π― What it does: A time-difference-based InfoNCE (TD-InfoNCE) learning framework is proposed and implemented to estimate the future state occupancy distribution of the target policy in offline or data-limited scenarios, thereby driving goal-conditioned reinforcement learning (Goal-Conditioned RL) algorithms.
π― What it does: It is proven that SimCLR (InfoNCE) is equivalent to performing spectral clustering on the similarity graph generated by data augmentation, and this perspective is extended to CLIP in multimodal learning; a Kernel-InfoNCE loss is proposed based on the maximum entropy principle, using a mixed form of exponential kernel instead of Gaussian kernel, with experiments validating its performance superior to the original SimCLR on CIFAR-10/100 and TinyImageNet.
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningImage
π― What it does: A reinforcement learning method based on contrastive learning is proposedβContrastive Preference Learning (CPL), which learns optimal policies directly from human preference data, eliminating the need for a reward model and RL optimization steps.
Controlled Text Generation via Language Model Arithmetic
Jasper Dekoninck (ETH Zurich), Martin Vechev (ETH Zurich)
CodeGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A framework called model arithmetic is proposed, which utilizes weighted KL divergence to combine multiple LLMs and attribute models, supporting fine-grained text control through formulas, and incorporates a novel joint operator and extended speculative sampling for efficient inference.
Controlling Vision-Language Models for Multi-Task Image Restoration
Ziwei Luo (Uppsala University), Thomas B. SchΓΆn
CodeRestorationTransformerVision Language ModelDiffusion modelContrastive LearningImage
π― What it does: Train a degradation-aware CLIP (DA-CLIP) controller to enable the pre-trained CLIP image encoder to output high-quality features from degraded images and predict the type of degradation, then integrate these embeddings with an image restoration network to enhance multi-task image restoration performance.
Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model
Zihan Zhong (Tsinghua University), Chun Yuan (Tsinghua University)
CodeSegmentationTransformerMixture of ExpertsImageAgriculture Related
π― What it does: This paper proposes Conv-LoRA, a PEFT method that fine-tunes only a few parameters on the SAM pre-trained ViT encoder to improve multi-domain semantic segmentation.
π― What it does: This paper proposes the Convolutional Deep Kernel Machine (DKM) and designs an efficient cross-domain inducing point approximation method, achieving Bayesian representation learning for infinitely wide convolutional networks.
π― What it does: Proposes the COPlanner framework, which combines conservative model rollouts and optimistic environment exploration to mitigate model errors.
Copula Conformal prediction for multi-step time series prediction
Sophia Huiwen Sun (University of California San Diego), Rose Yu (University of California San Diego)
CodeTime Series
π― What it does: The CopulaCPTS algorithm is proposed, which uses copula and induced consistency prediction to generate confidence intervals for multi-step time series.
CoRe-GD: A Hierarchical Framework for Scalable Graph Visualization with GNNs
Florian GrΓΆtschla (ETH Zurich), Roger Wattenhofer (ETH Zurich)
CodeGraph Neural NetworkGraphBenchmark
π― What it does: This paper presents CoRe-GD, a scalable graph visualization framework that utilizes hierarchical graph refinement and position reconnection techniques to learn and optimize the pressure function of graphs and output node layouts.
COSA: Concatenated Sample Pretrained Vision-Language Foundation Model
Sihan Chen (University of Chinese Academy of Sciences), Jing Liu
CodeRetrievalTransformerVision Language ModelContrastive LearningImageVideoText
π― What it does: During the pre-training phase, pseudo video-paragraph corpora are constructed by online concatenation of multiple images and corresponding texts, thereby jointly learning static and event-level temporal representations of vision and language.
Counting Graph Substructures with Graph Neural Networks
Charilaos Kanatsoulis, Alejandro Ribeiro (University of Pennsylvania)
CodeGraph Neural NetworkGraph
π― What it does: This paper analyzes and proves that message-passing graph neural networks (GNNs) can generate equivariant representations through appropriate activation and normalization layers using random uninformative node inputs, thereby counting and enumerating substructures such as cycles, cliques, quasi-cliques, and connected components in graphs, and can generalize to unknown graphs.
CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets
Lifan Yuan (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelTextTabularRetrieval-Augmented GenerationChain-of-Thought
π― What it does: The CRAFT framework is proposed to automatically generate and retrieve a set of tools for specific tasks to enhance the reasoning and execution capabilities of LLMs.
π― What it does: A cross-image object-level self-supervised learning method called CrIBo is proposed, which enhances dense visual representation learning by using object-level nearest neighbor guidance during the training process.
Critical Learning Periods Emerge Even in Deep Linear Networks
Michael Kleinman (Stanford University), Stefano Soatto (University of California Los Angeles)
CodeTabularOrdinary Differential Equation
π― What it does: This study investigates the phenomenon of critical learning periods in deep linear networks and verifies its dependence on depth and data structure through analytical differential equations and numerical simulations.
π― What it does: A cross-modal contextual diffusion model called CONTEXTDIFF is proposed, which enhances text-guided visual generation and editing by incorporating cross-modal context into the forward and backward diffusion processes.
Curiosity-driven Red-teaming for Large Language Models
Zhang-Wei Hong (Massachusetts Institute of Technology), Pulkit Agrawal
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: By combining reinforcement learning with curiosity-driven exploration, this paper automates the generation of diverse and effective red team test cases to evaluate the robustness of large language models (LLMs) in producing toxic, erroneous, or harmful content.
π― What it does: This paper proposes a distance-aware fair adversarial training (DAFA) based on inter-class similarity, which enhances the robustness of difficult-to-classify categories by assigning them greater loss weights and adversarial margins.
π― What it does: A multi-stage dataset distillation (PDD) framework is proposed, which recursively generates small synthetic subsets at different training stages, capturing the complete training dynamics step by step based on the previous stage's subset, thereby enhancing the generalization performance of the distilled data.
Data-independent Module-aware Pruning for Hierarchical Vision Transformers
Yang He (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
CodeClassificationCompressionTransformerImage
π― What it does: A data-independent module-aware pruning method called DIMAP is proposed for hierarchical vision Transformers (Swin Transformer), aimed at compressing model parameters and computational load without relying on input data.
DATS: Difficulty-Aware Task Sampler for Meta-Learning Physics-Informed Neural Networks
Maryam Toloubidokhti (Rochester Institute of Technology), Linwei Wang (Rochester Institute of Technology)
CodeOptimizationMeta LearningTabularPhysics Related
π― What it does: DATS is proposed, a difficulty-aware task sampler that prioritizes training higher difficulty PDE tasks during meta-learning of PINN and dynamically allocates residual points.
π― What it does: A domain-agnostic latent diffusion model DDMI is designed for synthesizing high-quality implicit neural representations (INR), capable of generating continuous functions across various signal domains such as images, 3D shapes, videos, and NeRF.
π― What it does: This paper proposes the Vector Field Network (VFN) to achieve stronger geometric feature extraction under protein frame encoding without atomic information, thereby improving the quality and diversity of de novo protein design.
Debiased Collaborative Filtering with Kernel-Based Causal Balancing
Haoxuan Li (Peking University), Peng Cui (Tsinghua University)
CodeRecommendation SystemTabular
π― What it does: A kernel-based causal balancing method is proposed, which estimates the propensity score by learning a kernel function that can approximate all balancing functions in collaborative filtering, thereby eliminating selection bias in the observed data.
Tomasz Limisiewicz (Charles University), TomΓ‘Ε‘ Musil (Charles University)
CodeTransformerLarge Language ModelText
π― What it does: Proposes and implements an algorithm for debiasing through model adaptation (DAMA), which modifies the LLaMA language model without architectural changes to reduce gender bias.
Decoding Natural Images from EEG for Object Recognition
Yonghao Song (Tsinghua University), Xiaorong Gao (Chinese Academy of Sciences)
CodeClassificationRecognitionContrastive LearningMultimodalityTime Series
π― What it does: A self-supervised EEG-image contrastive learning framework (NICE) has been constructed to decode natural images from EEG signals and achieve zero-shot object recognition.
π― What it does: A sampling strategy called Decomposed Diffusion Sampling (DDS) is proposed, which combines the DDIM sampling of diffusion models with the Krylov subspace method (conjugate gradient CG) to iteratively solve inverse problems quickly and with high quality by using CG in the tangent space of the denoised points.
π― What it does: This study investigates how to alleviate congestion in online markets by controlling a subset of product display features on the platform, thereby enhancing social welfare.
π― What it does: A new DOCKGEN blind ligand docking benchmark is proposed, analyzing and enhancing the generalization performance of docking methods based on diffusion models;
π― What it does: This study investigates the predictive behavior of neural networks on out-of-distribution (OOD) inputs under high-dimensional input conditions, finding that they often tend towards the 'Optimal Constant Solution (OCS)', which is a constant prediction that minimizes training loss while ignoring the input.
π― What it does: Two anomaly detection models based on deep orthogonal hypersphere compression (DOHSC and DO2HSC) are proposed, which achieve decision boundaries that better conform to the hypersphere assumption through an orthogonal projection layer, and introduce dual hypersphere compression to alleviate high-dimensional bubble effects and sparsity issues.
π― What it does: The research proposes an improved heuristic method based on deep reinforcement learning, using graph neural networks to represent complete job shop scheduling solutions and directly selecting neighborhood operations through learned policies to enhance solution quality.
DeepZero: Scaling Up Zeroth-Order Optimization for Deep Model Training
Aochuan Chen (Michigan State University), Sijia Liu (Michigan State University)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: This paper proposes the DeepZero framework, aimed at achieving zero-order (gradient-free) optimization training of deep networks from scratch.
CodeOptimizationExplainability and InterpretabilityTransformerLarge Language ModelImageText
π― What it does: A method is proposed to extract generalizable interaction primitives between multiple models to explain the reasoning logic of DNNs.
Defining Expertise: Applications to Treatment Effect Estimation
Alihan HΓΌyΓΌk, Mihaela van der Schaar (University of Cambridge)
CodeTabular
π― What it does: This paper studies the role of decision-makers' expert knowledge in causal inference, proposing two types of experts: predictive and prognostic, and treating them as prior biases in treatment effect estimation.
π― What it does: By reconstructing the metadata of CLIP and combining substring matching with balanced sampling, the MetaCLIP algorithm is proposed, which trims the massive pool of image-text pairs from the web into a high-quality, distribution-balanced dataset, and validates its superiority on multi-scale ViT models.
Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition
Faisal Hamman (University of Maryland), Sanghamitra Dutta (University of Maryland)
CodeOptimizationFederated LearningTabular
π― What it does: This paper studies the trade-off between local and global fairness in federated learning, utilizing the PID method from information theory to decompose the sources of unfairness, and proposes the AGLFOP optimization framework to evaluate the theoretical limits of accuracy and fairness.
π― What it does: Proposes DDSM (Denoising Diffusion Step-aware Models), which uses networks of different sizes at different steps of the diffusion process to reduce redundant computations;
Byeongjun Park (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
CodeRestorationGenerationDiffusion modelImage
π― What it does: A method called Denoising Task Routing (DTR) is proposed, which establishes dedicated information channels for different denoising tasks (corresponding to different time steps) in diffusion models through channel masking, explicitly addressing the negative transfer problem in multi-task learning.
π― What it does: The research proposes Decomposed Prompt Tuning (DEPT) for parameter-efficient fine-tuning, addressing the issue of overly long input sequences caused by soft prompts;
Detecting Machine-Generated Texts by Multi-Population Aware Optimization for Maximum Mean Discrepancy
Shuhai Zhang (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: A multi-population perception optimization method for maximum mean discrepancy (MMD) called MMD-MP is proposed and implemented, and paragraph-level and sentence-level machine-generated text detectors are constructed based on this method.
π― What it does: A pre-training framework for RGB-D, called DFormer, is proposed, which can learn transferable RGB-D representations and be used directly in downstream tasks;
Diagnosing Transformers: Illuminating Feature Spaces for Clinical Decision-Making
Aliyah R. Hsu (University of California Berkeley), Bin Yu (University of California Berkeley)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerSupervised Fine-TuningTextBiomedical Data
π― What it does: A framework called SUFO (Supervised Probing + Unsupervised similarity + Feature Dynamics + Outlier analysis) is proposed to systematically explain the feature space and its evolution process of fine-tuned transformers in clinical tasks.
π― What it does: A mechanism is proposed to implant invisible memory in text-to-image diffusion models to detect unauthorized use of specific data during training or fine-tuning.
Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking
Kaifeng Lyu (Princeton University), Wei Hu (University of Michigan)
CodeTabular
π― What it does: This paper theoretically proves that under the conditions of large initialization and small weight decay, the training of neural networks experiences an implicit bias separation from kernel mode to rich model in two stages, with a sharp performance transition (grokking) occurring at 1/Ξ» log Ξ±.
DiffEnc: Variational Diffusion with a Learned Encoder
Beatrix Miranda Ginn Nielsen (Technical University of Denmark), Ole Winther (Technical University of Denmark)
CodeGenerationDiffusion modelImage
π― What it does: This paper proposes DiffEncβa variational diffusion model that incorporates a time-dependent encoder into the diffusion process, maintaining the sampling time while enhancing the model's flexibility.
π― What it does: A differentiable Euler Characteristic Transform (DECT) is proposed, enabling end-to-end learning and optimization of shape features.
Differentially Private Synthetic Data via Foundation Model APIs 1: Images
Zinan Lin (Microsoft Research), Sergey Yekhanin (Microsoft Research)
CodeData SynthesisSafty and PrivacyDiffusion modelImageBiomedical Data
π― What it does: Using the inference API of large pre-trained models, a training-free differential privacy synthetic data generation framework called Private Evolution (PE) is proposed, which generates synthetic samples that closely resemble the distribution of private data through gradual evolution.
π― What it does: A training-independent Circular One-Way Diffusion (COW) method is proposed, achieving high-fidelity generation of visual and textual conditions by injecting visual conditions into a pre-trained diffusion model and cyclically perturbing the reconstruction.
Diffusion-TS: Interpretable Diffusion for General Time Series Generation
Xinyu Yuan (Hefei University of Technology), Yan Qiao (Hefei University of Technology)
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerDiffusion modelTime SeriesFinance Related
π― What it does: A time series generation framework based on diffusion models, Diffusion-TS, is proposed, utilizing a transformer architecture and decoupled seasonal-trend-error decomposition to achieve high-quality synthesis of multivariate long-term time series.
π― What it does: This paper presents DiffusionNAGβa conditional neural architecture generation framework based on diffusion models, which can directly guide the generation of neural network architectures that meet target attributes (such as high accuracy and low latency) using a predictor.
Dirichlet-based Per-Sample Weighting by Transition Matrix for Noisy Label Learning
HeeSun Bae (KAIST), Il-chul Moon
CodeClassificationImage
π― What it does: This paper proposes a resampling method called RENT based on the noise transfer matrix to improve the learning effectiveness with noisy labels.
Discovering modular solutions that generalize compositionally
Simon Schug (ETH Zurich), Angelika Steger (ETH Zurich)
CodeKnowledge DistillationMeta Learning
π― What it does: This paper studies the method of achieving compositional generalization through modular hypernetworks in multi-task learning, using a teacher-student framework to validate theory and practice.
π― What it does: A time-aware RL objective function is proposed, which adapts to different training durations by incorporating training cycle information into the learning algorithm.
π― What it does: Proposes the DisenBooth framework, which utilizes identity-preserving embeddings and identity-agnostic embeddings for decoupled fine-tuning, achieving theme-driven text-to-image generation.
Dissecting learning and forgetting in language model finetuning
Xiao Zhang (Tsinghua University), Ji Wu (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
π― What it does: Decomposes the learning and forgetting processes of domain fine-tuned language models, assessing the variations in biases related to themes, styles, and factual knowledge.
Distinguished In Uniform: Self-Attention Vs. Virtual Nodes
Eran Rosenbluth (RWTH Aachen University), Martin Grohe (RWTH Aachen University)
CodeGraph Neural NetworkTransformerGraph
π― What it does: This paper theoretically proves and experimentally compares the unified expressive power of Graph Transformer (GT) and Message Passing Graph Neural Network with Virtual Nodes (MPGNN+VN), finding that neither is a global universal approximator, and each cannot fully simulate the functions of the other.
Distributional Preference Learning: Understanding and Accounting for Hidden Context in RLHF
Anand Siththaranjan (University of California), Dylan Hadfield-Menell (Massachusetts Institute of Technology)
CodeAdversarial AttackReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: The study investigates how preference learning aggregates information in the presence of hidden contexts (such as diverse preferences, competing objectives, and unconscious factors) and proposes Distributional Preference Learning (DPL) to explicitly model the effects of hidden contexts; it also validates the mitigation effect of DPL on 'jailbreak' attacks in the RLHF tasks of large language models.
Distributionally Robust Optimization with Bias and Variance Reduction
Ronak Mehta (University of Washington), Zaid Harchaoui (University of Washington)
CodeOptimizationTabular
π― What it does: A stochastic gradient algorithm named Prospect is proposed for solving distributionally robust optimization problems with spectral risk measures (such as CVaR, extremile, ESRM);
Divide and not forget: Ensemble of selectively trained experts in Continual Learning
Grzegorz RypeΕΔ (Warsaw University of Technology), BartΕomiej Twardowski (Universitat AutΓ²noma de Barcelona)
CodeClassificationKnowledge DistillationConvolutional Neural NetworkMixture of ExpertsImage
π― What it does: This paper proposes a sample-free continuous learning method called SEED, which utilizes a limited number of expert networks and fine-tunes a single expert for each new task, thereby reducing forgetting and promoting expert diversity.
DMBP: Diffusion model-based predictor for robust offline reinforcement learning against state observation perturbations
Zhihe YANG, Yunjian Xu (Chinese University of Hong Kong)
CodeReinforcement LearningDiffusion modelTabular
π― What it does: A diffusion model-based predictor (DMBP) is proposed to denoise state observation disturbances in offline reinforcement learning, thereby enhancing policy robustness.
DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text
Xianjun Yang (University of California), Haifeng Chen (NEC Laboratories America)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: This paper proposes a zero-training, interpretable GPT text detection method called DNA-GPT, which utilizes text truncation and regeneration, and then judges whether the text is generated by LLM based on n-gram similarity or probability differences.
π― What it does: An efficient foundational model for multi-species genome pre-training, DNABERT-2, has been constructed, with improvements such as BPE tokenization and ALiBi positional encoding.
π― What it does: The study investigates the role of generated data in contrastive learning, finding that generated data can sometimes harm performance, and proposes an adaptive data augmentation strategy called AdaInf.
Does CLIPβs generalization performance mainly stem from high train-test similarity?
Prasanna Mayilvahanan (University of TΓΌbingen), Wieland Brendel (TΓΌbingen AI Center)
CodeClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImage
π― What it does: The paper quantifies the perceptual similarity between the CLIP training set and the OOD benchmark set, and based on this, conducts data pruning experiments to test whether high training-testing similarity is the main reason for CLIP's outstanding performance on OOD.
Does Writing with Language Models Reduce Content Diversity?
Vishakh Padmakumar (New York University), He He (New York University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Evaluate the impact of different LLMs (GPT-3, InstructGPT) on the diversity and homogenization of argumentative writing in controlled experiments, and quantify the contributions of model text and user text to diversity.
DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models
Yung-Sung Chuang (Massachusetts Institute of Technology), Pengcheng He (Microsoft)
CodeGenerationTransformerLarge Language ModelContrastive LearningText
π― What it does: A decoding strategy called DoLa is proposed, which does not require retrieval or fine-tuning, to enhance the factuality of large language models by comparing the logits of different Transformer layers.
Domain constraints improve risk prediction when outcome data is missing
Sidhika Balachandar (Cornell University), Emma Pierson (Cornell University)
CodeBiomedical DataElectronic Health Records
π― What it does: A Bayesian model is proposed to estimate disease risk and evaluate human testing decisions in selective labeling scenarios (only observing the results of tested subjects).
π― What it does: A molecular pre-trained language model called MOLGEN based on SELFIES was constructed and trained, guided by a chemical feedback mechanism for molecular generation.
π― What it does: This paper proposes Domain-Inspired Sharpness-Aware Minimization (DISAM), which enhances convergence speed and generalization performance in multi-source domain learning with domain shift by introducing domain-level convergence consistency constraints in the perturbation generation of SAM.
π― What it does: This paper proposes a Motion Coherent Augmentation (MCA) for video understanding, which achieves hue perturbation in the RGB space through an efficient SwapMix operation and incorporates Variation Alignment (VA) to align predictions of different appearances of the same video, encouraging the model to focus on motion information rather than static appearance.
π― What it does: This study investigates the problem of generating low-density (minority) samples using diffusion models and proposes a controllable sampling framework.
Don't Trust: Verify -- Grounding LLM Quantitative Reasoning with Autoformalization
Jin Peng Zhou (Cornell University), Yuhuai Wu (xAI)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Utilizing large language models to automatically translate natural language math problems and their solutions into Isabelle form, and using automated theorem provers to verify and filter out correct answers.
π― What it does: A diversity external sample sampling strategy named DOS is designed and implemented to select external samples that can construct a global compact discriminative boundary during the training process, thereby enhancing the model's ability to detect out-of-distribution (OOD) data.
Doubly Robust Proximal Causal Learning for Continuous Treatments
Yong Wu (Fudan University), Xinwei Sun (Fudan University)
CodeGenerative Adversarial NetworkTabular
π― What it does: A kernel-based double robust estimator is proposed to estimate the causal effects of continuous treatments within the proximal causal learning framework.
DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer
Junyuan Hong (University of Texas at Austin), Zhangyang Wang (University of Chicago)
CodeClassificationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This work proposes Differentially-Private Offsite Prompt Tuning (DP-OPT), an end-to-end privacy-preserving framework that trains discrete prompts using private data locally and infers on cloud LLMs;
CodeOptimizationSafty and PrivacyConvolutional Neural NetworkImageTabular
π― What it does: Proposes Clipless DP-SGD, which utilizes Lipschitz constraints on parameters to provide an analytically upper bound on gradient norms, thereby eliminating the clipping process for each sample gradient in traditional DP-SGD.
CodeRetrievalTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: A framework named DQ-LoRe is proposed, which first allows the LLM to generate Chain-of-Thought (CoT), and then uses the CoT along with the question to query the retriever, automatically selecting examples used in In-Context Learning (ICL).
π― What it does: We propose DragonDiffusion, a method for drag-and-drop image editing that utilizes the correspondence of image features in pre-trained diffusion models, supporting tasks such as object movement, scaling, replacement, and pasting without the need for additional training.
π― What it does: Introducing DreamCraft3D: It first generates high-quality 2D images, then hierarchically sculpts geometry and enhances textures, ultimately producing globally consistent and detail-rich 3D objects.
DreamSmooth: Improving Model-based Reinforcement Learning via Reward Smoothing
Vint Lee (University of California), Youngwoon Lee (Yonsei University)
CodeReinforcement LearningWorld ModelSequential
π― What it does: This study investigates the challenge of reward prediction in model-based reinforcement learning and proposes improving the prediction performance of the reward model through temporal smoothing of the reward signal (DreamSmooth).
π― What it does: By using Dropout technology in neural networks to quickly explore the Rashomon set, we can efficiently estimate predictive multiplicity.
π― What it does: This paper proposes DRSM (De-Randomized Smoothed MalConv), which achieves provable robustness detection for malicious binary files through window ablation and de-randomized smoothing.
DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines
Omar Khattab (Stanford University), Christopher Potts (Stanford University)
CodeTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the DSPy programming model, defines signatures, modules, and teleprompters, and implements a compiler for automatic optimization of the LM pipeline; experiments are conducted on mathematical word problems (GSM8K) and multi-hop question answering (HotPotQA).
π― What it does: This study proposes an end-to-end multimodal 3D lane detection framework, DV-3DLane, which can simultaneously utilize camera images and LiDAR point clouds, and learn and fuse features in both perspective view (PV) and bird's eye view (BEV) spaces.
Hang Xu (Institute of Automation, Chinese Academy of Sciences), Jian Cheng (Institute of Automation, Chinese Academy of Sciences)
CodeOptimizationReinforcement LearningSequential
π― What it does: A dynamic discounting CFR framework DDCFR is proposed, which quickly approaches Nash equilibrium in uncertain information games using a learnable discount scheme.
π― What it does: Proposed and implemented a Dynamic Neural Response Tuning (DNRT) mechanism, which includes Response Adaptive Activation (RAA) and Aggregated Response Regularization (ARR), and embedded it into various ANN architectures, significantly improving model performance.
Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs
Yuxin Zhang (Xiamen University), Rongrong Ji (Xiamen University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A training-free sparse LLM fine-tuning method (DS/ban_circle T) is proposed, which enhances the performance of sparse models by iteratively adding and removing weights (prune-and-grow) on a pre-sparsified model, using only matrix multiplication to compute reconstruction error, without the need for backpropagation or weight updates.
Dynamics-Informed Protein Design with Structure Conditioning
Urszula Julia Komorowska (University of Cambridge), Mateja Jamnik (University of Cambridge)
CodeProtein Structure PredictionDiffusion modelScore-based ModelBiomedical Data
π― What it does: Incorporating kinetic constraints (based on the lowest non-trivial modes) into the protein generation process using diffusion models, and further achieving joint conditional generation of structure and dynamics.
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Proposed and implemented the DYVAL dynamic evaluation protocol, which dynamically generates inference task samples using directed acyclic graphs.
π― What it does: This paper presents EasyTPP, a unified open-source benchmark library for standardizing and transparently training, evaluating, and comparing temporal point process models.
π― What it does: A geometric deep learning framework called EBMDock based on an energy model is proposed, which describes the probability distribution of protein-protein docking using statistical potential energy and finds the lowest energy conformation through Langevin dynamics sampling.
π― What it does: This paper proposes DA-Fusion, a semantic enhancement method for image-to-image based on Stable Diffusion, which uses a pre-trained diffusion model to semantically edit real images and generate diverse data augmentation samples.
π― What it does: Refining large-scale web datasets through deduplication, CLIP-score filtering, and density-based pruning (DBP) based on concept complexity, significantly reducing computational costs and improving model performance in large-scale CLIP training.
π― What it does: Proposed and implemented a local curvature distribution (LCP) structure encoding based on discrete Ricci curvature, and used it as node features input into various graph neural networks.
π― What it does: A method for generating graphs through gradual local expansion (from a single node to a complete graph) is proposed, utilizing a denoising diffusion model to recover the edge and node features at each step, and introducing a local PPGN layer to maintain high expressive power.