carnegie Mellon University)
CodeDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringBiomedical DataBenchmark
π― What it does: The SCPILOT framework is proposed to achieve 'omics-native reasoning' of large language models in single-cell RNA sequencing analysis, automating cell type annotation, developmental trajectory reconstruction, and transcription factor target prediction, while providing transparent reasoning pathways.
scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy
Ashesh, Florian Jug (Human Technopole)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: The scSPLIT method is proposed for the decomposition and removal of bleed-through in fluorescence microscopy images, capable of simultaneously handling structural separation of single-channel images.
π― What it does: A reward-guided hierarchical diffusion model DIFFT is proposed to automatically generate the task-optimal feature transformation sequence.
SD-VLM: Spatial Measuring and Understanding with Depth-Encoded Vision-Language Models
Pingyi Chen (Zhejiang University), Jieping Ye (Alibaba Cloud Computing)
CodeObject DetectionDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
π― What it does: This study investigates the insufficient spatial reasoning of VLM, proposing the MSMU dataset and deep position encoding to enhance spatial measurement and understanding.
π― What it does: A gradient-based self-distillation framework SDPGO is proposed, which utilizes proximal gradient optimization to dynamically weight features and achieves self-distillation through sequential iterative learning.
π― What it does: A model called SDTagNet is proposed for online HD map construction using the full textual annotations from OpenStreetMap and all map elements.
SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning
Xinbin Yuan (Nankai University), Bo Li (Vivo Mobile Communication Company)
CodeReinforcement LearningVision Language ModelImage
π― What it does: A training set consisting of only 3,000 high-quality samples was constructed, and reinforcement learning was used for self-evolution refinement, enhancing the visual localization capability of the GUI agent.
Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning
Yaorui Shi (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
CodeRetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: The AutoRefine framework is proposed in retrieval-augmented reasoning, improving the reasoning quality of LLMs by incorporating an explicit knowledge refinement step between retrieval steps.
SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
Hwiwon Lee (University of Illinois Urbana-Champaign), LINGMING ZHANG
CodeLarge Language ModelAgentic AIBenchmark
π― What it does: An automated evaluation framework for LLM security agents, SEC-bench, has been constructed, capable of automatically collecting, reproducing, and verifying vulnerabilities from real C/C++ CVEs.
SECA: Semantically Equivalent and Coherent Attacks for Eliciting LLM Hallucinations
Buyun Liang (University of Pennsylvania), Rene Vidal
CodeGenerationOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes a zero-order optimization method named SECA, which induces hallucinations in large language models (LLMs) by generating semantically equivalent and coherent rephrased prompts.
See&Trek: Training-Free Spatial Prompting for Multimodal Large Language Model
Pengteng Li, Hui Xiong
CodeRecognitionObject DetectionAutonomous DrivingLarge Language ModelPrompt EngineeringVision Language ModelSimultaneous Localization and MappingImageVideoMultimodalityBenchmark
π― What it does: This paper presents SEE&Trek, a training-free and GPU-free spatial prompting framework designed to enhance the spatial understanding capabilities of multimodal large language models under purely visual conditions.
π― What it does: The NEURO framework is proposed, which combines visual perception networks with robust optimization to achieve end-to-end training of visual navigation agents.
Seemingly Redundant Modules Enhance Robust Odor Learning in Fruit Flies
HaiYang Li, Yunliang Zang (Xiamen Intretech Inc)
CodeSpiking Neural NetworkReinforcement LearningTime Series
π― What it does: This study investigates the effects of lateral inhibition (LI) and spike frequency adaptation (SFA) as two sparsification mechanisms on olfactory discrimination learning under different noise levels in the fruit fly olfactory circuit, and compares their performance under noise-enhanced conditions through computational modeling.
SeerAttention: Self-distilled Attention Gating for Efficient Long-context Prefilling
Yizhao Gao (University of Hong Kong), Mao Yang (Microsoft Research)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsText
π― What it does: This paper proposes SeerAttention, a self-supervised sparse attention mechanism that learns block-level sparse gating (AttnGate) in each attention head of LLMs, and achieves efficient long text pre-filling inference by combining block-sparse FlashAttention kernels.
SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation
Yueyang Hu (University of Chinese Academy of Sciences), Hao Pan (Tsinghua University)
CodeSegmentationGraph Neural NetworkPoint Cloud
π― What it does: Utilizing the segmentation information generated by SAM, a segmentation graph is constructed to achieve few-shot 3D part segmentation.
Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models
Yiran Guo (Institute of Software, Chinese Academy of Sciences), Shuang Qiu (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: In response to the reinforcement learning training of large language models (LLM) in reasoning tasks, the Segment Policy Optimization (SPO) framework is proposed, utilizing intermediate-granularity segment-level advantage estimation to achieve more accurate credit allocation. The SPO-chain is designed for short chain-of-thought (CoT) scenarios, while the SPO-tree is designed for long CoT scenarios.
Selective Learning for Deep Time Series Forecasting
Yisong Fu (Institute of Computing Technology Chinese Academy of Sciences), Fei Wang (Institute of Computing Technology Chinese Academy of Sciences)
CodeAnomaly DetectionOptimizationTransformerTime Series
π― What it does: A new learning strategy called Selective Learning is proposed, which suppresses the model's overfitting to noise and anomalous moments by only calculating loss at moments that can generalize in deep time series forecasting.
Self Iterative Label Refinement via Robust Unlabeled Learning
Hikaru Asano (University of Tokyo), Yukino Baba (University of Tokyo)
CodeClassificationOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposes a self-iterative label improvement framework that uses two groups of unlabeled data with different positive-negative ratios to iteratively denoise and refine the pseudo-labels generated by LLM through robust UU learning.
π― What it does: A self-supervised learning framework Lfp2vec has been developed to achieve real-time localization of brain regions using raw local field potential (LFP) signals from multi-channel microelectrode arrays, which can be transferred to downstream tasks such as disease classification.
Self-Calibrating BCIs: Ranking and Recovery of Mental Targets Without Labels
Jonathan Grizou (GrizAI University of Glasgow), Tuukka Ruotsalo (LUT University University of Copenhagen)
CodeGenerationRetrievalOptimizationGenerative Adversarial NetworkImageMultimodalityTime Series
π― What it does: This paper proposes an unsupervised self-calibrating brain-computer interface framework named CURSOR, which can recover the target image that the user is thinking of solely through paired EEG and image embeddings, without the need for labels or a pre-trained decoder.
Self-Evolving Pseudo-Rehearsal for Catastrophic Forgetting with Task Similarity in LLMs
Jun Wang (Wuhan University), Bo Du (Wuhan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A Self-Evolving Pseudo-Rehearsal framework (SERS) is proposed to address the catastrophic forgetting problem in large language models through self-evolving pseudo-sample replay and task similarity-driven dynamic regularization.
Self-Refining Language Model Anonymizers via Adversarial Distillation
Kyuyoung Kim (KAIST), Jinwoo Shin (KAIST)
CodeSafty and PrivacyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Train small language models to achieve text anonymization without relying on external large models, and implement iterative improvements through self-assessment.
π― What it does: A self-supervised visual pre-training framework SΒ²VM is proposed for jointly encoding molecular images of drug pairs and reconstructing the original images, thereby enhancing drug-drug interaction (DDI) prediction.
π― What it does: This paper studies the theoretical relationship between self-supervised contrastive learning (CL) and its supervised counterpart (the supervised contrastive loss with only negative samples, NSCL). It proves that when the number of classes is sufficiently large, the losses of the two are similar, and provides a geometric characterization of the global minimization solution of NSCL. Furthermore, it presents a new upper bound on the few-shot error and validates the theoretical predictions on various datasets.
π― What it does: This paper presents DISCOVR, a self-supervised learning framework that combines video self-distillation and online image guidance, capable of learning rich spatiotemporal representations in unlabeled cardiac ultrasound videos and performing downstream tasks such as anomaly detection, classification, and segmentation.
π― What it does: A joint model called GraphIDS, which combines self-supervised graph neural networks and Transformer-based masked autoencoders, is proposed for network traffic anomaly detection.
Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration
Wenjie Li (Beijing University of Posts and Telecommunications), Zhanyu Ma (Beijing University of Posts and Telecommunications)
CodeRestorationDiffusion modelImage
π― What it does: To address the problem of restoring faces in old photographs, a self-supervised selective guidance diffusion model (SSDiff) is proposed, which utilizes weak guidance to generate pseudo-reference facial images and conducts staged structural and color guidance during the reverse diffusion process.
Self-Training with Dynamic Weighting for Robust Gradual Domain Adaptation
Zixi Wang (University of Electronic Science and Technology of China), Xin Lai (New Jersey Institute of Technology)
CodeDomain AdaptationImage
π― What it does: A self-training with dynamic weighting (STDW) method is proposed for progressive domain adaptation, addressing issues such as unsmooth knowledge transfer and instability during the transition phase.
π― What it does: This study investigates the effectiveness of small Transformers in enhancing multi-step Chain-of-Thought reasoning through self-verification reflection in non-natural language environments, providing both theoretical and experimental analysis.
Semantic Representation Attack against Aligned Large Language Models
Jiawei Lian (Northwestern Polytechnical University), Lap-Pui Chau (Hong Kong Polytechnic University)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A novel attack method aimed at aligned large language models is proposedβSemantic Representation Attack, which efficiently generates semantically consistent and natural attack prompts.
π― What it does: A zero-shot, inference-time text embedding vector subtraction framework called Semantic Surgery is proposed, which can dynamically remove unwanted concepts from text prompts without changing the model parameters.
Semantic-guided Diverse Decoding for Large Language Model
Weijie Shi (Hong Kong University of Science and Technology), Xiaofang Zhou (Hong Kong University of Science and Technology)
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
π― What it does: Proposes the SemDiD method, which achieves semantic-level diverse decoding in large language models, directly guiding generation in the embedding space;
SemCoT: Accelerating Chain-of-Thought Reasoning through Semantically-Aligned Implicit Tokens
Yinhan He (University of Virginia), Jundong Li (University of Virginia)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
π― What it does: Proposes the SemCoT framework, which utilizes semantically aligned implicit Chain-of-Thought (CoT) tokens to accelerate the reasoning process of large language models (LLMs).
Cody Melcher (University of Arizona), Erfan Yazdandoost Hamedani (University of Arizona)
CodeOptimizationTabular
π― What it does: This paper studies semi-infinite non-convex min-max optimization problems and proposes an incomplete dynamic barrier primal-dual algorithm (iDB-PD) to solve non-convex min-max problems with infinite constraints.
π― What it does: A semi-supervised graph anomaly detection method called RHO is proposed, which learns anomalous nodes in the graph using a small number of labeled normal nodes.
Semi-Supervised Regression with Heteroscedastic Pseudo-Labels
Xueqing Sun (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
CodeOptimizationImage
π― What it does: A pseudo-labeling framework for uncertainty-aware semi-supervised regression is proposed, which can dynamically adjust the heteroscedastic uncertainty of pseudo-labels through bi-level optimization, thereby mitigating the negative impact of incorrect pseudo-labels on the model.
SEMPO: Lightweight Foundation Models for Time Series Forecasting
Hui He (Beijing Institute of Technology), Guansong Pang (Singapore Management University)
CodeTransformerPrompt EngineeringTime Series
π― What it does: A lightweight foundational model SEMPO is proposed, utilizing energy-aware spectral decomposition and a mixture of prompts Transformer to achieve strong generalization in time series prediction with a small amount of pre-training data.
π― What it does: This paper proposes seq-JEPA, a world model that learns visual representations with both invariance and equivariance by autoregressively predicting action-observation sequences.
CodeGenerationData SynthesisComputational EfficiencyFlow-based ModelTime SeriesSequentialFinance RelatedPhysics Related
π― What it does: A sequence modeling method based on operator theory is proposedβSpectral Mean Flows, which embeds sequence distributions into Hilbert space using the linear operator mapping of HMM, and defines MMD gradient flows for generation in this space;
Sequential Attention-based Sampling for Histopathological Analysis
Tarun G (Indian Institute of Science), Devarajan Sridharan (Indian Institute of Science)
CodeComputational EfficiencyTransformerReinforcement LearningImageBiomedical Data
π― What it does: The SASHA framework is proposed, combining multi-attention multi-instance learning with reinforcement learning to achieve efficient sequential sampling and diagnosis of panoramic tissue sections.
Sequential Monte Carlo for Policy Optimization in Continuous POMDPs
Hany Abdulsamad (University of Amsterdam), Simo SΓ€rkkΓ€ (Aalto University)
CodeOptimizationReinforcement LearningSequential
π― What it does: A continuous POMDP policy optimization algorithm (P3O) based on the Feynman-Kac model and nested particle filtering is proposed, which can directly model and learn the joint Bayesian inference and decision-making process.
Chen Lu (Nanjing University), Chao Qian (Nanjing University)
CodeOptimizationReinforcement LearningSequential
π― What it does: In response to the inherent dependencies of multiple parameters in Dynamic Algorithm Configuration (DAC), the Seq-MADAC framework is proposed, and a Sequence Advantage Decomposition Network (SADN) is designed to achieve multi-agent dynamic configuration.
SeRL: Self-play Reinforcement Learning for Large Language Models with Limited Data
Wenkai Fang (Zhejiang University), Dacheng Tao (Nanyang Technological University)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes the SeRL framework, which achieves unsupervised reinforcement learning through LLM self-generated instructions and self-evaluated rewards in data-scarce situations;
SGN: Shifted Window-Based Hierarchical Variable Grouping for Multivariate Time Series Classification
Zenan Ying (University of Science and Technology of China), Wei Chen (University of Science and Technology of China)
CodeClassificationConvolutional Neural NetworkTransformerTime Series
π― What it does: This paper proposes SwinGroupNet (SGN) for multivariate time series classification, achieving structured modeling of inter-variable and temporal dependencies through modules such as variable grouping embedding, cross-scale window mixing, and periodic window shifting and merging.
π― What it does: Designed and evaluated a shallow flow matching (SFM) mechanism for coarse-to-fine text-to-speech synthesis models, allowing the flow model to sample from the intermediate state beyond noise, enhancing generation stability and naturalness.
SHAP zero Explains Biological Sequence Models with Near-zero Marginal Cost for Future Queries
Darin Tsui (Georgia Institute of Technology), Amirali Aghazadeh (Georgia Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryBiomedical Data
π― What it does: Proposes the SHAP zero algorithm, which utilizes sparse Fourier folding techniques to provide Shapley explanations for black-box biological sequence models, and achieves near-zero incremental cost for subsequent queries by constructing a global Fourier sketch in one go;
Shape it Up! Restoring LLM Safety during Finetuning
ShengYun Peng (Georgia Tech), Duen Horng Chau (Georgia Tech)
CodeSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: A dynamic safety shaping framework (DSS) is proposed, utilizing the token-level safety scoring STAR from the guardrail model to guide reinforcement learning on safe content during LLM fine-tuning, significantly enhancing safety while maintaining capabilities.
π― What it does: A self-supervised framework called ShapeEmbed is proposed, which encodes two-dimensional closed contours using a distance matrix and learns shape descriptors that are invariant to geometric transformations (translation, rotation, scaling, reflection, point indexing).