Conference on Neural Information Processing Systems Β· 1874 papers
Black-Box Forgetting
Yusuke Kuwana (Tokyo University of Science), Go Irie (Tokyo University of Science)
CodeOptimizationPrompt EngineeringVision Language ModelImage
π― What it does: A framework for selective forgetting (Black-Box Forgetting) on black-box pre-trained models is proposed, which allows the model to forget specified categories without changing the model parameters.
BLAST: Block-Level Adaptive Structured Matrices for Efficient Deep Neural Network Inference
Changwoo Lee (University of Michigan), Hun-Seok Kim (University of Michigan)
CodeCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelImageText
π― What it does: This paper proposes a Block-Level Adaptive Structured Matrix (BLAST) to replace the dense weights in the linear layers of large models, achieving inference acceleration.
BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models
Yibin Wang (Rutgers University), Hao Wang (Rutgers University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a Bayesian low-rank adaptation framework BLoB that jointly estimates the weight mean and covariance through backpropagation during the fine-tuning process of LoRA-parameterized LLMs.
π― What it does: This paper proposes a Bayesian learning framework based on dispersed block sparse priors (DivSBL), which can adaptively estimate block size and position, significantly improving the recovery accuracy of block sparse signals.
Block Transformer: Global-to-Local Language Modeling for Fast Inference
Namgyu Ho (KAIST), Se-Young Yun (KAIST)
CodeTransformerLarge Language ModelText
π― What it does: A block-based Transformer structure is designed, divided into block-level global attention and word-level local attention, to accelerate autoregressive language model inference.
π― What it does: This paper proposes a Bayesian Model Reduced Structured Pruning method (BMRS) that achieves threshold-free network structure pruning, further improving computational and energy efficiency.
π― What it does: A Boolean logic deep learning framework based on Boolean mutation is proposed, enabling training and inference within the Boolean domain.
BoNBoN Alignment for Large Language Models and the Sweetness of Best-of-n Sampling
Lin Gui (University of Chicago), Victor Veitch (University of Chicago)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
π― What it does: This paper analyzes the relationship between Best n-sampling (BoN) and alignment methods, proving that BoN is nearly optimal in the trade-off between win rate and KL divergence, and proposes the BoNBoN alignment method to approximate the BoN distribution without significantly increasing the discreteness.
CodeDomain AdaptationTransformerVision Language ModelImageMultimodality
π― What it does: This paper proposes BoostAdapter, which improves the adaptive capability of zero-training visual language models during testing by incorporating self-enhanced low-entropy samples into the cache.
Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing
Yadong Qu (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
CodeRecognitionTransformerImage
π― What it does: A semi-supervised scene text recognition framework based on 'watching + summarizing' is proposed, which significantly improves the recognition performance of artistic and distorted text through online generation strategies and character unidirectional alignment loss.
Boosting the Potential of Large Language Models with an Intelligent Information Assistant
Yujia Zhou (Tsinghua University), Zhicheng Dou (Renmin University of China)
CodeGenerationRetrievalReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: The ASSISTRAG framework is proposed, which combines a trainable intelligent information assistant with the main LLM to enhance the reasoning and accuracy of retrieval-augmented generation tasks.
Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning
Di Ming (Chongqing University of Technology), Xin Feng (Chongqing University of Technology)
CodeAdversarial AttackTransformerImage
π― What it does: An Adaptive Token Tuning (ATT) attack method is proposed to enhance the transferability of adversarial attacks on Vision Transformers (ViT).
Boosting Transferability and Discriminability for Time Series Domain Adaptation
Mingyang Liu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeDomain AdaptationKnowledge DistillationTime Series
π― What it does: Designed the ACON network for unsupervised domain adaptation in time series, combining multi-period frequency feature learning, time-frequency domain mutual learning, and time-frequency related subspace adversarial learning to enhance transferability and discriminability.
CodeClassificationRecognitionOptimizationTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: A transductive method named TransCLIP is proposed, which enhances zero/few-shot inference performance on visual language models by leveraging the structure of unlabeled data.
Boundary Decomposition for Nadir Objective Vector Estimation
Ruihao Zheng (Southern University of Science and Technology), Zhenkun Wang (Southern University of Science and Technology)
CodeOptimization
π― What it does: A boundary decomposition-based bi-level optimization method (BDNE) is proposed for accurately estimating the nadir objective vector of multi-objective optimization problems (MOPs).
π― What it does: This paper proposes the BiLAF framework, which achieves efficient active fine-tuning through one-shot sample selection under the pre-training-fine-tuning paradigm.
BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning
Jianming Pan (University of California), Jiang Bian (Microsoft Research Asia)
CodeOptimizationComputational EfficiencyFinance Related
π― What it does: A differentiable convex optimization framework BPQP is proposed, which can efficiently solve the gradients of the optimization layer in end-to-end learning.
π― What it does: This study investigates the practical utilization of stimulus reconstruction methods for brain signals, proposing the BrainBits metric to quantify this utilization and provide random baselines and reconstruction upper limits.
Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeClassificationMixture of ExpertsImage
π― What it does: A controllable long-tail learning framework PRL is proposed, which utilizes a hypernetwork to generate diverse expert models, and dynamically adjusts the head and tail class weights through a preference vector during testing, achieving dual adaptation to distribution shifts and user needs.
π― What it does: This paper proposes an AI-generated image detection method based on image patch shuffling and an end-to-end block-level classifier, aimed at eliminating the issue of insufficient cross-scene generalization caused by 'semantic artifacts'.
Breaking the False Sense of Security in Backdoor Defense through Re-Activation Attack
Mingli Zhu (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
CodeOptimizationAdversarial AttackImage
π― What it does: This study investigates the residual backdoors in post-training defense models and proposes backdoor reactivation attack schemes in white-box, black-box, and transfer scenarios.
BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO
Sebastian Dittert (Universitat Pompeu Fabra), Gianni De Fabritiis (Universitat Pompeu Fabra)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: The BricksRL platform has been built, integrating LEGO hardware, Pybricks, and TorchRL, enabling the training of reinforcement learning (RL) algorithms on real LEGO robots, and providing examples of various tasks and robot models.
Bridge the Modality and Capability Gaps in Vision-Language Model Selection
Chao Yi (Nanjing University), Han-Jia Ye (Nanjing University)
CodeClassificationRecognitionRetrievalTransformerVision Language ModelImageTextMultimodality
π― What it does: The research selects the most suitable visual language model from the VLM repository using only the textual information of the target dataset, proposing to improve selection accuracy by bridging the modality gap and capability gap.
CodeGenerationProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningGraph
π― What it does: A generative model based on Markov bridge, named Bridge-IF, is proposed for inverse protein folding (generating foldable protein sequences from a given backbone structure). It generates a deterministic prior sequence through a structure encoder, and then gradually refines this prior on the Markov bridge, ultimately obtaining high-quality sequences that meet structural constraints.
π― What it does: The FMCSC framework is proposed in federated multi-view clustering to address the issues of client differences and view differences caused by the mixture of single-view and multi-view clients.
Bridging Multicalibration and Out-of-distribution Generalization Beyond Covariate Shift
Jiayun Wu (Tsinghua University), Steven Wu
CodeDomain AdaptationOptimizationTabularBenchmark
π― What it does: A new model-free optimization framework is proposed, which extends multicalibration with a joint grouping function and achieves robust learning for out-of-distribution (OOD) generalization. The MC-PseudoLabel algorithm is introduced, aligning multicalibration with invariance.
π― What it does: This paper proposes a graph-theory-based framework that jointly addresses the problems of OOD generalization and detection, utilizing the decomposition of the graph's adjacency matrix to obtain a closed-form representation.
π― What it does: Under unsupervised local context interaction, artificial intelligence agents self-play through an information theory framework, co-evolving a shared vocabulary that is both semantic and pragmatic, and evaluating their human similarity using multi-objective loss.
π― What it does: This study investigates the differences between linear attention and Softmax attention, proposing two attributes for improvement: injectivity and local modeling.
Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models
Ling Yang (Peking University), Bin CUI
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: The Buffer of Thoughts (BoT) framework is proposed, which establishes a meta-buffer to store high-level thinking templates, supporting large language models (LLMs) in efficient, accurate, and robust reasoning in multi-task scenarios.
π― What it does: The study investigates the use of structured linear layers (low-rank, block shuffling, block dense) in Transformer large language models as a replacement for traditional FFN, and validates its feasibility under custom training scales.
Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment
Teng Xiao (Pennsylvania State University), Vasant G Honavar (Pennsylvania State University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes a language model alignment method called Cal-DPO, which calibrates the implicit rewards in contrastive preference learning to maintain the same scale as the true rewards, thereby avoiding the issue of rewards decreasing during training that is common in traditional methods.
CALANet: Cheap All-Layer Aggregation for Human Activity Recognition
Jaegyun Park (Chung Ang University), Jaesung Lee (Chung Ang University)
CodeRecognitionConvolutional Neural NetworkTime Series
π― What it does: A lightweight convolutional network called CALANet is proposed, which utilizes feature aggregation from all layers to enhance human action recognition accuracy.
Yiyang Zhou (University of North Carolina at Chapel Hill), Huaxiu Yao (National University of Singapore)
CodeGenerationOptimizationTransformerReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: Calibrating self-reward for large-scale visual language models (LVLM) to make the model pay more attention to image information when generating text, thereby reducing hallucinations and enhancing cross-modal alignment.
Can Graph Learning Improve Planning in LLM-based Agents?
Xixi Wu (Fudan University), Dongsheng Li (Microsoft Research Asia)
CodeGraph Neural NetworkTransformerLarge Language ModelTextGraph
π― What it does: This paper proposes modeling task planning as a graph decision problem and combines graph neural networks with large language models (LLM) to achieve subtask retrieval, significantly improving planning effectiveness.
Can Graph Neural Networks Expose Training Data Properties? An Efficient Risk Assessment Approach
Hanyang Yuan (Zhejiang University), Yang Yang (Finvolution Group)
CodeAdversarial AttackGraph Neural NetworkGraph
π― What it does: This paper studies attribute inference attacks based on graph neural networks and proposes an efficient method to generate approximate shadow models using model approximation.
Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeLarge Language ModelPrompt EngineeringContrastive LearningTextChain-of-Thought
π― What it does: This paper proposes a method for robustness assessment and enhancement against noisy reasoning (invalid or erroneous reasoning steps) in chain-of-thought prompts, constructing a dedicated NoRa dataset and designing a CD-CoT comparative denoising framework;
Can LLMs Learn by Teaching for Better Reasoning? A Preliminary Study
Xuefei Ning (Tsinghua University), Yu Wang (Tsinghua University)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper explores whether large language models (LLMs) can enhance reasoning abilities through a method called 'Learning by Teaching' (LbT). It proposes three teacher-student interaction-based improvement methods, M1, M2, and M3, and conducts experiments on mathematical reasoning, code synthesis, and text classification tasks.
Pei Yang (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeAdversarial AttackDiffusion modelImageAudio
π― What it does: This paper studies the vulnerability of content-agnostic digital watermarks and proposes a steganalysis attack based on simple averaging, which can effectively remove or forge watermarks in both gray-box and black-box scenarios.
Farzaneh Taleb (KTH Royal Institute of Technology), Danica Kragic (KTH Royal Institute of Technology)
CodeTransformerTabular
π― What it does: This study validates the potential of unsupervised models in predicting olfactory attributes by aligning the representations encoded by the pre-trained chemical structure Transformer (MoLFormer) with human olfactory perception.
π― What it does: A deepfake detection method called ProDet is proposed, which organizes three types of samples: real, blendfake, and deepfake into the latent space in a progressive manner, and enhances detection performance through joint training.
Cascade Speculative Drafting for Even Faster LLM Inference
Ziyi Chen (University of Illinois), Jie Huang (University of Illinois)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A new inference acceleration algorithm is proposed - Cascade Speculative Drafting (CS Drafting), which achieves faster LLM inference through two levels of cascading (vertical cascading and horizontal cascading).
π― What it does: Designed and implemented the CAT model, which automatically completes the segmentation of multiple abdominal organs and tumors using 3D anatomical volume visual cues and medical text prompts.
Causal Discovery from Event Sequences by Local Cause-Effect Attribution
Joscha CΓΌppers (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
CodeTime SeriesSequentialFinance Related
π― What it does: This study investigates how to discover causal structures from event sequences, proposing a causal model based on one-to-one event matching and introducing the CASCADE algorithm.
Causal discovery with endogenous context variables
Wiebke GΓΌnther (German Aerospace Center), Jakob Runge (German Aerospace Center)
CodeTabular
π― What it does: This paper proposes a constraint-based causal discovery method suitable for endogenous contextual variables (PC-AC), which can simultaneously recover the causal graph for each context and its joint graph.
Causal language modeling can elicit search and reasoning capabilities on logic puzzles
Kulin Shah (University of Texas at Austin), Rina Panigrahy (Google Research)
CodeTransformerLarge Language ModelText
π― What it does: This paper investigates whether causal language models (Transformers) can solve logical puzzles (such as Sudoku and chess puzzles) by training solely on the next-word prediction task, and explores the impact of training order on the model's reasoning ability.
π― What it does: A CausalDiff framework based on causal diffusion models has been developed, utilizing diffusion models for conditional image generation, and separating causal features associated with labels from non-causal features during the generation process, thereby achieving robust classification in the face of unseen adversarial attacks.
CodeOptimizationComputational EfficiencyNeural Architecture SearchTransformerReinforcement LearningTime Series
π― What it does: The CE-NAS framework is proposed, which utilizes reinforcement learning to dynamically allocate GPU resources, prioritizing low-carbon one/shot NAS during high/low carbon periods and high-energy-consuming Vanilla NAS, thereby significantly reducing carbon emissions while ensuring search quality.
π― What it does: A basic model for single-cell transcriptomics, scCello, utilizing cell ontology knowledge is proposed, and unsupervised pre-training of gene expression is achieved through a triple loss function.
π― What it does: A facial synthesis method based on diffusion models, CemiFace, is proposed, which generates semi-hard samples with moderate similarity to the identity center using a similarity control factor, aimed at training more discriminative facial recognition models.
Certified Adversarial Robustness via Randomized $\alpha$-Smoothing for Regression Models
Aref Miri Rekavandi (University of Melbourne), Benjamin I. P. Rubinstein
CodeOptimizationAdversarial AttackTabular
π― What it does: A randomized smoothing method based on Ξ±-trimmed smoothing is proposed, providing a probabilistic robustness proof against βp attacks for regression models and achieving certification in a black-box setting.
Certified Machine Unlearning via Noisy Stochastic Gradient Descent
Eli Chien (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
CodeOptimizationSafty and PrivacyImage
π― What it does: This paper proposes the use of Projected Noise Stochastic Gradient Descent (PNSGD) to achieve the unlearning functionality of machine learning models, providing an approximate no-learning guarantee under convex assumptions.
π― What it does: This paper is the first to apply randomized smoothing methods to deep equilibrium models (DEQ) and proposes Serialized Randomized Smoothing (SRS) to significantly accelerate the provable robustness evaluation of DEQ.
Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs
Xuan Zhang (Singapore Management University), Min Lin (Sea AI Lab)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: When performing chain reasoning with large language models, we first use Tree-of-Thought (ToT) search to obtain multiple lines of thought, and then pair the 'preferred' and 'deteriorated' thoughts generated at each step of the search tree, utilizing this preference data to fine-tune the model during the training phase (Chain of Preference Optimization, CPO).
π― What it does: A skeleton multi-entity action recognition framework based on Convex Hull Adaptive Translation (CHASE) is proposed, utilizing sample-adaptive coordinate origin relocation to alleviate distribution differences between different entities.
CodeObject DetectionRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodalityPoint Cloud
π― What it does: This study investigates 3D multimodal large language models, proposing object identifiers and object-level representations, unifying scene-language tasks into a question-answer format.
Cherry on Top: Parameter Heterogeneity and Quantization in Large Language Models
Wanyun Cui (Shanghai University of Finance and Economics), Qianle Wang (Shanghai University of Finance and Economics)
CodeOptimizationTransformerLarge Language ModelTextBenchmark
π― What it does: This study investigates the heterogeneity of parameters in large language models (LLMs) and proposes a mixed-precision quantization method called CherryQ, which retains a small number of key parameters (the 'cherry' parameters) at high precision while quantizing the remaining parameters to low bits.
ChronoEpilogi: Scalable Time Series Selection with Multiple Solutions
Etienne Vareille (CY Cergy Paris Universite), Vassilis Christophides
CodeTime Series
π― What it does: A scalable time series variable selection method called ChronoEpilogi is proposed, which can simultaneously identify all minimal predictive subsets (Markov Boundary) and provide an equivalent class representation of interchangeable variables.
CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models
Saurav Jha (University of New South Wales), Lina Yao (CSIRO Data61)
CodeClassificationKnowledge DistillationRepresentation LearningTransformerMixture of ExpertsVision Language ModelImageMultimodality
π― What it does: This paper proposes a CLIP-based probabilistic fine-tuning framework called CLAP for continual learning, which can learn task-specific distributions in a multi-task flow while retaining pre-trained knowledge.
π― What it does: This study investigates the issue of class distribution shift caused by unknown attributes in zero-shot learning and proposes a robust representation learning algorithm based on synthetic environments and environment balancing.
Classifier Clustering and Feature Alignment for Federated Learning under Distributed Concept Drift
Junbao Chen (Beijing Institute of Technology), Lu Huang (Beijing Institute of Technology)
CodeFederated LearningContrastive LearningImage
π― What it does: This paper proposes the FedCCFA framework, which combines classifier clustering and feature alignment to address the issues of distributed concept drift and data heterogeneity in federated learning.
CLIP in Mirror: Disentangling text from visual images through reflection
Tiancheng Wang (Beihang University), Baochang Zhang
CodeClassificationRecognitionAdversarial AttackTransformerVision Language ModelContrastive LearningImageText
π― What it does: A zero-shot dual-stream framework called MirrorCLIP is proposed, which inputs the original image and its horizontally flipped version into the CLIP encoder. It generates separate masks to achieve decoupling of text and visual features by utilizing the invariance of visual features to horizontal flipping and the difference in invariance of text features.
CLIPCEIL: Domain Generalization through CLIP via Channel rEfinement and Image-text aLignment
Xi Yu (Brookhaven National Laboratory), Yuewei Lin (Brookhaven National Laboratory)
CodeDomain AdaptationTransformerVision Language ModelContrastive LearningImageTextBenchmark
π― What it does: A lightweight adapter is added to the CLIP pre-trained model, utilizing channel refinement and image-text alignment to achieve domain-invariant and class-related feature extraction, thereby enhancing cross-domain generalization performance.
π― What it does: Two new data selection metrics are proposed: negCLIPLoss (using the negative value of the CLIP training loss as a quality score) and NormSim (using the p-norm similarity of visual embeddings to compare the target task data) to enhance data selection effectiveness in multimodal contrastive learning.
Cloud Object Detector Adaptation by Integrating Different Source Knowledge
Shuaifeng Li (University of Electronic Science and Technology of China), Xiatian Zhu (University of Surrey)
CodeObject DetectionDomain AdaptationKnowledge DistillationConvolutional Neural NetworkPrompt EngineeringVision Language ModelContrastive LearningImage
π― What it does: Proposes the Cloud Object Detector Adaptation (CODA) task, which utilizes the detection results from large cloud models and integrates knowledge from the open-source vision-language model CLIP to train detectors in an unlabeled target domain.
π― What it does: A customizable and natural vehicle detector adversarial camouflage method (CNCA) based on a pre-trained diffusion model is designed, which can generate diverse camouflage textures through user text prompts while maintaining high attack performance.
Co-occurrence is not Factual Association in Language Models
Xiao Zhang (Tsinghua University), Ji Wu (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The study investigates that language models tend to learn word co-occurrence statistics rather than real factual associations during the fine-tuning process, and proposes improvements to enhance the generalizability of factual knowledge.
Konstantinos P. Panousis (University of Athens), Diego Marcos (Inria)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkVision Language ModelImage
π― What it does: A Coarse-Fine Concept Bottleneck Model (CF-CBM) is proposed, which enhances interpretability and classification performance in visual tasks by simultaneously conducting concept discovery on the entire image and image patches.
π― What it does: This paper models collaborative learning as a bi-level optimization problem and proposes the COBO algorithm to achieve alternating optimization of client selection and model training.
Coevolving with the Other You: Fine-Tuning LLM with Sequential Cooperative Multi-Agent Reinforcement Learning
Hao Ma (School of Artificial Intelligence, University of Chinese Academy of Sciences), Min Chen (Institute of Automation, Chinese Academy of Sciences)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextSequential
π― What it does: A collaborative multi-agent reinforcement learning-based LLM fine-tuning framework called CORY is proposed, utilizing two agents, a pioneer and an observer, for knowledge transfer and role exchange to achieve co-evolution.
Abhinav Joshi (Indian Institute of Technology Kanpur), Ashutosh Modi (Indian Institute of Technology Kanpur)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes the COLD framework, which constructs a closed causal graph using script-based knowledge of daily activities and automatically generates nearly ten million causal query triples through this graph to systematically evaluate the causal reasoning capabilities of large language models.
ColJailBreak: Collaborative Generation and Editing for Jailbreaking Text-to-Image Deep Generation
Yizhuo Ma (Xi'an Jiaotong University), Qing Guo (Agency for Science Technology and Research)
CodeGenerationAdversarial AttackLarge Language ModelPrompt EngineeringDiffusion modelImageText
π― What it does: This paper proposes a collaborative generation and editing framework called ColJailBreak, which first generates safe images using secure natural language prompts, and then injects unsafe content through local editing to bypass the safety filters of commercial text-to-image models.
π― What it does: A Collaborative Neural Framework (CNF) is proposed to enhance the performance of neural vehicle routing methods on clean and adversarial instances through multi-model collaborative adversarial training.
Collaborative Cognitive Diagnosis with Disentangled Representation Learning for Learner Modeling
Weibo Gao (University of Science and Technology of China), Zheng Zhang (University of Science and Technology of China)
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningTabular
π― What it does: This paper proposes a collaborative cognitive diagnosis model named Coral, which achieves interpretable diagnosis of learners' knowledge abilities by jointly learning intrinsic and collaborative information.
CoLoR-Filter: Conditional Loss Reduction Filtering for Targeted Language Model Pre-training
David Brandfonbrener (Kempner Institute at Harvard University), Sham M. Kakade
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes CoLoR-Filter, an offline data selection method based on a two-model comparison, aimed at targeted pre-training of language models.
Color-Oriented Redundancy Reduction in Dataset Distillation
Bowen Yuan (University of Queensland), Zi Huang (University of Queensland)
CodeData SynthesisKnowledge DistillationImage
π― What it does: The AutoPalette framework is proposed to reduce color redundancy in synthesized images during dataset distillation through a palette network and color-guided initialization.
CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching
Dongzhi Jiang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeImage TranslationGenerationVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: This paper proposes an end-to-end fine-tuning strategy named CoMat, which enhances the alignment capability of text-to-image diffusion models by utilizing a concept matching mechanism from image to text.
Combining Observational Data and Language for Species Range Estimation
Max Hamilton (UMass Amherst), Subhransu Maji (UMass Amherst)
CodeLarge Language ModelText
π― What it does: Combining millions of citizen science observation data with Wikipedia text, the LE-SINR model is trained to achieve zero-shot prediction of species distribution ranges and improve accuracy in few-shot scenarios.
π― What it does: The CoMERA framework is proposed, which implements a training method based on rank-adaptive tensor compression, significantly reducing memory usage while achieving speedup in training on GPUs.
π― What it does: This paper studies the theoretical properties of node2vec (and its DeepWalk variant) under the (degree-corrected) stochastic block model and proves that the learned embeddings can be used for weakly consistent community detection via k-means.
Compact Proofs of Model Performance via Mechanistic Interpretability
Jason Gross (Massachusetts Institute of Technology), Lawrence Chan
CodeExplainability and InterpretabilityTransformerTabular
π― What it does: Using mechanism interpretation to perform reverse analysis on a single-layer single-head attention Transformer executing the Max-of-K task, various forms of formal proofs are constructed based on the obtained implementation details, and a lower bound on the model's accuracy is calculated.
Compositional Generalization Across Distributional Shifts with Sparse Tree Operations
Paul Soulos (Johns Hopkins University), Roland Fernandez (Microsoft Research)
CodeTransformerTextSequential
π― What it does: Proposes a Sparse Differentiable Tree Machine (sDTM) and uses Sparse Coordinate Trees (SCT) to represent tree structures, unifying neural and symbolic computation.
π― What it does: This paper proposes a composable PAC-Bayes framework that provides data-dependent consistency generalization bounds for heterogeneous layers and sub-models (such as graph neural networks and persistent layers like PersLay), and designs new regularization methods based on these bounds.
Compressing Large Language Models using Low Rank and Low Precision Decomposition
Rajarshi Saha (Stanford University), Mert Pilanci (Stanford University)
CodeCompressionTransformerLarge Language ModelText
π― What it does: The CALDERA algorithm is proposed, which performs low-precision low-rank decomposition (Q + LR) on LLM weight matrices and supports post-training quantization and low-rank adaptation.
Computerized Adaptive Testing via Collaborative Ranking
Zirui Liu (University of Science and Technology of China), Shijin Wang (iFLYTEK Co., Ltd)
CodeTabular
π― What it does: This paper proposes a Collaborative Computer Adaptive Testing (CCAT) framework, which utilizes collaborative students who have completed the entire question bank as a benchmark to improve question selection and ability estimation, thereby enhancing the consistency of student rankings.
Con4m: Context-aware Consistency Learning Framework for Segmented Time Series Classification
Junru Chen (Zhejiang University), Yang Yang (Zhejiang University)
CodeClassificationSegmentationTransformerTime Series
π― What it does: A segmented classification framework called Con4m is proposed for original multi-class time series with varying durations (MVD), utilizing contextual information to enhance segmentation classification performance and gradually harmonizing inconsistent boundary labels.
Concentrate Attention: Towards Domain-Generalizable Prompt Optimization for Language Models
Chengzhengxu Li (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
CodeDomain AdaptationOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: A prompt optimization objective based on attention concentration is proposed, with new loss and matching strategies designed for both soft prompts and hard prompts, thereby enhancing the generalization performance of language models in unknown domains.
π― What it does: A conditional controllable image fusion framework (CCF) based on a pre-trained denoising diffusion model is proposed, achieving training-free fusion through a conditional library and sampling adaptive condition selection.
Conditional Density Estimation with Histogram Trees
Lincen Yang (Leiden University), Matthijs van Leeuwen (Leiden University)
CodeTabular
π― What it does: A Conditional Density Tree (CDTree) is designed and implemented, which partitions the feature space using decision trees. Each leaf approximates the conditional density using histograms, and the tree structure and leaf histogram parameters are learned uniformly through the Minimum Description Length (MDL) principle.
Confidence Calibration of Classifiers with Many Classes
Adrien Le Coz (IRT SystemX), Faouzi Adjed (IRT SystemX)
CodeClassificationSupervised Fine-TuningImageText
π― What it does: A method is proposed to transform the confidence calibration problem of multi-class classifiers into a single binary classification problemβTop-versus-All (TvA), and improvements to existing calibration methods are made within this framework;
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: This study investigates the confidence modulation mechanisms within large-scale language models (LLMs) and identifies two types of neurons: entropy neurons, which adjust output entropy by writing into the null space of unembedding and utilizing LayerNorm; and token frequency neurons, which directly adjust the distance between the output distribution and the unigram distribution, thereby affecting the model's confidence. Their functionality is validated in the context of repeated sequences (induction).
Conformal Alignment: Knowing When to Trust Foundation Models with Guarantees
Yu Gui (University of Chicago), Zhimei Ren (University of Pennsylvania)
CodeGenerationData-Centric LearningSupervised Fine-TuningTextBiomedical Data
π― What it does: To address the alignment issue of large foundational model outputs, we propose the Conformal Alignment framework, which utilizes distribution-independent conformal prediction to select individual outputs, ensuring that the selected outputs are highly consistent with human evaluations under a given FDR target.
Conformal Prediction for Class-wise Coverage via Augmented Label Rank Calibration
Yuanjie Shi (Washington State University), Yan Yan (Washington State University)
CodeClassificationComputational EfficiencyImage
π― What it does: The RC3P algorithm is proposed, improving the traditional class-conditional CP, resulting in a significantly smaller prediction set while ensuring coverage for each category.
Alireza Javanmardi (Ludwig Maximilian University of Munich), Eyke HΓΌllermeier (Ludwig Maximilian University of Munich)
CodeClassificationImageText
π― What it does: This paper proposes a credible set prediction method based on conformal prediction to provide a set that includes the true distribution in classification tasks and quantifies randomness and knowledge uncertainty.
Baiting Chen (University of California Los Angeles), Lu Cheng (University of Illinois)
CodeRecurrent Neural NetworkTransformerTime SeriesFinance Related
π― What it does: This paper proposes the CT-SSF (Conformalized Time Series with Semantic Features) method, which constructs non-compliance scores in the semantic latent space of neural networks and dynamically adjusts weights to achieve confidence interval prediction for time series.
Conjugate Bayesian Two-step Change Point Detection for Hawkes Process
Zeyue Zhang (Renmin University of China), Feng Zhou (Renmin University of China)
CodeTime Series
π― What it does: A conjugate Bayesian two-step change point detection method CoBay-CPD is proposed to capture abrupt changes in the parameters of the Hawkes process over time;