π― What it does: This paper discusses the theoretical and practical advantages of using complex parameterization in Structured State Space Models (SSM), demonstrating that complex SSMs outperform real-valued SSMs in expressiveness and learnability.
Provably and Practically Efficient Adversarial Imitation Learning with General Function Approximation
Tian Xu (Nanjing University), Yang Yu (Nanjing University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequentialBenchmark
π― What it does: An online adversarial imitation learning framework OPTβAIL is designed, utilizing online no-regret reward optimization and optimistic Bellman error minimization to learn rewards and Q-values, thereby achieving theoretically provable efficient imitation learning under general function approximation.
Provably Safe Neural Network Controllers via Differential Dynamic Logic
Samuel Teuber (Karlsruhe Institute of Technology), Andre Platzer (Carnegie Mellon University)
CodeAutonomous DrivingSafty and PrivacyTabularOrdinary Differential Equation
π― What it does: A method called VerSAILLE, which combines differential dynamic logic (dL) with neural network verification (NNV), is proposed to provide safety proofs for neural network-based control systems (NNCS) over an infinite time horizon. Additionally, a Mosaic framework is introduced, extending the linear constraint NNV tools to polynomial nonlinear and non-standard queries while maintaining completeness.
Haiming Wang (Sun Yat-sen University), Xiaodan Liang (Pengcheng Laboratory)
CodeLarge Language ModelSupervised Fine-TuningText
π― What it does: By recursively constructing verifiable proof sketches in a hierarchical manner and using sorry placeholders, POETRY advances layer by layer and ultimately achieves a complete proof.
Jacob M. Chen (Johns Hopkins University), Katherine A. Keith (Williams College)
CodeTransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsElectrocardiogram
π― What it does: This paper proposes a novel method that utilizes a zero-shot model for inferring two proxy variables from preprocessed text and applies it to proximal causal inference (proximal g-formula) to estimate causal effects in the absence of completely unobserved key confounding variables.
π― What it does: A content-aware image relocalization method named PruneRepaint is proposed, which combines hierarchical semantic-guided seam carving and an adaptive repainting module to achieve image retargeting at arbitrary ratios.
Pruning neural network models for gene regulatory dynamics using data and domain knowledge
Intekhab Hossain (Harvard T.H. Chan School of Public Health), John Quackenbush (Harvard T.H. Chan School of Public Health)
CodeOptimizationExplainability and InterpretabilityBiomedical DataOrdinary Differential Equation
π― What it does: A domain knowledge-based network pruning framework called DASH is proposed for inferring sparse neural network models of gene regulatory networks.
π― What it does: The research proposes a model inversion attack guided by pseudo-private data (PPDG-MI), which enhances the attack effectiveness by dynamically fine-tuning the generator's prior distribution.
PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation
Weiqin Yang (Zhejiang University), Can Wang (Zhejiang University)
CodeRecommendation SystemTabular
π― What it does: Proposed Pairwise Softmax Loss (PSL) to improve traditional Softmax Loss and achieve a more compact DCG approximation in recommendation systems.
π― What it does: A post-training quantization (PTQ) method called PTQ4DiT is proposed for Diffusion Transformers (DiTs), which can compress the model to 8-bit weights/activations while maintaining generation quality, and further supports 4-bit weight quantization.
Pure Message Passing Can Estimate Common Neighbor for Link Prediction
Kaiwen Dong (University of Notre Dame), Nitesh V Chawla
CodeGraph Neural NetworkGraph
π― What it does: This study investigates whether Message Passing Neural Networks (MPNN) can capture link structural features, and based on this, proposes the MPLP (Message Passing Link Predictor) model to estimate and predict potential edges in a graph.
π― What it does: This study investigates the out-of-distribution fluid dynamics modeling problem and proposes a framework called PURE (Prompt Evolution with Graph ODE) to enhance the model's predictive performance under parameter and temporal distribution shifts.
π― What it does: This paper proposes PUREGEN, a dynamic random preprocessing method that utilizes Energy-Based Models (EBM) and Diffusion Models (DDPM) for unsupervised purification of data during the training phase, thereby defending against backdoor and triggerless training-time data poisoning attacks.
PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression
Vladimir Malinovskii (Yandex), Peter RichtΓ‘rik
CodeCompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes PV-Tuning, a fine-tuning framework for large language models with extremely low bit quantization (1-2 bits), which can simultaneously update continuous parameters (such as scaling factors and codebooks) and discrete parameters (weight quantization codes), significantly improving the accuracy of the quantized model.
Q-Distribution guided Q-learning for offline reinforcement learning: Uncertainty penalized Q-value via consistency model
Jing Zhang (Hong Kong University of Science and Technology), Bingyi Jing
CodeReinforcement LearningTabularBenchmark
π― What it does: This paper proposes an offline reinforcement learning method called Q-Distribution Guided Q-Learning (QDQ), which estimates the uncertainty of the Q value distribution using a consistency model and applies a lazy penalty to out-of-distribution (OOD) actions to balance excessive pessimism and exploration.
Q-VLM: Post-training Quantization for Large Vision-Language Models
Changyuan Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeCompressionComputational EfficiencyTransformerMixture of ExpertsVision Language ModelMultimodality
π― What it does: A post-training quantization framework Q-VLM is proposed for efficient multimodal inference of large visual language models, reducing model size and inference speed.
π― What it does: This paper proposes a QGFN method that combines GFlowNet with the action value function Q, achieving a controllable greedy strategy during sampling through an adjustable mixing parameter p, thereby improving the generation rate of high-reward samples while maintaining diversity.
π― What it does: A directly trained hierarchical spiking Transformer called QKFormer is designed, and a linear complexity Q-K attention mechanism along with a cross-scale SPEDS embedding module is proposed, aiming to enhance the performance and energy efficiency of Spiking Neural Networks (SNNs);
π― What it does: A visual Mamba model called QuadMamba based on a learnable quadtree scanning mechanism has been designed for visual tasks such as image classification, object detection, instance segmentation, and semantic segmentation.
Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner
Valentyn Melnychuk (Ludwig Maximilian University of Munich), Mihaela van der Schaar (University of Cambridge)
CodeFlow-based ModelTabular
π― What it does: A novel orthogonal learner (AU-learner) is proposed to estimate the Markov boundary of the conditional treatment effect distribution (CDTE), thereby quantifying the Alitayi uncertainty of treatment effects.
Quantifying and Optimizing Global Faithfulness in Persona-driven Role-playing
Letian Peng (University of California, San Diego), Jingbo Shang (University of California, San Diego)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: Quantitatively assess the global fidelity in persona-driven role-playing (PRP) and propose an Active-Passive Constraint (APC) scoring and an APC-based Direct Preference Optimization (DPO) scheme.
CodeOptimizationRepresentation LearningTransformerLarge Language ModelText
π― What it does: Proposes and theorizes the mechanism of weak-to-strong generalization, proving that the error of a strong model using labels generated by a weak model is at least one 'misfit' order of magnitude lower than that of the weak model;
Quantitative Convergences of Lie Group Momentum Optimizers
Lingkai Kong (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
CodeOptimizationTransformerImage
π― What it does: This paper proposes two momentum-based Lie group optimizers: Lie Heavy-Ball and Lie NAG-SC, and provides quantitative convergence rates under L-smooth and local strong convexity conditions, verifying the acceleration effect of Lie NAG-SC, while applying it to the optimization of visual Transformers with orthogonal constraints.
Quantum algorithm for large-scale market equilibrium computation
Po-Wei Huang (National University of Singapore), Patrick Rebentrost (National University of Singapore)
CodeOptimizationComputational EfficiencyTabularFinance Related
π― What it does: The first quantum equilibrium computation algorithm for the Fisher market is proposed, and iterative updates are achieved through error-proportional response dynamics.
Philipp Schleich (University of Toronto), Alan Aspuru-Guzik
CodeOptimizationImage
π― What it does: A Quantum Deep Equilibrium Model (QDEQ) is proposed, achieving parameter optimization through the use of deep equilibrium techniques on quantum models;
David Huk (University of Warwick), Mark Steel (University of Warwick)
CodeHyperparameter SearchTabular
π― What it does: A high-dimensional joint density estimation method combining Bayesian prediction and Vine copula is proposedβQuasi-Bayesian Vine (QB-Vine).
π― What it does: A method for achieving query-efficient relevant clustering under unknown and noisy similarity is proposed, along with two pure exploration (fixed confidence / fixed budget) online learning frameworks, corresponding to the KC-FC and KC-FB algorithms.
π― What it does: The QUEST framework is proposed, which implements quaternion constraints and self-punishment for multimodal contrastive learning through a shared and unique information decoder.
π― What it does: An efficient QWO module is proposed for quickly calculating the causal graph GΟ under a given permutation in linear Gaussian acyclic models (LiGAM), significantly accelerating the permutation search process.
Yujie Zhao (University of California), Huazheng Wang (Oregon State University)
CodeReinforcement LearningTabularSequential
π― What it does: This paper proposes a risk-aware preference reinforcement learning algorithm named RA-PbRL, aimed at addressing the PbRL problem where only complete trajectory preference feedback is available. It demonstrates the theoretical feasibility and lower bounds for achieving nested and static quantile risk objectives.
π― What it does: We propose a ray-decoupled training framework called Rad-NeRF, which utilizes sub-NeRF integration and a soft gating module to allocate different sub-models along the ray dimension, achieving better lighting and geometric consistency through deep mutual learning.
π― What it does: A general retrieval-enhanced graph learning framework RAGRAPH is proposed, which improves the generalization ability of pre-trained graph neural networks on unseen graph data by utilizing external graph data through retrieval and information injection.
π― What it does: A novel parallel time solver RandNet-Parareal is proposed, which learns the error between coarse and fine solvers through a random neural network, thereby achieving rapid correction in the Parareal iteration.
Randomized Exploration in Cooperative Multi-Agent Reinforcement Learning
Hao-Lun Hsu (Duke University), Pan Xu (Duke University)
CodeReinforcement Learning
π― What it does: A unified framework is proposed for implementing Thompson Sampling-based stochastic exploration algorithmsβCoopTS-PHE and CoopTS-LMCβin parallel MDP environments;
Randomized Sparse Matrix Compression for Large-Scale Constrained Optimization in Cancer Radiotherapy
Shima Adeli (Sharif University of Technology), Masoud Zarepisheh (Memorial Sloan Kettering Cancer Center)
CodeOptimizationBiomedical Data
π― What it does: A dose impact matrix compression method based on random sparsification is proposed, enabling the solution of large-scale constrained nonlinear optimization problems in cancer radiotherapy within clinical time windows.
RanDumb: Random Representations Outperform Online Continually Learned Representations
Ameya Prabhu (University of Oxford), Puneet K. Dokania (IIIT Hyderabad)
CodeClassificationRepresentation LearningImage
π― What it does: This paper presents RanDumbβa baseline for online continual learning that uses only random Fourier feature projections and a nearest class mean (with Mahalanobis distance) classifier, aimed at evaluating the comparison between random representations and traditional online learning representations.
π― What it does: The RankUp framework is proposed, which transforms regression tasks into ranking problems, using an Auxiliary Ranking Classifier (ARC) trained together with semi-supervised classification methods like FixMatch, and introduces an optional Regression Distribution Alignment (RDA) to further improve the quality of pseudo-labels.
RashomonGB: Analyzing the Rashomon Effect and Mitigating Predictive Multiplicity in Gradient Boosting
Hsiang Hsu (JPMorganChase Global Technology Applied Research), Chun-Fu Chen (JPMorganChase AI Research)
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: Analyzes the Rashomon effect in gradient boosting and proposes the RashomonGB method, which efficiently constructs a model ensemble using the residual Rashomon set, further evaluating prediction diversity, interpretability, and fairness.
CodeClassificationObject DetectionSegmentationData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodality
π― What it does: This paper proposes the RAVL method, which analyzes fine-grained image regions to automatically discover and eliminate the false correlations learned by visual-language models during fine-tuning, thereby enhancing the model's robustness in zero-shot tasks.
π― What it does: This paper proposes RCDN, an algorithm capable of recovering failed viewpoints and maintaining high accuracy in multi-vehicle collaborative perception scenarios where cameras fail or noise is severe, through collaborative neural rendering.
π― What it does: A lightweight recurrent unit RTU is proposed and evaluated for real-time recursive learning (RTRL) in online reinforcement learning.
Real-Time Selection Under General Constraints via Predictive Inference
Yuyang Huo (Nankai University), Changliang Zou (Nankai University)
CodeOptimizationSupervised Fine-TuningTabular
π― What it does: A real-time online sample selection algorithm II-COS is proposed, which can select samples that meet the target response interval while satisfying individual cost constraints and interaction diversity constraints.
Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network
Chengyu Fang (Shenzhen International Graduate School Tsinghua University), Xiu Li (Shenzhen International Graduate School Tsinghua University)
CodeRestorationKnowledge DistillationImage
π― What it does: This paper proposes the Collaborative Unfolding Network (CORUN) and the consistency pseudo-label-based Mean-Teacher framework (Colabator) for achieving high-quality image dehazing in real-world scenarios.
RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models
Xinchen Zhang (Tsinghua University), Bin CUI
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
π― What it does: The RealCompo framework is proposed, which utilizes large language models to generate layouts and integrates pre-trained text-image diffusion models with spatially aware diffusion models (layouts, key points, segmentation, etc.) through a dynamic balancer without training, to achieve a balance between image realism and multi-object composability.
π― What it does: In autonomous driving, simultaneous multi-agent behavior prediction and planning are achieved by introducing a behavior topology (BeTop) based on winding theory to explicitly supervise the consistency of multi-agent future interactions. Furthermore, a collaborative prediction and planning Transformer network, BeTopNet, is designed to control uncertainty and conflicts.
Reasons and Solutions for the Decline in Model Performance after Editing
Xiusheng Huang (Chinese Academy of Sciences), Kang Liu (Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Through experiments from both data and model perspectives, the reasons for the performance decline after knowledge editing were systematically analyzed, and a new sequence editing method called D4S was proposed to mitigate performance loss.
REBEL: Reinforcement Learning via Regressing Relative Rewards
Zhaolin Gao (Cornell University), Wen Sun (Princeton University)
CodeReinforcement LearningText
π― What it does: This paper proposes a reinforcement learning algorithm named REBEL, which achieves policy optimization by regressing on 'relative rewards', simplifying the implementation difficulty of traditional RL schemes.
π― What it does: In unsupervised speech recognition, an iterative training framework called REBORN is proposed, which alternately trains a segmentation model and a phoneme prediction model.
Reciprocal Reward Influence Encourages Cooperation From Self-Interested Agents
John Luoyu Zhou, Jonathan Kao
CodeReinforcement Learning
π― What it does: This paper proposes a reinforcement learning agent named Reciprocator, which encourages self-interested agents to cooperate through intrinsic reciprocal rewards.
π― What it does: An algorithm is proposed to recover complete 3D meshes from incomplete 3D point clouds of non-worn and manipulated garments using the ISP model, an extended diffusion deformation prior, and a UV mapping network.
π― What it does: This paper proposes a personalized image generation method that requires no training, utilizing Anchored Classifier Guidance to guide Rectified Flow, which can maintain identity consistency based on reference images provided by users.
π― What it does: A fully convolutional autoencoder (SynCx) is proposed, utilizing complex weights and recursive phase updates to achieve unsupervised object detection.
Recursive PAC-Bayes: A Frequentist Approach to Sequential Prior Updates with No Information Loss
Yi-Shan Wu (University of South Denmark), Yevgeny Seldin (University of Copenhagen)
CodeClassificationOptimizationImage
π― What it does: A recursive PAC-Bayes framework is proposed, allowing for sequential prior updates without losing confidence information, and corresponding PAC-Bayes bounds are provided;
π― What it does: This paper proposes an online batch selection method called REDUCR, which implements robust data downsampling using class priority re-weighting, significantly improving the generalization performance of the worst class while reducing the amount of training data.
π― What it does: By incorporating reference feature guidance into the self-attention of the diffusion model UNet, controllable consistency generation that is independent of training has been achieved.
Reference Trustable Decoding: A Training-Free Augmentation Paradigm for Large Language Models
Shi Luohe (Wuhan University), hai zhao
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: A training-free, retrieval-based decoding method (Reference Trustable Decoding, RTD) is proposed, which constructs a reference dataset and adjusts the output distribution based on similarity during the decoding phase, allowing LLMs to quickly adapt to downstream tasks while keeping the original parameters unchanged.
Referring Human Pose and Mask Estimation In the Wild
Bo Miao (University of Western Australia), Ajmal Saeed Mian
CodeObject DetectionSegmentationPose EstimationTransformerVision Language ModelImageTextMultimodality
π― What it does: Proposed and implemented the 'R-HPM' (Referring Human Pose and Mask Estimation) task, which can accurately predict the joint location information and segmentation mask of a specified person through text or location prompts (scribble/point).
ReFIR: Grounding Large Restoration Models with Retrieval Augmentation
Hang Guo (Tsinghua University), Shu-Tao Xia (Tsinghua University)
CodeRestorationRetrievalDiffusion modelImage
π― What it does: A training-free retrieval-enhanced framework called ReFIR is proposed, which uses high-quality reference images obtained through retrieval to help large image restoration models avoid 'hallucinations' and generate more realistic details.
ReFT: Representation Finetuning for Language Models
Zhengxuan Wu (Stanford University), Christopher Potts (Stanford University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes and evaluates a parameter-efficient fine-tuning method through low-rank linear subspace intervention (ReFT) on the hidden representations of pre-trained language models, focusing on LoReFT and its efficient variant DiReFT.
π― What it does: This paper proposes a post-processing algorithm for a regression model that does not require inferring sensitive attributes, ensuring fair predictions under demographic balance constraints.
π― What it does: A new Regularized Adaptive Dual Equilibrium (RAMDA) algorithm is proposed for training deep neural networks with structures (such as sparsity), and an implementable subproblem approximation solver is designed for this algorithm.
Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs
Rui Yang (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes enhancing the generalization of the reward model to unseen data by adding text generation regularization (GRM) in the hidden layers of the reward model, thereby alleviating the issue of reward over-optimization in RLHF.
Reinforced Cross-Domain Knowledge Distillation on Time Series Data
QING XU, Zhenghua Chen (Institute for Infocomm Research A*STAR)
CodeDomain AdaptationKnowledge DistillationReinforcement LearningTime Series
π― What it does: This paper proposes an end-to-end cross-domain knowledge distillation framework RCD-KD, which combines reinforcement learning to dynamically select target domain samples and an adversarial domain discriminator to achieve unsupervised domain adaptation for lightweight models in time series tasks.
Reinforcement Learning Gradients as Vitamin for Online Finetuning Decision Transformers
Kai Yan (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
CodeTransformerReinforcement Learning
π― What it does: This paper studies the issue of online fine-tuning of the Decision Transformer after offline pre-training and proposes combining TD3 reinforcement learning gradients with the original self-supervised training to enhance the online fine-tuning effect.
π― What it does: This paper proposes a macro regulator method called MaskRegulate based on reinforcement learning, which makes local adjustments on an existing macro layout instead of starting from scratch.
Reinforcement Learning with Adaptive Regularization for Safe Control of Critical Systems
Haozhe Tian (Imperial College London), Pietro Ferraro (Imperial College London)
CodeSafty and PrivacyReinforcement LearningSequential
π― What it does: The RL-AR algorithm is proposed, which achieves safe control of critical systems by introducing a safety regularizer (MPC) and an adaptive weight combination of the RL policy during the reinforcement learning process.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Proposes Bellman Backup with Action Decomposition (BAD) and combines it with PPO to form POAD, assigning rewards to each intra-action token of the language agent at a finer granularity;
π― What it does: A post-hoc rejection framework based on the idealized distribution density ratio is proposed, which determines when to reject predictions by learning an idealized distribution regularized by Ξ±-divergence that is aligned with the original data distribution.
Hugo Chateau-Laurent (Inria centre of the University of Bordeaux), Frederic Alexandre
CodeRetrievalReinforcement LearningImage
π― What it does: This paper proves that the Differentiable Neural Dictionary (DND) used in reinforcement learning is essentially an instance of the Universal Hopfield Network (UHN), and derives the energy function through continuous approximation; it then evaluates its performance in terms of capacity and retrieval on MNIST, CIFAR-10, and Tiny ImageNet.
π― What it does: A ReLIZO zero-order optimization method is proposed, which reduces the number of queries and computational complexity while maintaining a constant sample size by estimating gradients through linear interpolation and allowing the reuse of previous queries.
Hyunseok Lee (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
CodeRecognitionGenerationTransformerLarge Language ModelReinforcement LearningText
π― What it does: A reward model-based alignment framework for detecting LLM-generated text, ReMoDetect, is proposed to distinguish between LLM-generated text and human text.
Remove that Square Root: A New Efficient Scale-Invariant Version of AdaGrad
Sayantan Choudhury (Mohammed Bin Zayed University of Artificial Intelligence), Eduard Gorbunov (Mohammed Bin Zayed University of Artificial Intelligence)
CodeOptimizationImageTabular
π― What it does: KATE is proposedβa variant of AdaGrad that removes the square root and achieves scale invariance, and it is proven to have a convergence rate comparable to AdaGrad/Adam on smooth non-convex problems.
π― What it does: A method is proposed to enhance the performance of single-step text-to-image models during the inference phase through Reward-guided Noise Optimization (ReNO).
π― What it does: This paper studies the issue of the lack of reparameterization invariance in approximate posteriors of Bayesian neural networks, proposing the use of Laplace diffusion with pseudo-Riemannian metrics to explain the geometric reasons behind the success of linearized Laplace, and providing a feasible sampling algorithm.
π― What it does: This study investigates graph-class incremental learning (GCIL) without task IDs, achieving complete no-replay and no-forgetting node classification through task analysis and graph prompts.
Representation Noising: A Defence Mechanism Against Harmful Finetuning
Domenic Rosati (Dalhousie University), Frank Rudzicz (Dalhousie University)
CodeRepresentation LearningAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A defense mechanism called Representation Noising (RepNoise) is proposed, aimed at resisting harmful fine-tuning attacks (HFA) by eliminating intermediate representation information of harmful inputs, making it difficult for attackers to recover harmful capabilities even after obtaining the model weights.
Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design
Xiangxin Zhou (University of Chinese Academy of Sciences), Jianzhu Ma (Tsinghua University)
CodeDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data
π― What it does: A generative model for dual-target drug design was developed, and a dual-target dataset based on synergistic drug combinations was constructed.
Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe
Albert Q. Jiang (University of Cambridge), Piotr MiΕoΕ (University of Warsaw)
CodeTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: This study investigates the optimal configuration for fine-tuning a pre-trained decoder language model to produce high-quality text embeddings under computational constraints using contrastive loss.
CodeGenerationAI Code AssistantTransformerLarge Language ModelText
π― What it does: This paper proposes viewing the generator-reorderer system as a communication system and proves that under multi-path redundancy generation and reordering, the probability of incorrect answers can decay exponentially/power-law with the number of candidates N, achieving asymptotically error-free generation.
π― What it does: A general framework named Resfusion is proposed, which introduces residual terms into the diffusion forward process and directly starts from the noisy degraded image to complete the reverse recovery.
ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search
Dan Zhang (Tsinghua University), Jie Tang (Tsinghua University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
π― What it does: Proposes ReST-MCTS, which utilizes tree search MCTS* and an automatically generated process reward model to perform dual training of the strategy and reward model for LLM in a self-supervised manner, generating high-quality reasoning trajectories.
Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective
Qishuai Wen (Beijing University of Posts and Telecommunications), Chun-Guang Li (Beijing University of Posts and Telecommunications)
CodeSegmentationCompressionTransformerImage
π― What it does: A Transformer decoder based on compression theory (DEPICT) is proposed, which completes semantic segmentation through PCA and a gradient-free expansion of coding rate, designing fully attention-based self-attention and cross-attention modules.
Rethinking Fourier Transform from A Basis Functions Perspective for Long-term Time Series Forecasting
Runze Yang (Shanghai Jiao Tong University), li jianxun
CodeTransformerTime Series
π― What it does: This paper proposes and implements Fourier Basis Mapping (FBM), treating the discrete Fourier transform as an expansion of sine and cosine basis functions, generating features that incorporate both time and frequency information. It achieves long-term time series prediction through three mapping networks: linear, MLP, and Transformer, while addressing the issues of inconsistent starting periods and sequence lengths present in existing Fourier-based methods.
π― What it does: A standardized human evaluation protocol for text-to-video generation models, T2VHE, is proposed to address the issues of repeatability, reliability, and practicality in existing evaluations.
Rethinking Imbalance in Image Super-Resolution for Efficient Inference
Wei Yu (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
CodeRestorationSuper ResolutionImage
π― What it does: This study investigates the imbalance problem of data distribution and model optimization in image super-resolution, proposing a weight balancing framework WBSR, which includes Hierarchical Equitable Sampling (HES) and Balanced Diversity Loss (BDLoss), and presents a gradient projection dynamic inference strategy.
Rethinking Inverse Reinforcement Learning: from Data Alignment to Task Alignment
Weichao Zhou (Boston University), Wenchao Li (Boston University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningGenerative Adversarial Network
π― What it does: A new Inverse Reinforcement Learning (IRL) framework is proposed, prioritizing task alignment over traditional data alignment, utilizing expert demonstrations as weak supervision signals, and deriving a set of candidate reward functions to better reflect task objectives.
π― What it does: A framework for OOD detection called ImOOD is proposed, which can correct the bias caused by class imbalance through regularization during the training phase, improving the accuracy of OOD detection under long-tail data distribution.
π― What it does: A learnable initial membrane potential (IMP) is proposed to enhance the expressive capability of spiking neural networks, combined with LTS processing and label smoothing TET loss to improve performance on static and event data.
Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective
Chengsen Wang (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
CodeTransformerTime Series
π― What it does: A general plugin framework GLAFF is proposed, which utilizes global information from timestamps to enhance the robustness of mainstream time series forecasting models.