π― What it does: Proposes the MF-GIA framework, which leverages pre-trained graph neural networks to achieve cross-domain, modality-agnostic few-shot context learning, enabling parameter-free updates for node/edge classification on unseen domains.
π― What it does: Designed and implemented a completely unsupervised CT image pathology detection framework called Screener, which estimates pathological abnormalities using feature maps obtained through dense self-supervised learning and obscured-invariant conditional variables.
MoDr: Mixture-of-Depth-Recurrent Transformers for Test-Time Reasoning
Xiaojing Zhang (DataCanvas), Zhanxing Zhu (University of Southampton)
CodeComputational EfficiencyRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose MoDr Transformer, which dynamically selects the most suitable branch for the next step generation by introducing multi-branch LoRA and hard gate routing in Huginn's recurrent module, achieving multi-path 'deep reasoning',
π― What it does: Propose the Mixture-of-Length (MoL) framework, which adaptively controls the answer length based on question difficulty to achieve efficient and accurate context-based question answering.
MolecularIQ: Characterizing Chemical Reasoning Capabilities Through Symbolic Verification on Molecular Graphs
Christoph Bartmann (Johannes Kepler University Linz), Sohvi Luukkonen (Johannes Kepler University Linz)
CodeDrug DiscoveryLarge Language ModelMixture of ExpertsGraphBiomedical DataBenchmark
π― What it does: Designed and released a fully symbolically verifiable molecular structure reasoning benchmark called MOLECULARIQ, covering three task categories: counting, indexing, and constraint generation, while conducting fine-grained evaluation of the molecular reasoning capabilities of multiple LLMs.
MolLangBench: A Comprehensive Benchmark for Language-Prompted Molecular Structure Recognition, Editing, and Generation
Feiyang Cai (Clemson University), Feng Luo (University of Delaware)
CodeRecognitionGenerationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityGraphBenchmark
π― What it does: Create the MolLangBench benchmark to evaluate models' cross-modal capabilities in molecular structure recognition, molecular editing under language prompts, and generation tasks.
Samar Fares (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
CodeGenerationDiffusion modelImageText
π― What it does: Designed and implemented a trackable watermark framework (MOLM) for text-to-image diffusion models, which achieves watermark embedding and extraction by dynamically routing lightweight low-rank adapters (LoRA) with key-based parameter perturbation.
MoM: Linear Sequence Modeling with Mixture-of-Memories
Jusen Du (Tsinghua University), Yu Cheng (Chinese University of Hong Kong)
CodeText
π― What it does: Propose the Mixture-of-Memories (MoM) architecture, which uses multiple independent memory states and assigns inputs to specific memories via a router, thereby enhancing the memory capacity of linear sequence models and reducing memory interference.
MoMa: A Simple Modular Learning Framework for Material Property Prediction
Botian Wang (Tsinghua University), Hao Zhou (Tsinghua University)
CodeRepresentation LearningGraph Neural NetworkSupervised Fine-TuningGraphPhysics Related
π― What it does: Proposes the MoMa framework, which first trains specialized modules on multiple tasks and centralizes them in a Hub, then uses Adaptive Module Combination (AMC) to weight and fine-tune these modules for each downstream material property prediction task.
MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE
Geng Zhang (National University of Singapore), Yang You (National University of Singapore)
CodeComputational EfficiencyMixture of ExpertsTextBenchmark
π― What it does: This paper proposes a new Mixture-of-Novices-and-Experts (MoNE) expert pruning method, which replaces redundant experts with lightweight novices requiring no computational overhead using a small amount of calibration data, thereby significantly compressing the MoE model while maintaining high performance.
Monocular Normal Estimation via Shading Sequence Estimation
Zongrui Li (Nanyang Technological University), Song Bai (ByteDance)
CodeGenerationConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderImageVideo
π― What it does: Propose the RoSE method, which transforms monocular normal estimation into shadow sequence estimation, generating shadow sequences using a video generation model and solving for normals via OLS.
π― What it does: Construct a geospatial representation learning framework named MoRA centered on human mobility graphs, integrating multimodal data including POI, satellite imagery, and demographic statistics, and outputting a unified 128-dimensional regional embedding;
Mordal: Automated Pretrained Model Selection for Vision Language Models
Shiqi He (University of Michigan), Mosharaf Chowdhury (University of Michigan)
CodeRepresentation LearningHyperparameter SearchVision Language ModelMultimodality
π― What it does: Under aligned data for downstream tasks, automatically identify the optimal combination of pre-trained vision encoders and language models to construct the best Vision-Language Model (VLM).
More Than What Was Chosen: LLM-based Explainable Recommendation Beyond Noisy User Preferences
Chung Park (SK Telecom), Jaegul Choo (Korea Advanced Institute of Science & Technology)
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Propose the C-APO framework, which simultaneously models revealed preferences (RP) and coherent preferences (CP) in recommendation systems, and jointly optimizes recommendations and explanations through conflict-aware adaptive weights.
MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes
Yu Ying Chiu (University of Washington), Sydney Levine (New York University)
CodeLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Constructed and publicly released two novel moral reasoning benchmarksβMORE BENCH (1,000 scenarios, 23,018 human-written evaluation criteria) and MORE BENCHTHEORY (150 scenarios, covering five ethical frameworks)βand evaluated the reasoning quality of large models' Chain of Thought (CoT) and final answers using an LLM judge.
MoSA: Mosaic Shared Adaptation of Large Language Models
Xiequn Wang (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose a new parameter-efficient fine-tuning method called Mosaic Shared Adaptation (MoSA), which performs full-rank updates on pre-trained model weights through randomized shared scalars;
π― What it does: Proposed the MOSAIC framework for multi-agent personalized image generation, addressing the challenges of identity preservation and semantic consistency.
Motion-Aligned Word Embeddings for Text-to-Motion Generation
Ke Han (University of Trento), Nicu Sebe (University of Trento)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodality
π― What it does: Designed and implemented the MATE framework, aligning vocabulary with human motion semantics by fine-tuning the embedding layer of large language models, thereby improving word-level understanding and generation quality in text-to-motion tasks.
π― What it does: Propose MindHier, a hierarchical recursive fMRI-to-image reconstruction framework that progressively generates images using multi-level brain signal features;
David Anugraha (Stanford University), Genta Indra Winata (Capital One)
CodeReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
π― What it does: Proposed a multilingual, rating-scale-independent reward reasoning model called MR3, supporting evaluation and reasoning in 72 languages;
MRAD: Zero-Shot Anomaly Detection with Memory-Driven Retrieval
Chaoran Xu (Chinese Academy of Sciences), Zhengtao Zhang (Chinese Academy of Sciences)
CodeAnomaly DetectionPrompt EngineeringVision Language ModelImageBiomedical Data
π― What it does: Propose a memory retrieval-based zero-shot anomaly detection framework, MRAD, which directly utilizes frozen CLIP features and a two-layer feature-label memory bank for similarity retrieval, enabling both image-level classification and pixel-level segmentation.
π― What it does: Propose the MTVCraft framework, which directly quantizes raw 4D motion sequences (3D joint coordinates over time) into discrete 4D motion tokens and drives arbitrary character image animation using a 4D motion attention-based Diffusion Transformer (MV-DiT).
Multi-Action Self-Improvement For Neural Combinatorial Optimization
Laurin Luttmann (Leuphana University LΓΌneburg), Lin Xie (Brandenburg University of Technology)
CodeOptimizationTransformerGraphBenchmark
π― What it does: Propose the MACSIM framework, extending self-improvement to multi-agent joint actions by generating task assignments for all agents through a single forward pass and using set prediction loss to enhance sample efficiency and coordination.
π― What it does: Efficient offline multi-agent reinforcement learning (MARL) is achieved by first learning a multi-agent joint policy through flow matching, then distilling it into a single-step decentralized policy via behavioral cloning and Q-guidance.
Hongduan Tian (TMLR Group Hong Kong Baptist University), Bo Han (TMLR Group Hong Kong Baptist University)
CodeTransformerAgentic AITextChain-of-Thought
π― What it does: Propose a multi-agent debate framework called MAD-M2, which enhances the reasoning robustness of multi-agent debates by actively evaluating and masking erroneous memories generated in the previous round.
π― What it does: Propose a multi-condition adaptive composite selection method (MCCS) to achieve precise sample selection within target intervals and control the false discovery rate (FDR).
Multi-Domain Riemannian Graph Gluing for Building Graph Foundation Models
Li Sun (Beijing University of Posts and Telecommunications), Philip S. Yu (University of Illinois Chicago)
CodeDomain AdaptationRepresentation LearningGraph Neural NetworkMixture of ExpertsContrastive LearningGraph
π― What it does: Proposed a Riemannian geometry-based multi-domain graph pre-training framework called GRAPHGLUE, which integrates graph data from different domains into a smooth unified Riemannian manifold and enables cross-domain knowledge transfer through this manifold.
Multi-LLM Adaptive Conformal Inference for Reliable LLM Response
Kangjun Noh (Yonsei University), Kyungwoo Song (KAIST)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelText
π― What it does: Propose a synthetic reasoning method called MACI that combines multi-model fusion with group conditional calibration, aiming to ensure the authenticity of LLM responses in high-risk domains and achieve group conditional coverage guarantees.
Multi-modal Data Spectrum: Multi-modal Datasets are Multi-dimensional
Divyam Madaan (New York University), Sumit Chopra (New York University)
CodeExplainability and InterpretabilityRepresentation LearningData-Centric LearningLarge Language ModelImageTextMultimodalityBenchmark
π― What it does: Conduct large-scale experiments on 23 visual question answering benchmarks to systematically measure the extent to which multimodal models rely on unimodal (image, text) and interactive modalities;
Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models
Lujain Ibrahim (University of Oxford), Laura Weidinger (Google DeepMind)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Design and implement AnthroBench, a multi-turn dialogue automated evaluation framework to quantify anthropomorphic behaviors in large language models (LLMs).
Multimodal Aligned Semantic Knowledge for Unpaired Image-text Matching
Laiguo Yin (Shandong University), Lizhen Cui (Shandong University)
CodeRetrievalVision Language ModelContrastive LearningMultimodality
π― What it does: This paper proposes Multimodal Aligned Semantic Knowledge (MASK) for unpairedεΎζ matching, bridging words and visual prototypes through word vectors, and suppressing distribution variance via prototype consistency contrastive learning;
Multimodal Classification via Total Correlation Maximization
Feng Yu (Nanjing University of Science and Technology), Jianfeng Lu (Nanjing University of Science and Technology)
CodeClassificationMultimodality
π― What it does: Propose a hyperparameter-free TCMax loss that avoids modality competition and enhances cross-modal collaborative learning by maximizing the total correlation between multimodal features and labels
Multimodal Dataset Distillation Made Simple by Prototype-Guided Data Synthesis
Junhyeok Choi (Pohang University of Science and Technology), Minwoo Chae (Pohang University of Science and Technology)
CodeData SynthesisRetrievalKnowledge DistillationTransformerVision Language ModelDiffusion modelContrastive LearningMultimodality
π― What it does: Developed a learning-free multi-modal dataset distillation framework called PDS, which creates small-scale but information-rich multi-modal datasets by generating synthetic samples that conform to image-text prototypes using CLIP embeddings and unCLIP decoders.
π― What it does: This paper proposes a multimodal dataset distillation framework called PTM-ST based on phased teacher models and shortcut trajectories, which can extract knowledge from teacher models in different training stages and generate higher quality synthetic datasets.
Multimodal LLM-assisted Evolutionary Search for Programmatic Control Policies
Qinglong Hu (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision-Language-Action ModelMultimodality
π― What it does: Proposed and implemented a framework called MLES that combines multimodal large language models with evolutionary search to directly evolve executable and interpretable procedural control strategies through environmental interaction.
Multimodal Prompt Optimization: Why Not Leverage Multiple Modalities for MLLMs
Yumin Choi (KAIST), Sung Ju Hwang (KAIST)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringImageVideoTextMultimodality
π― What it does: Propose Multimodal Prompt Optimization (MPO) and design the MPO framework, which automatically jointly optimizes text and non-text prompts on MLLM to enhance the performance of multimodal tasks.
Multiverse Mechanica: A Testbed for Learning Game Mechanics via Counterfactual Worlds
Robert Ness, Lars Kunze
CodeExplainability and InterpretabilityData-Centric LearningDiffusion modelContrastive LearningWorld ModelImageVideoBenchmark
π― What it does: Developed the Multiverse Mechanica game testing platform, formalizing game mechanics as causal counterfactual inference, followed by causal consistency fine-tuning of the world model on this platform.
NAB: Neural Adaptive Binning for Sparse-View CT reconstruction
Wangduo Xie (KU Leuven), Matthew B. Blaschko (KU Leuven)
CodeBiomedical DataComputed Tomography
π― What it does: Proposes a self-supervised CT reconstruction method based on Neural Adaptive Binning (NAB), leveraging differentiable binning encoding to capture rectangular shape priors in industrial CT.
Naming to Learn: Class Incremental Learning for Vision-Language Model with Unlabeled Data
Qiwei Li (Peking University), Jiahuan Zhou (Peking University)
CodeClassificationRepresentation LearningVision Language ModelMultimodalityBenchmark
π― What it does: Propose a visual-language model incremental learning method called N2L, which addresses the problem of noisy pseudo-labels in unlabeled incremental learning by providing only class names and unannotated images for each incremental task.
Natural Language PDDL (NL-PDDL) for Open-world Goal-oriented Commonsense Regression Planning in Embodied AI
Xiaotian Liu (University of Toronto), Scott Sanner (University of Toronto)
CodeOptimizationRobotic IntelligenceLarge Language ModelVision Language ModelTextMultimodality
π― What it does: Propose the NL-PDDL framework, integrating natural language descriptions with classical PDDL, supporting goal-directed planning in open-world scenarios;
π― What it does: Proposed a data-free model fusion framework called FLEXMERGE, which can merge multi-task models at any size (including non-integer sizes).
NC-Bench and NCfold: A Benchmark and Closed-Loop Framework for RNA Non-Canonical Base-Pair Prediction
Heqin Zhu (University of Science and Technology of China), S Kevin Zhou
CodeTransformerBiomedical DataBenchmark
π― What it does: Constructed the NC-Bench benchmark dataset targeting RNA atypical base pairing, and proposed a closed-loop dual-branch framework NCfold for prediction
π― What it does: Propose NEF-NET V2, constructing an ECG panoramic reconstruction system with arbitrary length and arbitrary perspective, supporting multi-device and on-site calibration;
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: Investigate the negative pre-activation of smooth activation functions in large language models and explore their role in syntactic processing, with a focus on sparse Wasserstein neurons;
π― What it does: Proposed a continuous spatiotemporal motion graph (NeMo-map) that uses an implicit neural network to predict motion distributions at any position and time.
Nemotron-CC-Math: A 133 Billion-Token-Scale High Quality Math Pretraining Dataset
Rabeeh Karimi mahabadi, Bryan Catanzaro (NVIDIA)
CodeData SynthesisData-Centric LearningTransformerLarge Language ModelText
π― What it does: Constructed and publicly released a high-quality mathematical pre-training dataset named Nemotron-CC-Math with a scale of 133B tokens, and trained models to verify its effectiveness.
π― What it does: Propose the Neon method, which improves the quality of image generation models through negative extrapolation (reverse self-training degradation direction).
NePTune: A Neuro-Pythonic Framework for Tunable Compositional Reasoning on Vision-Language
Danial Kamali (Michigan State University), Parisa Kordjamshidi (Michigan State University)
CodeLarge Language ModelVision Language ModelMultimodality
π― What it does: This paper proposes NePTune, a neural symbolic framework that combines LLM-generated Python programs with VLM for visual-language compositional reasoning.
NetArena: Dynamic Benchmarks for AI Agents in Network Automation
Yajie Zhou (University of Maryland), Zaoxing Liu (University of Maryland)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIBenchmarkChain-of-Thought
π― What it does: Proposed the NETARENA framework for dynamically generating benchmark tasks for network automation and real-time verification of the correctness, safety, and latency performance of LLM agents in simulation environments.
π― What it does: This paper proposes a 3D mesh compression framework based on sparse implicit representation (SIR) and sparse neural compression (SNC), converting mesh surfaces into sparse SDFs retaining only the regions near the surface and compressing them into low-bitstreams via sparse convolutional autoencoders.
Neural Graduated Assignment for Maximum Common Edge Subgraphs
Chaolong Ying (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
CodeOptimizationGraph Neural NetworkGraphBiomedical Data
π― What it does: Propose a trainable temperature gradient assignment (Neural Graduated Assignment, NGA) framework based on neural networks for approximately solving the maximum common edge subgraph (MCES) problem.
Neural Latent Arbitrary Lagrangian-Eulerian Grids for Fluid-Solid Interaction
Shilong Tao (Peking University), Yunhuai Liu (Peking University)
CodeTransformerMeshTime SeriesSequentialPhysics Related
π― What it does: Designed and trained a deep learning framework called Fisale, based on multi-scale potential ALE grids and partitioned coupling modules, for predicting two-way fluid-structure interaction problems.
CodeAnomaly DetectionGraph Neural NetworkLarge Language ModelText
π― What it does: Construct an attributed graph using attention maps and activation information, perform graph neural network (GNN) information propagation on the internal computation trajectory of LLMs, and complete hallucination detection in generated text.
π― What it does: Proposed a decomposition-based reinforcement learning framework that generates Pareto fronts for multi-objective flexible job shop scheduling problems using a single neural network.
π― What it does: Accelerate polynomial non-negativity determination by predicting compact bases via Transformer, combined with repair mechanisms and iterative SDP solving;
Neural Synchrony Between Socially Interacting Language Models
Zhining Zhang (Peking University), Heng Ji (University Of Illinois Urbana Champaign)
CodeRepresentation LearningTransformerLarge Language ModelTextBenchmark
π― What it does: Studied the internal representation synchrony of large language models (LLMs) in social interactions, proposed and used the SyncR^2 metric to measure neural synchrony during LLM interactions;
π― What it does: Propose NeuralOS, an end-to-end model that generates operating system GUI frames through neural networks, capable of predicting screen images based on mouse and keyboard inputs;
Yanchen Wang (Columbia University), Jiajun Wu (Stanford University)
CodeExplainability and InterpretabilityBiomedical DataMagnetic Resonance Imaging
π― What it does: Proposes NEURONA, a neuro-symbolic framework that maps fMRI signals to interpretable conceptual representations and achieves precise decoding of multi-entity relationships in visual stimuli through symbolic execution.
Neuron-Level Analysis of Cultural Understanding in Large Language Models
Taisei Yamamoto (University of Tokyo), Hitomi Yanaka (University of Tokyo)
CodeExplainability and InterpretabilityTransformerTextBenchmark
π― What it does: Identify and validate neurons affecting cultural understanding in large language models through gradient attribution and control dataset screening methods, and explore their distribution and functions.
NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents
Tianshi Zheng (Hong Kong University of Science and Technology), Simon See (NVIDIA)
CodeTransformerLarge Language ModelAgentic AIBenchmarkPhysics Related
π― What it does: Proposes NEWTONBENCH, an interactive benchmark for evaluating the capability of large language models (LLMs) in discovering scientific laws.
π― What it does: Proposes the NewtonGen framework, achieving the generation of physically consistent and controllable dynamic images from text descriptions.
NExT-OMNI: Towards Any-to-Any Omnimodal Foundation Models with Discrete Flow Matching
Run Luo (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Tat-Seng Chua (National University of Singapore)
CodeGenerationRetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningFlow-based ModelAuto EncoderImageVideoTextMultimodalityAudio
π― What it does: Developed NExT-OMNIβa fully open-source omnimodal foundation model capable of cross-modal understanding, generation, and retrieval between any modalities;
NextQuill: Causal Preference Modeling for Enhancing LLM Personalization
Xiaoyan Zhao (Chinese University of Hong Kong), Tat-Seng Chua (National University of Singapore)
CodeRecommendation SystemExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the NextQuill framework, which personalizes and aligns large language models through causal preference modeling, identifying and aligning preference components driven by user history in model predictions and real responses.
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale
Chunrui Han (StepFun), Yibo Zhu (StepFun)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelFlow-based ModelAuto EncoderImageVideoTextMultimodalityBenchmark
π― What it does: Propose a fully autoregressive text-to-image generation model named NextStep-1, which concatenates text with continuous image tokens into a unified sequence and progressively predicts image patches using a lightweight flow matching head.
Shanwen Mao (Harbin Institute Of Technology), Jie Liu (State Key Laboratory Of Smart Farm Technologies And Systems)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the NuBitQ framework and OCP plugin to achieve layer-wise adaptive non-uniform quantization, and suppress block-level anomaly errors through multi-level compensation.
Noise-Adaptive Diffusion Sampling for Inverse Problems Without Task-Specific Tuning
Yingzhi Xia (Institute of High Performance Computing, Agency for Science, Technology and Research), ZAIWANG GU
CodeRestorationDiffusion modelImage
π― What it does: This paper proposes the HMC sampling method N-HMC and its adaptive variant NA-N-HMC, which perform backward diffusion in the noise space to solve inverse problems without requiring task-specific hyperparameters.
Noise-Aware Generalization: Robustness to In-Domain Noise and Out-of-Domain Generalization
Siqi Wang (Boston University), Bryan A. Plummer (Boston University)
CodeDomain AdaptationImage
π― What it does: Studied the generalization problem in the presence of label noise and multi-domain data (Noise-Aware Generalization, NAG), and proposed an algorithm called DL4ND that uses cross-domain comparison for noise detection and label correction.
π― What it does: Proposes a positive-unlabeled (PU) learning framework named NcPU, combining a noise-robust non-contrastive learning loss (NoiSNCL) and phantom label disambiguation (PLD), to learn more discriminative representations without requiring auxiliary negative samples or prior class proportion assumptions;
π― What it does: The study investigates how to balance memory usage and accuracy across dimensions such as weight precision, KV cache compression, generation length, and parallel sampling during inference under a fixed GPU memory budget, systematically evaluating the effectiveness of various compression and expansion strategies in inference tasks.
Not All Models Suit Expert Offloading: On Local Routing Consistency of Mixture-of-Expert Models
Jingcong Liang (Fudan University), zhongyu wei
CodeComputational EfficiencyLarge Language ModelMixture of ExpertsText
π― What it does: This paper investigates the local routing consistency of expert routing in Mixture-of-Experts (MoE) large language models during inference, and proposes two quantitative metrics: Segment Routing Best Performance (SRP) and Segment Cache Best Hit Rate (SCH).
Not Search, But Scan: Benchmarking MLLMs on Scan-Oriented Academic Paper Reasoning
Rongjin Li (Beijing University of Posts and Telecommunications), Haihong E (Beijing University of Posts and Telecommunications)
CodeTransformerLarge Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes ScholScan, a benchmark for academic paper reasoning oriented towards scanning, requiring models to fully scan papers without target prompts and identify errors.
Justin Chen (Salesforce AI Research), Chien-Sheng Wu (Salesforce AI Research)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Designed an algorithm called NuRL that self-generates hints during reinforcement learning (RL) training to 'push the boundaries' of large language model (LLM) reasoning, helping the model learn difficult problems that it could not successfully explore otherwise.
π― What it does: This paper proposes a data-agnostic continuous model fusion method called NUFILT, which can sequentially merge multiple fine-tuned models into a single high-performance model without accessing task data.
NurValues: Real-World Nursing Values Evaluation for Large Language Models in Clinical Context
Ben Yao (Hong Kong Polytechnic University), Jing Qin (Hong Kong Polytechnic University)
CodeLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkChain-of-Thought
π― What it does: Constructed the NurValues benchmark using nursing behavior cases collected from five months of longitudinal interviews in real hospitals, generating two levels of difficulty datasets to evaluate the alignment capability of LLMs with the five major nursing values.
NΓΌwa: Mending the Spatial Integrity Torn by VLM Token Pruning
Yihong Huang (Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)), Qi Tian (Hong Kong Polytechnic University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
π― What it does: Proposed the Nuwa two-stage visual token pruning framework, balancing the performance of vision-language models with efficient inference;
π― What it does: Investigated the robustness of LLM latent space monitors, proposed and evaluated obfuscated activations attacks capable of bypassing monitors;
π― What it does: Designed and implemented an online distillation and shape prior-based remote sensing image generation model (OF-Diff), achieving high-fidelity, controllable layout-to-image generation without relying on real image references, and further enhancing diversity and semantic consistency through DDPO.
Object-Centric Refinement for Enhanced Zero-Shot Segmentation
Srinivasa Rao Nandam (University of Surrey), Muhammad Awais (University of Surrey)
CodeSegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageBenchmark
π― What it does: Proposed OC-ZSS, which improves CLIP's patch representations by utilizing self-supervised guided object prompts and a two-stage object refinement attention module to achieve zero-shot semantic segmentation.
π― What it does: Propose a one-time, no-training pruning framework called OBS-Diff for large-scale text-to-image diffusion models, which can significantly reduce model computational and memory overhead while maintaining visual quality.
Aravind Venugopal, Jeff Schneider (Princeton University)
CodeReinforcement LearningFlow-based ModelTime SeriesPhysics Related
π― What it does: Through offline goal-conditioned reinforcement learning, the occupancy measure learned from the world model is converted into a global reward function, thereby addressing the credit assignment problem under sparse rewards.
OCR-Reasoning Benchmark: Unveiling the True Capabilities of MLLMs in Complex Text-Rich Image Reasoning
Mingxin Huang (South China University Of Technology), Lianwen Jin (Huawei Technologies Co Ltd)
CodeLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes the OCR-Reasoning benchmark to systematically evaluate the ability of multimodal large language models in text-rich image reasoning tasks.
CodeComputational EfficiencyLarge Language ModelReinforcement LearningText
π― What it does: Developed OCTAX, a CHIP-8 simulation platform based on JAX, providing a GPU-accelerated classic arcade game environment for reinforcement learning experiments.
π― What it does: Proposed the OD 3 framework, which performs data distillation on object detection datasets in a non-optimized manner, generating extremely small but efficient synthetic datasets.
π― What it does: Proposed a continuous spatiotemporal graph model called ODEBRAIN based on neural ordinary differential equations (Neural ODE), for dynamic brain network modeling of EEG signals and achieving epilepsy seizure detection and abnormal EEG classification.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextOrdinary Differential Equation
π― What it does: Proposed a unified activation scheduling framework based on ordinary differential equations (ODEs) and implemented a novel scheduling method called ODESTEER to enhance large language model (LLM) alignment effectiveness.
ODI-Bench: Can MLLMs Understand Immersive Omnidirectional Environments?
Liu Yang (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
CodeRecognitionVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed the ODI-Bench benchmark for panoramic image (ODI) understanding, containing 2000 high-quality panoramic images and over 4000 QA pairs, covering 10 fine-grained tasks, and benchmarked 20 multimodal large language models (MLLMs) under both closed-ended and open-ended evaluations;
π― What it does: Propose an offline safe reinforcement learning algorithm COX-Q, combining cost-constrained optimistic exploration and distributed value learning to maintain safety during data collection and deployment while improving sample efficiency.
π― What it does: Propose Q-Augmented Dual-Feature Fusion Decision Transformer (QDFFDT), integrating global sequence features and local Markovian features in offline reinforcement learning, and introducing Q-learning for value-guided decision-making.
OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!
Jingdi Lei (Nanyang Technological University), Soujanya Poria (Nanyang Technological University)
CodeSafty and PrivacyLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Evaluate the operational safety of large language models in specific agent tasks, propose a multilingual, multi-domain evaluation suite called OFFTOPICEVAL, and conduct large-scale benchmarking on 20 open-source/closed-source models.
CodeGenerationData SynthesisLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelVideoMultimodalityBenchmarkAudio
π― What it does: Designed and implemented the Omni-Detective data generation pipeline, two fine-grained description models (Audio-Captioner and Omni-Captioner), and proposed the Omni-Cloze benchmark for unified evaluation of multimodal fine-grained perception capabilities.
Omni-IML: Towards Unified Interpretable Image Manipulation Localization
Chenfan Qu (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeAnomaly DetectionExplainability and InterpretabilityLarge Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: Developed a general-purpose image manipulation localization model called Omni-IML and constructed a large-scale, structured interpretable dataset named Omni-273k.
Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences
Zhuoran Jin (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageVideoTextMultimodalityBenchmarkAudio
π― What it does: Propose the Omni-Reward framework, constructing Omni-RewardBench, Omni-RewardData, and Omni-RewardModel to achieve multi-modal free-form preference reward modeling.
π― What it does: Propose Omni-View, a unified model integrating 3D scene understanding and generation, enhancing multi-view image understanding through texture and geometry generation modules.