π― What it does: This paper proposes a method to generate 3D meshes from source domain point clouds through implicit reconstruction, and then render these meshes into point clouds according to the target domain LiDAR parameters. This allows for the generation of point clouds in the style of the target domain without using any target domain labels, and utilizes these zero-shot data to enhance cross-domain 3D detection performance.
π― What it does: This paper proposes a general adversarial attack framework under a multi-agent reinforcement learning environment and presents a provably convergent adversarial training algorithm based on time scale separation.
Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective
Ming-Yu Chung (National Taiwan University), Tsung-Yi Ho (Chinese University of Hong Kong)
CodeKnowledge DistillationAdversarial AttackImage
π― What it does: This study investigates the feasibility of backdoor attacks during the Knowledge Induction Process (KIP) of dataset distillation, constructs a theoretical framework based on kernel methods, and designs two types of triggers (simple-trigger and relax-trigger) to implant backdoors into the distilled dataset.
Rethinking Branching on Exact Combinatorial Optimization Solver: The First Deep Symbolic Discovery Framework
Yufei Kuang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkReinforcement LearningTabular
π― What it does: A deep symbolic discovery framework called Symb4CO is proposed for the branch selection task in Mixed Integer Linear Programming (MILP), aimed at automatically learning efficient and interpretable branching strategies, and compiling the learned symbolic expressions into rules that can be directly deployed on pure CPU.
π― What it does: The study focuses on multivariate time series forecasting and proposes a module called LIFT that utilizes locally stable lead-lag relationships.
Rethinking Channel Dimensions to Isolate Outliers for Low-bit Weight Quantization of Large Language Models
Jung Hwan Heo (University of Southern California), Dongsoo Lee (NAVER Cloud)
CodeAnomaly DetectionOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes a novel per-IC quantization method and constructs the AdaDim adaptation framework to achieve low-bit weight quantization (3-bit/4-bit) while maintaining LLM performance.
Rethinking Complex Queries on Knowledge Graphs with Neural Link Predictors
Hang Yin (Tsinghua University), Yangqiu Song (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkGraph
π― What it does: This paper evaluates and redefines the reasoning methods for complex query answers (EFO 1) in knowledge graphs, pointing out the limitations of existing query embedding techniques in terms of syntax and expressiveness. It extends to complete EFO 1 queries and proposes a neural-symbolic reasoning algorithm called FIT, which combines fuzzy logic to achieve end-to-end reasoning for any EFO 1 query.
π― What it does: A high-probability information-theoretic generalization bound based on loss entropy is proposed, replacing the traditional mutual information bound, and providing two types of PAC confidence bounds: data-independent and data-dependent.
π― What it does: This paper studies the robustness of Graph Neural Networks (GNN) under label poisoning and points out six major evaluation flaws in previous research; by correcting the evaluation framework, it systematically assesses various attack methods and proposes two new efficient attacks (linear proxy attack and meta-learning attack), revealing that binary label poisoning is often more effective.
Rethinking Model Ensemble in Transfer-based Adversarial Attacks
Huanran Chen (Beijing Institute of Technology), Jun Zhu (Tsinghua University)
CodeObject DetectionAdversarial AttackConvolutional Neural NetworkTransformerVision Language ModelImage
π― What it does: This paper researches and implements a new model ensemble-based adversarial attack method aimed at improving the attack success rate in black-box environments.
Rethinking the Power of Graph Canonization in Graph Representation Learning with Stability
Zehao Dong (Washington University in St. Louis), Yixin Chen (Washington University in St. Louis)
CodeRepresentation LearningGraph Neural NetworkGraphBiomedical Data
π― What it does: This paper constructs the GC-GNN framework by using the discrete node colors generated by graph canonization as positional encoding, significantly enhancing the expressive power of GNNs, and further proposes the Universal Graph Canonization (UGC-GNN) to address the trade-off between expressiveness and stability.
π― What it does: A uniformity metric based on the squared Wasserstein distance is proposed and introduced as an auxiliary loss in various self-supervised learning methods to enhance the uniformity of representations and suppress dimensional collapse.
π― What it does: This paper proposes a retrieval-based separable representation learning framework (VDR) that utilizes natural language as a supervisory signal to map data and text into the same lexical space, achieving dimension-level separation through sparse word vectors.
π― What it does: This paper proposes a stochastic process framework that considers the uncertainty of reaction feasibility and availability in chemical retro-synthesis planning, and based on this, defines a success synthesis probability (SSP) evaluation metric. Subsequently, a greedy search algorithm called Retro-Fallback is designed to maximize this metric.
π― What it does: This paper proposes a generative framework called RetroBridge based on Markov bridges, which directly generates possible precursor molecules from target molecules to achieve single-step retro-synthesis prediction.
Retroformer: Retrospective Large Language Agents with Policy Gradient Optimization
Weiran Yao (Salesforce AI Research), Silvio Savarese (Salesforce AI Research)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposes the Retroformer framework, which utilizes a trainable retrospective language model to self-reflect and correct the prompts of LLM agents, learning policy gradients from environmental rewards;
π― What it does: A lightweight, multilingual text similarity model called RETSim has been developed, which can efficiently retrieve and cluster approximately duplicate texts in the presence of spelling errors, character-level attacks, and adversarial transformations, and is used for dataset deduplication and spam clustering.
Reverse Forward Curriculum Learning for Extreme Sample and Demo Efficiency
Stone Tao (University of California San Diego), Hao Su (University of California San Diego)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement Learning
π― What it does: A reverse + forward curriculum learning framework is proposed, which can train complex robot control strategies with only a small number of demonstrations under sparse rewards.
Revisiting Data Augmentation in Deep Reinforcement Learning
Jianshu Hu (Shanghai Jiao Tong University), Paul Weng (Duke Kunshan University)
CodeReinforcement LearningImageVideo
π― What it does: Theoretical and experimental analysis of data augmentation techniques in deep reinforcement learning on image benchmarks is conducted, proposing a novel actor-critic algorithm that includes explicit/implicit regularization and tangent propagation, along with a unified implementation framework.
π― What it does: A mini-batch Learning-to-Match (m-LTM) framework for audio-text retrieval is proposed, combining Mahalanobis distance and partial optimal transport to enhance the quality of the embedding space and noise tolerance.
Revisiting Plasticity in Visual Reinforcement Learning: Data, Modules and Training Stages
Guozheng Ma (Tsinghua University), Dacheng Tao (Nanyang Technological University)
CodeReinforcement LearningImage
π― What it does: This paper systematically studies the plasticity loss of neural networks in visual reinforcement learning, exploring the impact of data augmentation, modules, and training phases on plasticity, and proposes an adaptive replay ratio method.
Reward Model Ensembles Help Mitigate Overoptimization
Thomas Coste (University of Cambridge), David Krueger (University College London)
CodeOptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
π― What it does: This study investigates the use of reward model ensemble and conservative optimization (WCO, UWO) in RLHF to eliminate over-optimization, validating its effectiveness through two strategies: BoN and PPO.
Reward-Consistent Dynamics Models are Strongly Generalizable for Offline Reinforcement Learning
Fan-Ming Luo (Nanjing University), Yang Yu (Nanjing University)
CodeReinforcement Learning
π― What it does: The MOREC method is proposed, which learns a 'dynamic reward' in offline reinforcement learning and uses it to filter transitions generated by the model, thereby enhancing the authenticity of model rollouts and the performance of the policy.
π― What it does: A rigid protein-protein docking method called ElliDock was developed, which predicts the interface based on elliptical paraboloid and aligns the docking transformation directly from two unbound proteins.
π― What it does: Ring-A-Bell, a model-agnostic red team tool, has been developed to offline generate adversarial prompts to evaluate the security mechanisms of text-image diffusion models (such as filters and concept elimination).
RLCD: Reinforcement Learning from Contrastive Distillation for LM Alignment
Kevin Yang (Meta AI), Yuandong Tian (Meta AI)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes a new language model alignment methodβRLCD (Reinforcement Learning from Contrastive Distillation), which automatically labels preference pairs by generating positive and negative contrastive prompts and is used in the RLHF process.
π― What it does: This paper studies a distributed strategy merging method for robot teams and proposes the FLEET-MERGE algorithm to enable multi-robot learning without the need to share raw data.
Robust Adversarial Reinforcement Learning via Bounded Rationality Curricula
Aryaman Reddi (Technical University of Darmstadt), Carlo D'Eramo (University of WΓΌrzburg)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: A quantized adversarial reinforcement learning algorithm QARL based on entropy regularization is proposed to achieve adversarial robustness in high-dimensional control tasks.
Robust Model-Based Optimization for Challenging Fitness Landscapes
Saba Ghaffari (University of Illinois Urbana-Champaign), Saurabh Sinha (Georgia Institute of Technology)
CodeOptimizationDrug DiscoveryAuto EncoderSequentialBiomedical Data
π― What it does: A property-prioritized variational autoencoder (PPGVAE) is proposed and implemented for model-driven optimization, capable of overcoming the local optimum challenges caused by high imbalance and separation in protein design and continuous optimization problems.
Robustifying and Boosting Training-Free Neural Architecture Search
Zhenfeng He (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeNeural Architecture Search
π― What it does: This paper proposes a RoBoT algorithm that utilizes Bayesian optimization to learn a weighted linear combination of various training-free metrics to construct a more robust evaluation measure, and further enhances NAS performance through greedy search.
π― What it does: This study investigates the robustness of AI image detection methods, proposing diffusion purification attacks for low-disturbance watermarks and model substitution adversarial attacks for high-disturbance watermarks, demonstrating an irreconcilable trade-off between the robustness and reliability of classifier-based deepfake detectors.
RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies
Hao Cheng (University of California), Liang Sun (Alibaba Group)
CodeAnomaly DetectionTime Series
π― What it does: A unified theory and method for time series forecasting with anomalies (TSFA) is proposed, and the RobustTSF algorithm is constructed.
π― What it does: Using a single out-of-distribution (OoD) image to generate proxy data, the model is fine-tuned through backpropagation without using the original training data, injecting a verifiable backdoor watermark to ensure that the model maintains performance on the original task while achieving intellectual property protection.
π― What it does: A CVaR-based safe collaborative filtering algorithm, SAFER2, is proposed, significantly improving the recommendation quality for low-satisfaction (tail) users.
Safe Offline Reinforcement Learning with Feasibility-Guided Diffusion Model
Yinan Zheng (Tsinghua University), Jingjing Liu (Tsinghua University)
CodeSafty and PrivacyReinforcement LearningDiffusion modelTabularBenchmark
π― What it does: A secure offline reinforcement learning framework FISOR is proposed, which transforms hard safety constraints into feasibility-dependent objectives, and utilizes HJ reachability analysis, expectile regression, weighted advantage learning, and diffusion models to achieve the decoupled training of three objectives: safety, reward maximization, and offline data regularization.
Safe RLHF: Safe Reinforcement Learning from Human Feedback
Josef Dai, Yaodong Yang (Peking University)
CodeSafty and PrivacyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes Safe RLHF, an improved RLHF training framework that decouples human preferences for 'usefulness' and 'harmlessness', and dynamically balances the two through safety constraints and Lagrangian methods to achieve a safer and more useful LLM.
SafeDreamer: Safe Reinforcement Learning with World Models
Weidong Huang (Peking University), Yaodong Yang (Peking University)
CodeSafty and PrivacyReinforcement LearningWorld ModelImage
π― What it does: This paper presents SafeDreamer, a safe reinforcement learning framework that combines Lagrangian methods with world model planning, achieving near-zero cost performance in visual and low-dimensional input tasks.
Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions
Federico Bianchi (Stanford University), James Zou (Stanford University)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This study investigates the impact of adding a small number of safety examples in instruction fine-tuning on the safety of large language models, and validates its effectiveness on various safety evaluation datasets.
SALMON: Self-Alignment with Instructable Reward Models
Zhiqing Sun (Carnegie Mellon University), Chuang Gan (UMass Amherst)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: By using only a small number of human-defined principles and 6 examples, the SALMON method achieves self-alignment on large language models, resulting in the Dromedary-2 model;
CodeTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
π― What it does: This paper presents SALMONN, a multimodal large language model that integrates speech, audio events, and music to achieve general auditory capabilities.
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
Chongyu Fan (Michigan State University), Sijia Liu (Michigan State University)
CodeClassificationGenerationDiffusion modelImage
π― What it does: A machine model forgetting method based on gradient weight saliency (SalUn) is proposed and implemented, applicable to image classification and generation tasks.
Sample-Efficient Quality-Diversity by Cooperative Coevolution
Ke Xue (Nanjing University), Chao Qian (Nanjing University)
CodeOptimizationReinforcement LearningSequential
π― What it does: A Cooperative Coevolution Quality-Diversity (CCQD) framework is designed, which significantly improves the sampling efficiency of QD algorithms by splitting the policy network into representation and decision layers, maintaining two subpopulations that co-evolve.
π― What it does: This paper proposes a new GAN training framework called Slicing Adversarial Network (SAN), which ensures the metric consistency between the generated distribution and the real distribution by enforcing directional optimality in the discriminator, and can transform various existing GANs into SAN with two lightweight modifications to the discriminator.
π― What it does: A method is proposed to compress activation maps into wide activations with local sparse structures through Structured Activation Sparsification (SAS), and to achieve efficient matrix multiplication using NVIDIA Sparse Tensor Core, improving model accuracy without increasing the multiplication count.
Scalable Language Model with Generalized Continual Learning
Bohao PENG (Chinese University of Hong Kong), Jiaya Jia
CodeClassificationRetrievalOptimizationTransformerLarge Language ModelTextSequentialRetrieval-Augmented Generation
π― What it does: A scalable language model (SLM) is proposed to achieve continuous learning for sequence tasks, avoiding the use of experience replay, optimization constraints, or task IDs during inference.
Scalable Modular Network: A Framework for Adaptive Learning via Agreement Routing
Minyang Hu (Institute of Computing Technology Chinese Academy of Sciences), Xilin CHEN
CodeClassificationMeta LearningTransformerMixture of ExpertsImage
π― What it does: A scalable modular network (SMN) is proposed, which dynamically selects and combines specialized modules based on different inputs through an agreement router, and supports the introduction of new modules after pre-training to enhance adaptability.
Hyunho Kim (Sungkyunkwan University), Jong-Seok Lee (Sungkyunkwan University)
CodeTabular
π― What it does: This paper studies a scalable monotonic neural network (SMNN) that ensures monotonicity for specified inputs and achieves efficient training.
π― What it does: Proposes a Scalable Neural Network Kernel (SNNK) as an alternative to traditional Feedforward Layers (FFL), achieving more efficient forward computation by using only dot product kernels after decoupling inputs and parameters;
π― What it does: This paper first conducts theoretical and experimental analysis of the existing Activation Shaping method (ASH), proving that its core enhancement of OOD detection is due to scaling the activations rather than pruning; it then proposes a post-hoc method called SCALE that only uses scaling, further improving OOD detection performance without reducing ID accuracy; based on this, the scaling concept is transferred to the training phase, designing Intermediate Tensor Shaping (ISH) for activation weighting optimization during training, thus achieving higher OOD detection performance with lower training costs.
Xiaoran Liu (Fudan University), Dahua Lin (Shanghai AI Lab)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study investigates the extrapolation performance of LLMs based on RoPE when exceeding the pre-training context length and proposes a scaling law that can uniformly describe the extrapolation capability of RoPE.
Scaling physics-informed hard constraints with mixture-of-experts
Nithin Chalapathi (University of California), Aditi S. Krishnapriyan (University of California)
CodeMixture of ExpertsTime SeriesPhysics Related
π― What it does: A hard physical constraint framework based on Mixture-of-Experts is proposed, achieving strict adherence to physical laws in neural PDE solvers through local optimization.
π― What it does: A supervised local learning method named AugLocal is proposed, which enhances the collaboration between local layers and subsequent layers by uniformly sampling subsets of the subsequent layers in the auxiliary network of each hidden layer, achieving performance close to BP on large-scale networks.
CodeTransformerLarge Language ModelPrompt EngineeringContrastive LearningVideoTextMultimodality
π― What it does: Utilize large language models to generate descriptions of state changes corresponding to steps, and learn a structured state space through visual-language alignment, thereby completing process planning in instructional videos.
π― What it does: A novel offline reinforcement learning algorithm named SRPO (Score Regularized Policy Optimization) is proposed, which utilizes the score function of a pre-trained diffusion behavior model to regularize the policy at the gradient level, enabling deterministic policy extraction without diffusion sampling.
π― What it does: The FoldFlow series model is proposed, utilizing flow matching to learn continuous dynamics on SE(3), generating protein backbone structures from prior distributions, covering high-quality, designable, and diverse samples of 300 amino acids.
π― What it does: A search-based offline imitation learning method called SEABO is proposed, which automatically annotates rewards using expert demonstrations and unlabeled data, and integrates with offline RL algorithms.
π― What it does: The SEAL framework is proposed, which constructs a representative degradation set using clustering methods to systematically evaluate real super-resolution methods.
Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation
Niels MΓΌndler (ETH Zurich), Martin Vechev (ETH Zurich)
CodeGenerationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: This paper studies the contradictory hallucinations that occur when large language models generate text and proposes a complete framework from triggering to detection to mitigation.
Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
Akari Asai (University of Washington), Hannaneh Hajishirzi (IBM Research AI)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Train a language model called SELF-RAG that can decide during the inference process whether to retrieve documents, evaluate the relevance and support of the retrieval results, and self-critique the generated content.
Self-Supervised Contrastive Learning for Long-term Forecasting
Junwoo Park (KAIST), Edward Choi (KAIST)
CodeContrastive LearningTime Series
π― What it does: This paper proposes a contrastive learning loss called AutoCon based on global autocorrelation, combined with an improved decomposition architecture, to achieve self-supervised learning for long-period variations outside the sliding window, thereby enhancing long-period time series prediction performance.
π― What it does: Compress the unlabeled dataset into a small number of synthetic samples, and use these samples for self-supervised pre-training before transferring to various downstream tasks.
π― What it does: The HERO framework is proposed, which does not rely on predefined meta-paths in self-supervised heterogeneous graph learning, jointly learning homogeneous representations of similar nodes and heterogeneous representations of different types of nodes.
Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment
Bowen Gao (Institute for AI Industry Research, Tsinghua University), Yanyan Lan (Institute for AI Industry Research, Tsinghua University)
CodeDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningBiomedical Data
π― What it does: A self-supervised pre-training framework called ProFSA is proposed, based on the alignment of protein fragments and their surrounding environments, to learn high-quality binding site representations.
SelfCheck: Using LLMs to Zero-Shot Check Their Own Step-by-Step Reasoning
Ning Miao (University of Oxford), Tom Rainforth (University of Oxford)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes SelfCheck, a zero-shot, step-by-step self-checking framework based on LLMs, aimed at identifying errors in its own reasoning chain and improving answer quality.
π― What it does: Utilize optical flow features to learn semantic representations of dynamic scenes, and supervise semantics using volume density as an opacity prior;
Yang bai, Chun-Mei Feng (Institute of High Performance Computing)
CodeRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
π― What it does: This paper proposes the use of dynamic sentence-level prompts to enhance the retrieval performance of queries composed of reference images and relative descriptions (CIR), replacing traditional late fusion and pseudo-word methods.
π― What it does: Treating the autoregressive sequence generation problem as an imitation learning task, the SequenceMatch method is proposed, which trains the model by minimizing various divergences of occupancy metrics (such as ΟΒ² divergence) and incorporates backspace actions to alleviate cumulative errors.
Lukas Muttenthaler (Technische UniversitΓ€t Berlin), Klaus Robert Muller
CodeClassificationImage
π― What it does: A training framework based on ensemble learning called Odd Out (OKO) is proposed to improve the accuracy and calibration of classification models.
π― What it does: Proposes the Sharpness-Aware Data Poisoning Attack (SAPA), which approximates the attack effect of the worst retrained model by utilizing the sharpness of the DNN loss landscape during the generation of poisoning samples;
π― What it does: Smaller models (1.3B, 2.7B) are obtained from the large pre-trained model (LLaMA2-7B) through structured pruning, and further pre-training is conducted based on this.
SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore
Sewon Min (University of Washington), Luke Zettlemoyer (University of Washington)
CodeRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Designed and evaluated a language model SILO that separates parametric and non-parametric components to reduce legal risks associated with training data.
π― What it does: By learning differentiable k-forms, a message-passing-free geometric learning framework is obtained for integrating k-simplices in embedded spaces.
SineNet: Learning Temporal Dynamics in Time-Dependent Partial Differential Equations
Xuan Zhang (Texas A&M University), Shuiwang Ji (University of Pittsburgh)
CodeConvolutional Neural NetworkTime SeriesPhysics Related
π― What it does: A multi-stage U-Net structure named SineNet is designed and proposed for learning the temporal dynamics of time-dependent partial differential equations (PDEs).
Skeleton-of-Thought: Prompting LLMs for Efficient Parallel Generation
Xuefei Ning (Tsinghua University), Yu Wang (Tsinghua University)
CodeGenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposes the Skeleton-of-Thought (SoT) framework, which first allows the LLM to generate an answer skeleton, and then expands the skeleton points in parallel to reduce inference latency.
Sliced Denoising: A Physics-Informed Molecular Pre-Training Method
Yuyan Ni (Chinese Academy of Sciences), Yanyan Lan (Tsinghua University)
CodeDrug DiscoveryTransformerGraphPhysics Related
π― What it does: This paper proposes a new molecular pre-training method called Sliced Denoising (SliDe), which simulates molecular energy distribution by adding Gaussian noise to bond lengths, bond angles, and torsion angles, and efficiently regresses force fields using a random slicing technique.
π― What it does: Proposes the use of control variates to reduce the Monte Carlo variance of the sliced Wasserstein (SW) distance, and provides two Gaussian approximation-based upper and lower bound control variates;
SliceGPT: Compress Large Language Models by Deleting Rows and Columns
Saleh Ashkboos (ETH Zurich), James Hensman (Microsoft)
CodeCompressionTransformerLarge Language ModelText
π― What it does: A post-training sparsification method called SliceGPT is proposed, which uses orthogonal transformations to slice the Transformer matrix into smaller dense matrices, thereby reducing parameters and embedding dimensions, compressing LLM while maintaining high performance.
π― What it does: This paper proposes a kernel smoothing-based calibration error metric SmoothECE and its corresponding smoothed reliability diagram, to replace the traditional binning ECE and reliability diagram;
π― What it does: A method for evaluating text-to-image generation models based on community implicit feedback is proposed and validatedβSocial Reward;
π― What it does: This paper proposes Social-Transmotion, a general Promptable Transformer model that can utilize various visual cues (trajectories, 3D/2D poses, bounding boxes) for human trajectory prediction.
SocioDojo: Building Lifelong Analytical Agents with Real-world Text and Time Series
Junyan Cheng (Dartmouth College), Peter Chin (Dartmouth College)
CodeRecommendation SystemAnomaly DetectionOptimizationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityTime SeriesFinance RelatedRetrieval-Augmented Generation
π― What it does: Constructed the SocioDojo open-source lifelong learning environment and proposed the Hyperportfolio task to evaluate agents' capabilities in social analysis and decision-making;
Soft Mixture Denoising: Beyond the Expressive Bottleneck of Diffusion Models
Yangming Li (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeRestorationGenerationDiffusion modelImage
π― What it does: Proposes the Soft Mixture Denoising (SMD) model, which improves the backward denoising process of diffusion models and addresses the expression bottleneck caused by the traditional single Gaussian prior.
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Robert Huben (EleutherAI), Lee Sharkey (MATS)
CodeExplainability and InterpretabilityRepresentation LearningTransformerAuto EncoderText
π― What it does: Using sparse autoencoders for dictionary learning on the internal activations of language models to extract sparse, unambiguous features, reduce polysemanticity, and reveal the causal mechanisms of the model.
π― What it does: An iterative magnitude pruning method based on Stochastic Weight Averaging with Multiple Particles (SWAMP) is proposed, which improves the sparsification effect of traditional IMP.
π― What it does: This paper proposes a visual Transformer named SparseFormer, which achieves sparse perception and recognition of images by using a minimal number of tokens (e.g., 9 to 81) in the latent space and employing sparse feature sampling.
π― What it does: A Spatial Awareness Transformer (SAT) is proposed for embodied agents, incorporating spatial embeddings into experiential memory and designing location-based hierarchical reading and adaptive memory allocation.
π― What it does: This paper proposes the 'Space-Time Approximation (STA)' method, which for the first time achieves the conversion of pre-trained Transformers (such as CLIP's ViT-B/32) into event-driven Spiking Neural Networks (SNN) in a training-independent manner, while maintaining performance close to the original model on tasks such as zero-shot classification.
Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation
Yuan Yuan (Tsinghua University), Yong Li (Tsinghua University)
CodeGenerationDomain AdaptationMeta LearningTransformerPrompt EngineeringDiffusion modelTime Series
π― What it does: A generative pre-training framework GPD based on diffusion models is proposed for cross-city spatiotemporal prediction with few-shot learning.
π― What it does: A novel implicit neural representation network SPDER is designed and implemented, utilizing sine multipliers and sublinear decay activation functions, achieving high-precision continuous representations of images, audio, and video without the need for pre-encoding or hyperparameter tuning.
π― What it does: This paper proposes SpeechTokenizer, a unified speech tokenizer, and constructs the SLMTokBench benchmark, subsequently implementing a Unified Speech Language Model (USLM) based on this tokenizer.
π― What it does: A directly trainable spiking neural network based on Transformer, called MetaβSpikeFormer, is proposed, which can simultaneously handle visual tasks such as classification, detection, and segmentation.
SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores
Zhiyu Mei (Tsinghua University), Yi Wu (Tsinghua University)
CodeReinforcement LearningVideo
π― What it does: A scalable distributed reinforcement learning system SRL has been designed and implemented, supporting CPU/GPU parallel training at the level of thousands of cores, achieving high throughput in various environments.
Stabilizing Backpropagation Through Time to Learn Complex Physics
Patrick Schnell (Technical University of Munich), Nils Thuerey (Technical University of Munich)
CodeOptimizationReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningTime SeriesPhysics Related
π― What it does: This paper addresses the issues of gradient explosion and vanishing in the long-term unrolling training of physical simulators and neural networks. It proposes a backpropagation method using gradient stop and rotational correction to achieve more balanced and stable update vectors, and validates its effectiveness in control tasks.
William Rudman (Brown University), Carsten Eickhoff (University of TΓΌbingen)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a differentiable, mini-batch stable regularization method called I-STAR based on IsoScoreβ, aimed at adjusting the isotropy of the embedding space of large language models during the fine-tuning process.
π― What it does: This paper proposes three types of stable Neural Stochastic Differential Equations (Neural SDE) β Langevin, linear noise, and geometric SDE, and embeds controlled paths into the drift term to address issues of irregular sampling and missing values in time series data encountered in reality.
Stack Attention: Improving the Ability of Transformers to Model Hierarchical Patterns
Brian DuSell (ETH Zurich), David Chiang (University of Notre Dame)
CodeTransformerText
π― What it does: Introducing differentiable stacks (superposition stack and nondeterministic stack) into the Transformer as an attention mechanism to enhance the model's ability to recognize and model hierarchical patterns and context-free languages (CFL) without the need for syntactic supervision.