NeurIPS 2025 Papers — Page 53
Conference on Neural Information Processing Systems · 5275 papers
When majority rules, minority loses: bias amplification of gradient descent
François Bachoc (University of Lille), Jean-Michel Loubes (University of Toulouse)
ClassificationOptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: This paper develops a formal framework to analyze majority-minority learning tasks in machine learning, revealing how standard training biases towards the majority group and produces stereotypical predictors that neglect minority features.
When Models Don’t Collapse: On the Consistency of Iterative MLE
Daniel Barzilai (Weizmann Institute of Science), Ohad Shamir (Weizmann Institute of Science)
Data SynthesisOptimization
🎯 What it does: This paper theoretically analyzes the maximum likelihood estimation (MLE) iterative training process in the context of gradually generating synthetic data and accumulating it into the original data, exploring the conditions for model collapse and strategies to avoid it.
When Models Know More Than They Can Explain: Quantifying Knowledge Transfer in Human-AI Collaboration
Quan Shi (Princeton University), Karthik R Narasimhan
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This study proposes the KITE framework and conducts a two-phase human-machine collaboration experiment with 118 participants to systematically evaluate the knowledge transfer capability of LLMs in coding and mathematical tasks.
When One Moment Isn't Enough: Multi-Moment Retrieval with Cross-Moment Interactions
Zhuo Cao (University of Queensland), Sen Wang (University of Queensland)
RetrievalRecurrent Neural NetworkContrastive LearningVideoText
🎯 What it does: This paper proposes the Multi-Moment Retrieval (MMR) task, constructs the QV-M² dataset annotated by humans, and designs the FlashMMR framework based on this.
When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding
Yan Shu (University of Trento), Nicu Sebe (University of Trento)
RecognitionTransformerLarge Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper addresses the issue of 'semantic illusion' in large multimodal models during scene text recognition and understanding, proposing a plug-and-play error correction framework that requires no training and can be applied only during the inference phase, and based on this, a new evaluation benchmark is created.
When Thinking Drifts: Evidential Grounding for Robust Video Reasoning
Mi Luo (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This study investigates the issue of visual thinking drift in chain-of-thought (CoT) during video reasoning tasks and proposes a Visual Evidence Reward (VER) mechanism. Through reinforcement learning (RL), it fine-tunes a multimodal large language model (MLLM) to better ground the model in video content when generating reasoning chains.
When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs
Xiaomin Li (Harvard University), Anurag Beniwal (Amazon)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study investigates the impact of explicit reasoning (Chain-of-Thought) on large language models (LLMs) in instruction-following tasks, finding that reasoning often reduces adherence accuracy and proposes four mitigation strategies.
When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization
Vivek Ramanujan (University of Washington), Ali Farhadi (University of Washington)
GenerationCompressionComputational EfficiencyTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the trade-off between compression and generation in a two-stage image generation framework, and improves the visual tokenizer by incorporating causal regularization into the encoder.
Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models
Donghoon Ahn (Korea Advanced Institute of Science and Technology), Seungryong Kim (Korea Advanced Institute of Science and Technology)
GenerationTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: This paper proposes attention perturbation at the granularity of a single attention head in the Diffusion Transformer (DiT), constructing the HeadHunter framework to iteratively select attention heads that meet the target, and introducing SoftPAG to adjust the perturbation intensity through linear interpolation, achieving fine-grained control over generation quality and visual style.
Where Does It Exist from the Low-Altitude: Spatial Aerial Video Grounding
Yang Zhan (Northwestern Polytechnical University), Yuan Yuan (Northwestern Polytechnical University)
Object DetectionObject TrackingTransformerVideoMultimodality
🎯 What it does: This paper proposes the task of Spatial Aerial Video Grounding (SAVG) and constructs a large-scale UAV-SVG dataset.
Where Graph Meets Heterogeneity: Multi-View Collaborative Graph Experts
Zhihao Wu (Zhejiang University), Haishuai Wang (Zhejiang University)
Graph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A multi-view collaborative graph expert framework MvCGE based on Mixture-of-Experts is proposed, which aggregates various graph experts at each layer using dynamic graph-aware routing, thus balancing consistency and complementarity in multi-view graph learning.
Which Algorithms Have Tight Generalization Bounds?
Michael Gastpar (École Polytechnique Fédérale de Lausanne), Thomas Weinberger (École Polytechnique Fédérale de Lausanne)
🎯 What it does: The study investigates whether there exists a tight generalization upper bound for learning algorithms under a given set of distributions, and provides necessary, sufficient, and necessary-sufficient conditions for estimability;
Which Data Attributes Stimulate Math and Code Reasoning? An Investigation via Influence Functions
Siqi Kou (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: By using influence functions for fine-grained attribution analysis of training data for large language models in mathematical and code reasoning, the impact of different data attributes on reasoning performance is revealed, and a data reweighting strategy is improved based on this.
Whitened Score Diffusion: A Structured Prior for Imaging Inverse Problems
Jeffrey Alido (Boston University), Lei Tian (Boston University)
RestorationGenerationDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: The Whitened Score Diffusion (WS-DM) model is proposed and trained to learn whitened scores, supporting arbitrary Gaussian forward noise, serving as a generative prior for image inverse problems.
Who Reasons in the Large Language Models?
Jie Shao (Nanjing University), Jianxin Wu (Nanjing University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates the source of reasoning capabilities in large language models, proposing that the output projection layer (o_proj) is a key module.
Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers
Xin Zhao (Institute of Information Engineering, Chinese Academy of Sciences), Yilong Chen (Institute of Information Engineering, Chinese Academy of Sciences)
GenerationAdversarial AttackTransformerMixture of ExpertsText
🎯 What it does: A backdoor attack framework called BadSwitch is proposed for Mixture-of-Experts (MoE) large models, embedding triggers into specific expert subsets of the model by utilizing expert routing preferences.
Who You Are Matters: Bridging Interests and Social Roles via LLM-Enhanced Logic Recommendation
Qing Yu (Wuhan University), Lixin Zou (Wuhan University)
Recommendation SystemKnowledge DistillationTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Utilizing multi-modal LLM to extract user roles and project topic tags, and using LLM to infer user-project behavior logic, a TagCF framework is constructed to enhance recommendation performance.
Whole-Body Conditioned Egocentric Video Prediction
Yutong Bai (University of California Berkeley), Jitendra Malik (University of California Berkeley)
GenerationPose EstimationTransformerDiffusion modelVideoSequential
🎯 What it does: Train an autoregressive conditional diffusion transformer model to predict first-person perspective videos using complete 3D human pose sequences.
Whose Instructions Count? Resolving Preference Bias in Instruction Fine-Tuning
Jiayu Zhang (Peking University), Xuan Zhang (Carnegie Mellon University)
TransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Addressing preference bias in instruction fine-tuning, robust instruction fine-tuning is achieved through Dynamic Cross-layer Preference Correction (DCPC).
Why 1 + 1 < 1 in Visual Token Pruning: Beyond Naive Integration via Multi-Objective Balanced Covering
Yangfu Li (East China Normal University), Yue Lu (East China Normal University)
CompressionOptimizationTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a training-independent visual token pruning method called MoB, which dynamically allocates retained tokens based on multi-objective balanced coverage theory to simultaneously meet the needs of visual retention and prompt alignment.
Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
Yiyou Sun (University of California), Dawn Song (University of California)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: A framework based on subsequence association is proposed to systematically track and explain the hallucination phenomenon of large language models.
Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training
Tony Bonnaire (PSL University), Marc Mezard
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This paper studies the impact of diffusion model training dynamics on memorization and generalization, revealing the existence of two time scales during the training process, which allows the model to achieve high-quality generation within the early stopping window without memorizing the training samples.
Why Do Some Language Models Fake Alignment While Others Don't?
Abhay Sheshadri (Anthropic), Fabien Roger (Anthropic)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: Evaluated the phenomenon of 'alignment masquerade' in 25 LLMs and found that only 5 models exhibited significant compliance gaps.
Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation
Sungmin Cha (New York University), Kyunghyun Cho (Genentech)
GenerationKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper explores the precision-recall trade-off of knowledge distillation in generative models through simulations of Gaussian mixture models and large-scale language model experiments.
Why Masking Diffusion Works: Condition on the Jump Schedule for Improved Discrete Diffusion
Alan Nawzad Amin, Andrew Gordon Wilson (New York University)
GenerationData SynthesisProtein Structure PredictionDiffusion modelImageText
🎯 What it does: A new discrete diffusion model design space called Schedule Conditioned Diffusion (SCUD) is proposed, which enhances model fitting by embedding the jump time distribution of the forward process into the backward process.
Why Playing Against Diverse and Challenging Opponents Speeds Up Coevolution: A Theoretical Analysis on Combinatorial Games
Alistair Benford (University of Birmingham), Per Kristian Lehre (University of Birmingham)
OptimizationGraph
🎯 What it does: The theoretical analysis of the running time of the competitive co-evolutionary evolutionary algorithm (UMDA) on the combinatorial game Reciprocal LeadingOnes (RLO) is conducted, providing an upper bound of O(n^2 log^3 n) and a lower bound of Ω(n^2 log n), while also verifying the performance improvement of the newly introduced random mutation operator.
Why Popular MOEAs are Popular: Proven Advantages in Approximating the Pareto Front
Mingfeng Li (Harbin Institute of Technology), Benjamin Doerr (École Polytechnique)
OptimizationBenchmark
🎯 What it does: A mathematical runtime analysis of the approximation performance of NSGA-II, NSGA-III, SMS-EMOA, and SPEA2 on the large Pareto front benchmark (LARGEFRONT 'ε) was conducted, proving that they require O(n² log n) fitness evaluations to obtain ε-approximate solutions in expectation, which is better than the simple MOEA GSEMO that only uses dominance checks (the latter requires exponential evaluations).
Wide-Horizon Thinking and Simulation-Based Evaluation for Real-World LLM Planning with Multifaceted Constraints
Dongjie Yang (Shanghai Jiao Tong University), hai zhao
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A multi-faceted planning framework MAoP and an Agent-based Travel-Sim simulation evaluation are proposed to enhance the performance of LLM in multi-constraint real-world planning (such as travel planning).
Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search
Yuichi Inoue (Sakana AI), Takuya Akiba (Sakana AI)
OptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes the Adaptive Branching Monte Carlo Tree Search (AB-MCTS) framework, which utilizes the diversity of LLMs and external feedback to dynamically decide whether to 'expand width' or 'dig deep' during inference, thereby improving the quality of task responses.
WildCAT3D: Appearance-Aware Multi-View Diffusion in the Wild
Morris Alper (Tel Aviv University), Tom Monnier (Meta AI)
GenerationData SynthesisDepth EstimationDiffusion modelImage
🎯 What it does: A multi-view diffusion model named WildCAT3D is proposed, which can learn from natural scene images captured from the web and generate a new panoramic view consistent with a single input view during inference, while also supporting appearance control and interpolation.
Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs
Hao Kang (Georgia Institute of Technology), Tsachy Weissman (Stanford University)
Recommendation SystemOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextTime SeriesBenchmarkFinance Related
🎯 What it does: Systematically analyze the delay-quality trade-off in real-time decision tasks of LLMs, design two benchmarks (HFTBench and StreetFighter), and propose an adaptive mixed-precision framework FPX;
WISA: World simulator assistant for physics-aware text-to-video generation
Jing Wang (Sun Yat-Sen University), Xiaodan Liang (University of Science and Technology Beijing)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelVideoTextPhysics Related
🎯 What it does: The WISA framework is proposed, which guides the generation of videos that comply with physical laws by breaking down physical principles into textual descriptions, qualitative categories, and quantitative attributes, and constructs an 80K physical video dataset WISA-80K.
Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting
Chenchen Tan (Monash University), Longxiang Gao (Qilu University of Technology)
TransformerLarge Language ModelText
🎯 What it does: Proposes the Attention-Shifting framework to achieve selective forgetting in LLMs, reducing access to sensitive knowledge without compromising overall performance.
With Limited Data for Multimodal Alignment, Let the STRUCTURE Guide You
Fabian Gröger (École Polytechnique Fédérale de Lausanne), Maria Brbic
ClassificationRetrievalContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a method for multimodal alignment using frozen unimodal pretrained models in low-sample scenarios.
WKV-sharing embraced random shuffle RWKV high-order modeling for pan-sharpening
Man Zhou (University of Science and Technology of China), Bo Huang (The University of Hong Kong)
Image TranslationRestorationTransformerImageMultimodality
🎯 What it does: This paper proposes a high-order model RSRWKV based on RWKV for multimodal image fusion (pan-sharpening). It extracts features from PAN and multispectral images through a dual-branch encoder, and then incorporates modules such as random shuffling, WKV sharing, and high-order gating in the spatial mixer and channel mixer, ultimately recovering high-resolution multispectral images in the decoder.
WMCopier: Forging Invisible Watermarks on Arbitrary Images
Ziping Dong (Zhejiang University), Kui Ren (Zhejiang University)
Data SynthesisAdversarial AttackDiffusion modelImage
🎯 What it does: Watermark forgery attacks using diffusion models in a black-box environment
Wonder Wins Ways: Curiosity-Driven Exploration through Multi-Agent Contextual Calibration
Yiyuan Pan (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningMultimodalityGraph
🎯 What it does: This paper proposes and implements CERMIC, a pluggable multi-agent contextual calibration module designed to enhance intrinsic curiosity exploration in sparse reward, partially observable, and non-communicative MARL scenarios.
Word-Level Emotional Expression Control in Zero-Shot Text-to-Speech Synthesis
tianrui wang, Jianwu Dang (Shenzhen Institute of Advanced Technology)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: A two-stage self-supervised training framework called WeSCon is designed to enable pre-trained zero-shot TTS models to achieve word-level emotion and speech rate control.
World Models as Reference Trajectories for Rapid Motor Adaptation
Carlos Stein Brito (NightCity Labs), Daniel C McNamee
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: A dual-mode framework called Reflexive World Models (RWM) is proposed, utilizing the latent trajectories predicted by the world model as a reference to achieve rapid dynamics adaptation without rewards.
World-aware Planning Narratives Enhance Large Vision-Language Model Planner
Junhao Shi (Fudan University), Xipeng Qiu (Fudan University)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodality
🎯 What it does: A World-Aware Planning Narrative Enhancement (WAP) framework is proposed, enabling large visual-language models (LVLMs) to achieve complex, multi-step, long-sequence task planning under closed-loop conditions using only visual observations and natural language instructions through multi-dimensional narratives and curriculum learning.
WorldMem: Long-term Consistent World Simulation with Memory
Zeqi Xiao (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: This paper proposes a long-term world simulation framework called WORLDMEM based on a memory bank, aimed at maintaining spatial and temporal consistency during the generation process.
WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception
Zhiheng Liu (University of Hong Kong), Linjie Yang (ByteDance Seed)
GenerationData SynthesisRobotic IntelligenceTransformerDiffusion modelAuto EncoderOptical FlowVideo
🎯 What it does: Proposes the WorldWeaver framework for long-term video generation using a combination of RGB and depth/optical flow perception conditions.
Wukong's 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models
Minghao Yin (University of Hong Kong), Kai Han (University of Hong Kong)
GenerationData SynthesisTransformerFlow-based ModelImageTextPoint Cloud
🎯 What it does: We propose a training-free high-fidelity texture 3D deformation framework called WUKONG, which can generate smooth and detail-rich 3D deformation sequences given only image or text prompts.
X-Field: A Physically Informed Representation for 3D X-ray Reconstruction
Feiran Wang, Yan Yan
RestorationOptimizationBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a three-dimensional representation called X-Field based on the penetration and attenuation characteristics of X-rays, using a concentric ellipsoid model to reconstruct X-ray views and CT volumes.
X-Mahalanobis: Transformer Feature Mixing for Reliable OOD Detection
Tong Wei (Southeast University), Min-Ling Zhang (Southeast University)
Anomaly DetectionTransformerImage
🎯 What it does: This paper proposes an OOD detection method called X-Maha based on multi-layer feature fusion of Transformer, utilizing the total variance of features at each layer as importance weights, and calculating the Mahalanobis distance as the OOD score after mixing features.
X-Scene: Large-Scale Driving Scene Generation with High Fidelity and Flexible Controllability
Yu Yang (Zhejiang University), Gim Hee Lee (National University of Singapore)
GenerationData SynthesisAutonomous DrivingGraph Neural NetworkLarge Language ModelDiffusion modelImageVideoRetrieval-Augmented Generation
🎯 What it does: The X-Scene framework is proposed to achieve the generation of large-scale 3D driving scenes, supporting multi-granularity control, high-fidelity geometric vision, and spatially consistent expansion.
xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories
Maurice Kraus (TU Darmstadt), Kristian Kersting (TU Darmstadt)
Recurrent Neural NetworkPrompt EngineeringTime SeriesBenchmark
🎯 What it does: A multi-view time series forecasting model called xLSTM-Mixer is proposed, which combines linear prediction with xLSTM hybrid.
XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation
Bowen Chen (ByteDance), Xinglong Wu (ByteDance)
GenerationData SynthesisTransformerDiffusion modelImageMultimodalityBenchmark
🎯 What it does: A multi-agent control framework XVerse based on DiT text stream modulation is proposed, which can achieve fine-grained, editable generation of multiple agents' identities, poses, styles, lighting, and other semantic attributes while maintaining the integrity of the image structure.
YEAST: Yet Another Sequential Test
Alexey Kurennoy (Meta), Ana Peleteiro Ramallo (Preply)
Tabular
🎯 What it does: A new continuous sequential testing method called YEAST is proposed for real-time monitoring of A/B experiments, making decisions immediately upon first reaching a threshold.
Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding
Yue Guan (Shanghai Jiao Tong University), Jingwen Leng (Shanghai Jiao Tong University)
TransformerLarge Language ModelText
🎯 What it does: Proposes the Yggdrasil system, which addresses the challenge of compatibility between dynamic speculative decoding and static graphs at compile time, achieving low-latency tree-based speculative decoding.
YOLOv12: Attention-Centric Real-Time Object Detectors
Yunjie Tian (University at Buffalo), David Doermann (University at Buffalo)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: In the YOLO series, an attention mechanism is incorporated to propose the YOLOv12 framework, achieving efficient real-time object detection through modules such as Area Attention and R-ELAN.
You Can Trust Your Clustering Model: A Parameter-free Self-Boosting Plug-in for Deep Clustering
Hanyang Li (Southeast University), Junhui Hou (City University of Hong Kong)
Representation LearningContrastive LearningImageBenchmark
🎯 What it does: A parameter-free self-enhancing plugin DCBoost is proposed, which can improve the global structure and clustering performance of existing deep clustering models.
You Only Communicate Once: One-shot Federated Low-Rank Adaptation of MLLM
Binqian Xu (National University of Singapore), Xiangbo Shu (Institute of High-Performance Computing)
Federated LearningLarge Language ModelMultimodality
🎯 What it does: A truly one-shot federated low-rank adaptation (OFL) scheme called YOCO is proposed, which guides the local LoRA training of a multimodal large language model (MLLM) using implicit global supervision, thus completing model adaptation in a single round of communication.
You Only Spectralize Once: Taking a Spectral Detour to Accelerate Graph Neural Network
Yi Li (University of Texas at Dallas), Bingzhe Li (University of Texas at Dallas)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes a GNN training scheme called YOSO that combines one-time spectralization and compressed sensing, performing a single learnable graph Fourier transform at the input layer and using compressed sensing to recover the complete embedding at the output layer, significantly reducing non-computational costs.
Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator
Beier Luo (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilizing unlabeled data for confidence calibration of post-trained LLMs
Zebra-Llama: Towards Extremely Efficient Hybrid Models
Mingyu Yang (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method to quickly construct a low-memory, high-efficiency hybrid model (Zebra-Llama) from existing large language models (such as Llama, Qwen), achieving almost no performance loss in knowledge transfer through fine initialization and hierarchical distillation.
ZEBRA: Towards Zero-Shot Cross-Subject Generalization for Universal Brain Visual Decoding
Haonan Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
RecognitionGenerationDomain AdaptationTransformerDiffusion modelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A zero-shot cross-subject brain visual decoding framework called ZEBRA is proposed, which reconstructs images from fMRI without the need for adaptation to new subjects.
ZeCO: Zero-Communication Overhead Sequence Parallelism for Linear Attention
Yuhong Chou (Hong Kong Polytechnic University), Zejun MA
TransformerLarge Language ModelText
🎯 What it does: This paper proposes the ZeCO sequence parallel method and the All-Scan communication primitive, achieving zero communication overhead for linear attention models across multiple devices, supporting near-linear scale training.
Zero-Shot Blind-Spot Image Denoising via Cross-Scale Non-Local Pixel Refilling
Qilong Guo (National University of Singapore), Hui Ji (National University of Singapore)
RestorationImage
🎯 What it does: A zero-shot blind spot denoising method based on cross-scale pixel completion is proposed, along with a theoretical risk analysis.
Zero-Shot Context Generalization in Reinforcement Learning from Few Training Contexts
James Chapman (University of California, Los Angeles), Guido Montufar
Reinforcement Learning
🎯 What it does: Proposes a Contextualized Bellman Equation (CEBE) and a Sample Enhancement method (CSE) to improve the zero-shot generalization ability of DRL under limited training contexts.
Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework
Ali Zafari (Rutgers University), Shirin Jalali (Rutgers University)
RestorationCompressionAuto EncoderImage
🎯 What it does: A zero-shot image denoising framework ZS-NCD is proposed, which utilizes a neural compression model for direct unsupervised training on a single noisy image, and then obtains denoised results through overlapping block averaging.
Zero-Shot Detection of LLM-Generated Text via Implicit Reward Model
Runheng Liu (Beijing Institute of Technology), Zhijing Wu (Beijing Institute of Technology)
Object DetectionReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a zero-shot LLM text detection method based on an implicit reward model (IRM), which constructs reward scores directly from publicly available instruction-tuned and base models without the need for additional preference data or training.
Zero-Shot Performance Prediction for Probabilistic Scaling Laws
Viktoria Schram (University of Melbourne), Trevor Cohn (University of Melbourne)
TransformerText
🎯 What it does: This paper proposes using a latent variable multi-output Gaussian process (Ma GP) for zero-shot prediction of learning curves in NLP tasks, and derives probabilistic scaling laws through Monte Carlo simulations, while combining active learning to reduce uncertainty.
Zero-shot protein stability prediction by inverse folding models: a free energy interpretation
Jes Frellsen, Wouter Boomsma
Protein Structure PredictionBiomedical Data
🎯 What it does: This paper theoretically derives the connection between the probability of the unsupervised inverse folding model and the thermodynamic free energy of proteins, explains the commonly used log-odds prediction method, and proposes improvements such as incorporating contributions from the unfolded state, using multi-sample folding conformations, and utilizing the generative model BioEmu. These improvements were subsequently validated on multiple experimental datasets.
Zero-Shot Trajectory Planning for Signal Temporal Logic Tasks
Ruijia Liu (Shanghai Jiao Tong University), Xiang Yin (Shanghai Jiao Tong University)
OptimizationRobotic IntelligenceDiffusion modelSequential
🎯 What it does: A zero-shot STL trajectory planning framework is designed to generate executable trajectories that satisfy complex STL tasks using task-agnostic trajectory data under unknown dynamics.
Zero-shot World Models via Search in Memory
Federico Malato (University of Eastern Finland), Ville Hautamaki
GenerationReinforcement LearningAuto EncoderWorld ModelImage
🎯 What it does: This paper proposes a zero-shot world model based on similarity search and VAE random representation, capable of predicting the transition dynamics of image environments and achieving long-term predictions without any learning training.
ZeroPatcher: Training-free Sampler for Video Inpainting and Editing
Shaoshu Yang, Ran He
RestorationGenerationDiffusion modelVideo
🎯 What it does: ZeroPatcher is a training-independent sampler that utilizes a pre-trained text-to-video diffusion model for video restoration and editing.
ZeroS: Zero‑Sum Linear Attention for Efficient Transformers
Jiecheng Lu (Georgia Institute of Technology), Shihao Yang (Georgia Institute of Technology)
Computational EfficiencyTransformerImageTextTime Series
🎯 What it does: This paper proposes Zero-Sum Linear Attention (ZeroS), which enhances the expressive power of Transformers while maintaining O(N) complexity by removing the zero-order constant term from softmax and re-weighting the remaining residuals to obtain zero-sum weights.
ZeroSep: Separate Anything in Audio with Zero Training
Chao Huang (University of Rochester), Chenliang Xu (Tencent America)
GenerationData SynthesisDiffusion modelAudio
🎯 What it does: A zero-shot audio source separation framework called ZeroSep is proposed, which utilizes a pre-trained text-guided audio diffusion model to achieve the separation of mixed audio.
Zeroth-Order Optimization Finds Flat Minima
Liang Zhang (ETH Zurich), Niao He (ETH Zurich)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the implicit regularization of zero-order optimization (two-point estimators) under convex smooth functions, proving that it tends towards flat local minima with minimal Hessian trace;
ZEUS: Zero-shot Embeddings for Unsupervised Separation of Tabular Data
Patryk Marszałek (Jagiellonian University), Marek Śmieja (Jagiellonian University)
Data-Centric LearningTransformerTabular
🎯 What it does: This paper presents ZEUS, a zero-shot Transformer model for clustering tabular data without fine-tuning.
ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding
LinshuangDiao, Dayong Ren (Nanjing University)
ClassificationSegmentationContrastive LearningPoint Cloud
🎯 What it does: A point cloud pre-training framework called ZigzagPointMamba is proposed, which combines a zigzag scanning path and a semantic semi-twin mask strategy to improve self-supervised learning.
zip2zip: Inference-Time Adaptive Tokenization via Online Compression
Saibo Geng (École Polytechnique Fédérale de Lausanne), Robert West (École Polytechnique Fédérale de Lausanne)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The zip2zip framework is proposed, which utilizes online LZW compression to dynamically expand the vocabulary during inference, achieving adaptive tokenization for large language models.
Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs
Xudong Li (Xiamen University), Rongrong Ji (Xiamen University)
OptimizationTransformerImage
🎯 What it does: This paper studies a multi-level preference optimization framework called CcDPO, aimed at enhancing multi-image understanding capabilities.
ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS
Weijie Wang (Zhejiang University), Bohan Zhuang (Zhejiang University)
CompressionTransformerGaussian SplattingPoint Cloud
🎯 What it does: Proposes the ZPressor module, which compresses multi-view inputs to enhance the scalability of the feed-forward 3D Gaussian Splatting model.