International Conference on Learning Representations Β· 1064 papers
OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views
Francis Engelmann (ETH Zurich), Federico Tombari (Google)
CodeSegmentationVision Language ModelNeural Radiance FieldImage
π― What it does: Developed the OpenNeRF method, which utilizes neural radiance fields (NeRF) and pixel-level CLIP features to achieve open-set 3D semantic segmentation. It can generate continuous 3D representations from RGB images (with optional depth) with known camera poses and perform zero-shot segmentation for any text or image queries.
Krishna Acharya (Georgia Institute of Technology), Juba Ziani (Georgia Institute of Technology)
CodeOptimizationReinforcement LearningTabularBiomedical Data
π― What it does: In the online prediction problem, a new oracle-efficient algorithm is proposed for scenarios with intersecting subsequences (i.e., individuals belonging to multiple groups) to simultaneously ensure sublinear regret (groupwise regret) relative to the best model for all groups.
π― What it does: This paper proposes Order-Preserving GFlowNets (OP-GFNs), a method that samples candidate sets by learning a reward function consistent with a given (partial) ranking without the need for explicit scalar rewards, applicable to both single-objective and multi-objective optimization.
π― What it does: Using a pre-trained CLIP model, we learn positive and negative prompt words to construct a positive-negative classifier for more accurately detecting out-of-distribution (OOD) samples.
Out-of-Variable Generalisation for Discriminative Models
Siyuan Guo (Max Planck Institute for Intelligent Systems), Bernhard SchΓΆlkopf (Max Planck Institute for Intelligent Systems)
CodeDomain AdaptationTabularBiomedical Data
π― What it does: The study investigates the generalization of discriminative models when the variable sets of the source and target environments do not completely overlap, and proposes using the higher-order moments of residuals to identify the target prediction function.
CodeClassificationDomain AdaptationKnowledge DistillationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
π― What it does: This study investigates the overfitting defects of vision-language models in out-of-distribution (OOD) generalization and proposes the OGEN method, which synthesizes unknown class features through a class-conditional feature generator and implements adaptive self-distillation for regularization, enhancing the OOD generalization performance of the fine-tuned model.
π― What it does: This paper proposes a prompt-based replay-free class incremental learning method, combining Virtual Outlier Regularization (VOR) to tighten the classifier's decision boundary, and introducing a single prompt scheme (OnePrompt) to simplify the prompt pool, ultimately forming the OVOR method.
Hongcheng Guo (Beihang University), Zhoujun Li (Cloudwise Research)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: A large language model tailored for IT operations tasks, named OWL, has been developed, along with the OWL-Instruct instruction set and the OWL-Bench bilingual evaluation benchmark; the HMCE long-context extension method and the Mixture-of-Adapter parameter-efficient fine-tuning strategy have been proposed to enhance the model's performance in operational scenarios.
π― What it does: A deep imbalance clustering method based on progressive pseudo-labels is proposed, utilizing progressive partial optimal transport to generate high-quality pseudo-labels that are sensitive to class imbalance, thereby learning more robust representations and clustering.
π― What it does: A point-level supervised instance segmentation framework P2Seg is proposed, which utilizes a mutual distillation module (MDM) for bidirectional knowledge transfer between semantic and instance information, generating high-quality pseudo labels and training the instance segmentation network.
π― What it does: A new algorithm is proposed that enables parallel evaluation and training of nonlinear sequence models (such as RNNs and neural ordinary differential equations) without changing the architecture of the sequence model, significantly accelerating the training speed.
Parametric Augmentation for Time Series Contrastive Learning
Xu Zheng (Florida International University), Dongsheng Luo (Florida International University)
CodeClassificationOptimizationRepresentation LearningRecurrent Neural NetworkGenerative Adversarial NetworkContrastive LearningTime Series
π― What it does: This paper proposes an adaptive parameterized data augmentation framework for time series contrastive learning, called AutoTCL, which achieves information retention and view diversity through factorization and learns the optimal augmentation strategy during the pre-training phase.
π― What it does: A dynamic rebalancing method based on Pareto multi-objective optimization (PLOT) is proposed to address gradient conflicts and representation learning degradation in long-tail recognition.
π― What it does: This paper proposes a Particle Guidance framework that utilizes time-evolving joint potential energy to achieve non-independent and more diverse sampling in diffusion models.
Partitioning Message Passing for Graph Fraud Detection
Wei Zhuo (Shenzhen Campus of Sun Yat-sen University), Jia Chen (GrabTaxi Holdings Pte. Ltd.)
CodeAnomaly DetectionGraph Neural NetworkGraphFinance Related
π― What it does: This paper proposes the Partitioning Message Passing (PMP) framework, which utilizes neighbor label information to distinguish between similar and dissimilar neighbors. During the information aggregation phase, it employs different aggregation weights that are adaptively generated by the central node to address the issues of label imbalance and similarity/dissimilarity in graph fraud detection.
Path Choice Matters for Clear Attributions in Path Methods
Borui Zhang (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: A path selection-based explanation method called SAMP is proposed to address the issue of unclear explanations caused by path uncertainty in traditional path methods.
Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models
Hritik Bansal (University of California Los Angeles), Aditya Grover (University of California Los Angeles)
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper studies the impact of sparse feedback protocols (scoring and ranking) on the alignment and evaluation of large language models (LLMs), and systematically assesses the differences between the two protocols in terms of preference consistency and evaluation consistency.
PeFLL: Personalized Federated Learning by Learning to Learn
Jonathan Scott (Institute of Science and Technology Austria), Christoph H Lampert
CodeFederated LearningMeta LearningImage
π― What it does: We propose PeFLL, a personalized federated learning algorithm based on meta-learning, which can generate personalized models for any client through a single forward inference on the server side;
π― What it does: The PerceptionCLIP method is proposed, which first uses CLIP to infer the contextual attributes of images (such as background, orientation, etc.), and then uses these attributes as conditions for zero-shot classification of images.
π― What it does: This paper measures and predicts perceptual scales such as spatial frequency, direction, and bandwidth through differential scaling experiments, and explores the relationship between perceptual scales and image power spectra by combining it with Fisher information theory.
Periodicity Decoupling Framework for Long-term Series Forecasting
Tao Dai (Shenzhen University), Shu-Tao Xia (Tsinghua University)
CodeTransformerTime Series
π― What it does: A periodic decoupling framework (PDF) is proposed, which splits the original 1D time series into short-term and long-term 2D variation sequences, and then performs parallel modeling to achieve long-period forecasting.
Renrui Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeSegmentationImageVideo
π― What it does: By providing only a reference image and its mask, the Segment Anything Model (SAM) is personalized to automatically segment user-specified visual concepts.
π― What it does: A system called PF-LRM is proposed, which can simultaneously predict camera pose and 3D object shape (NeRF) from a very small number of uncalibrated images.
Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement
Linlu Qiu (Massachusetts Institute of Technology), Xiang Ren (University of Southern California)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Through the iterative hypothesis refinement (generation, selection, refinement rules) method, the system evaluates the capabilities of language models in inductive reasoning tasks;
PhyloGFN: Phylogenetic inference with generative flow networks
Ming Yang Zhou, Yoshua Bengio (University of Montreal)
CodeTransformerFlow-based ModelSequential
π― What it does: We propose PhyloGFN, which utilizes Generative Flow Networks (GFlowNet) to simultaneously infer tree topology and branch lengths in the full tree space, achieving both Bayesian and minimum parsimony phylogenetic inference.
PINNACLE: PINN Adaptive ColLocation and Experimental points selection
Gregory Kang Ruey Lau (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeTime SeriesPhysics Related
π― What it does: The PINNACLE algorithm is proposed, which jointly optimizes the selection of all types of training points (experimental points, PDE collocation points, IC/BC collocation points) in the Physics-Informed Neural Network (PINN) and automatically adjusts the ratio of collocation points based on training progress.
PINNsFormer: A Transformer-Based Framework For Physics-Informed Neural Networks
Zhiyuan Zhao (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)
CodeTransformerTime SeriesSequentialPhysics Related
π― What it does: This paper presents PINNsFormer, a Transformer-based framework designed to capture temporal dependencies and approximate PDE analytical solutions within Physics-Informed Neural Networks.
PlaSma: Procedural Knowledge Models for Language-based Planning and Re-Planning
Faeze Brahman (Allen Institute for Artificial Intelligence), Yejin Choi (Allen Institute for Artificial Intelligence)
CodeOptimizationExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelText
π― What it does: This paper proposes the PLASMA framework, which utilizes small language models to achieve interpretable and high-quality programmatic planning and replanning through symbolic program knowledge distillation and verification-guided decoding during inference.
Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents
Yang Deng (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: A pluggable dialogue strategy planning plugin (PPDPP) is proposed, enabling large language models to achieve better goal attainment in active dialogue tasks through learning strategies.
Plug-and-Play: An Efficient Post-training Pruning Method for Large Language Models
Yingtao Zhang (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a post-training pruning framework that can directly perform one-time pruning on large language models, balancing model compression and hardware acceleration.
π― What it does: PnP Inversion is proposed, which decouples the source and target branches of the diffusion model, allowing for improved image editing quality with just three lines of code.
π― What it does: A novel unsupervised deep learning framework called Point2SSM is proposed, which can directly generate corresponding statistical shape models from raw point clouds.
π― What it does: This paper proposes a clean-label backdoor attack framework for face forgery detection models called Poisoned Forgery Face, which can implant a backdoor during the training phase and induce the model to incorrectly identify forged faces as real faces through a trigger.
PolyVoice: Language Models for Speech to Speech Translation
Qian qian Dong, Yuxuan Wang (ByteDance)
CodeTransformerLarge Language ModelPrompt EngineeringChain-of-ThoughtAudio
π― What it does: Proposes the PolyVoice framework, which uses three decoder-only language models to achieve speech-to-speech translation and supports non-written languages.
π― What it does: This paper studies how to constrain invariance/covariance in the latent space through a category theory framework when pooling multi-source medical imaging data, in order to address covariate shift and imbalance issues.
PoSE: Efficient Context Window Extension of LLMs via Positional Skip-wise Training
Dawei Zhu (Peking University), Sujian Li (Peking University)
CodeRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By implementing a jump offset on position indices during training, the PoSE method achieves the goal of extending the LLM context window to hundreds of thousands or even infinite lengths while maintaining the original context window's training length.
π― What it does: This paper proposes a post-processing method called MBS based on instance-level bias scores to achieve multiple group fairness constraints (DP, EOp, EO, and their combinations) without retraining the model, and provides analytical improvement rules for the Bayesian optimal classifier.
π― What it does: A Grounded Point Colorization (GPC) pre-training framework is proposed, which utilizes the corresponding colored images of point clouds for colorization learning of LiDAR 3D detectors, enhancing their perception of semantic structures.
π― What it does: A Random Orthogonal Projection Image Modeling (ROPIM) framework is proposed for self-supervised pre-training of Vision Transformers, replacing traditional binary masking methods.
Pre-training with Synthetic Data Helps Offline Reinforcement Learning
Zecheng Wang (New York University), Keith W. Ross (New York University)
CodeTransformerReinforcement LearningTabular
π― What it does: This study investigates the pre-training of offline reinforcement learning models (Decision Transformer and CQL) using simple synthetic data, demonstrating that this method can enhance performance, even surpassing pre-training with large language corpora.
π― What it does: A new class-incremental learning method called Prediction Error Classification (PEC) is proposed, which trains a student network for each class to fit the output of a random teacher network, using the student's prediction error against the teacher as the class score.
Junwei Su (University of Hong Kong), Chuan Wu (University of Hong Kong)
CodeGraph Neural NetworkGraphTime Series
π― What it does: This study addresses and solves the performance degradation issue caused by temporal discontinuity in Memory-based Dynamic Graph Neural Networks (MDGNN) as the temporal batch size increases, proposing a training framework named PRES (Predict-to-Smooth);
Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting
Enyi Jiang (University of Illinois Urbana-Champaign), Sanmi Koyejo (Stanford University)
CodeDomain AdaptationFederated LearningImage
π― What it does: A novel aggregation rule for domain adaptation between source clients and target clients in federated learning is proposed, and performance is further enhanced through adaptive weighting.
CodeGenerationSafty and PrivacyLarge Language ModelText
π― What it does: A differential privacy-based in-context learning (DP-ICL) framework is proposed and implemented, utilizing chunked example sets for parallel inference and noise aggregation to achieve privacy protection in text classification and language generation tasks.
Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation
Xinyu Tang (Princeton University), Robert Sim (Microsoft Research)
CodeGenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes an algorithm for synthesizing a small number of examples (few-shot) from private datasets while maintaining differential privacy (DP) guarantees, and using them as demonstrations for in-context learning (ICL) under large language models (LLM);
π― What it does: The ProSMin method is proposed, which minimizes the dimensional collapse of representations and enhances representation quality during the pre-training phase by using a loss function based on appropriate scoring rules through probability knowledge distillation between the online network and the target network.
Procedural Fairness Through Decoupling Objectionable Data Generating Components
Zeyu Tang (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
CodeTabular
π― What it does: A framework for program fairness based on causal graphs is proposed, which decouples objectionable causal paths from neutral paths during the data generation process using 'reference points' and 'value instantiation rules', predicting only with neutral paths to avoid hidden impacts of procedural unfairness.
π― What it does: Proposed the Progressive Fourier Neural Representation (PFNR), a sequential video compilation method that utilizes learnable sparse subnetworks to encode videos in Fourier space, avoiding catastrophic forgetting.
π― What it does: This paper proposes Prompt Gradient Projection (PGP), which combines prompt-tuning with gradient projection to suppress catastrophic forgetting in continual learning without task identifiers.
Boya Shi (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
CodeClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: A prompt learning framework utilizing Quaternion Networks (QNet) is proposed for fusing visual and textual modality features in zero-shot scenarios and generating high-quality prompts across multiple tasks.
Prompt Risk Control: A Rigorous Framework for Responsible Deployment of Large Language Models
Thomas P Zollo, Richard Zemel (Columbia University)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposes a Prompt Risk Control framework that utilizes distribution-free uncertainty quantification methods to provide a high-probability risk upper bound for prompt selection in large language models.
PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization
Xinyuan Wang (University of California San Diego), Zhiting Hu
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBiomedical Data
π― What it does: This paper presents PromptAgent, a framework that views prompt engineering as a strategic planning problem, generating expert-level prompts through MCTS and LLM self-reflection, which can automatically improve performance across multiple tasks.
PromptTTS 2: Describing and Generating Voices with Text Prompt
Yichong Leng (University of Science and Technology of China), Jiang Bian (Microsoft)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextAudio
π― What it does: Proposes PromptTTS 2, which uses a differential network and LLM to generate text prompts, addressing the one-to-many and data scale challenges of text prompt TTS;
Diego Gomez (University of Alberta), Marlos C. Machado (University of Alberta)
CodeOptimizationRepresentation LearningGraph
π― What it does: This paper proposes a new max-min objective function (ALLO) that achieves hyperparameter-free Laplacian representation learning through stop-gradient and augmented Lagrangian techniques, capable of simultaneously recovering the minimum eigenvector and corresponding eigenvalue of the Laplacian operator.
Protein Discovery with Discrete Walk-Jump Sampling
Nathan C. Frey (Genentech), Saeed Saremi (Genentech)
CodeGenerationDrug DiscoveryDiffusion modelScore-based ModelSequentialBiomedical Data
π― What it does: A new discrete generative model framework is proposedβDiscrete Walk-Jump Sampling (dWJS), which achieves high-quality and fast discrete sequence generation by training an energy model in a noise-smoothed space and returning to the original discrete data using least squares estimation.
π― What it does: A multi-assembly structure prediction framework based on prompt learning, PROMPTMSP, is proposed, which utilizes pre-training and prompt fine-tuning to achieve the stepwise assembly of multi-chain proteins.
Protein-Ligand Interaction Prior for Binding-aware 3D Molecule Diffusion Models
Zhilin Huang (Tsinghua University), Wenming Yang
CodeDrug DiscoveryDiffusion modelBiomedical Data
π― What it does: A 3D molecular diffusion model named IPDIFF is proposed, incorporating protein-ligand interaction priors in both forward and reverse diffusion processes to achieve molecule generation targeted at specific targets.
Prototypical Information Bottlenecking and Disentangling for Multimodal Cancer Survival Prediction
Yilan Zhang (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
CodeClassificationData-Centric LearningTransformerMultimodalityBiomedical Data
π― What it does: For multimodal cancer survival prediction, we propose the Prototypical Information Bottleneck and Prototypical Information Disentanglement (PIBD) framework, which first eliminates unimodal redundancy through the prototypical information bottleneck, and then utilizes prototype distribution to drive the decoupling of shared and specific information, thereby enhancing prediction performance.
π― What it does: This paper proposes a reinforcement learning exploration algorithm LMC-LSVI that directly samples the posterior distribution of the Q function using Langevin Monte Carlo (LMC), and extends it to deep Q networks (Adam LMCDQN), achieving scalable and theoretically provable efficient exploration.
Provably Robust Conformal Prediction with Improved Efficiency
Ge Yan (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
CodeClassificationComputational EfficiencyImage
π― What it does: A provably robust framework for synthetic prediction, RSCP+, is proposed, along with two methods: PTT, which is lossless after training, and RCT, which is robust for synthetic training, addressing the robustness proof flaws and inefficiencies of the original RSCP.
π― What it does: This study investigates a statistical method that utilizes exchangeability in black-box language models to test whether the test set is contaminated by pre-training data through log probability ratio comparison.
π― What it does: A pseudo-general method is proposed to achieve dynamic scene synthesis from a new viewpoint using a monocular video without scene-specific appearance optimization.
QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models
Yuhui Xu (Huawei Inc.), Qi Tian (Huawei Inc.)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A Quantized Perception Low-Rank Adaptation (QA-LoRA) method is proposed, balancing the parameter-efficient fine-tuning of large language models and low-bit quantization deployment.
QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models
Jing Liu (Monash University), Bohan Zhuang (SenseTime Research)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A low-bit LLM compression method based on post-training quantization, called QLLM, is proposed. It suppresses abnormal channels by splitting and merging activation values, and corrects quantization errors using low-rank parameter fine-tuning. It can complete 4-bit quantization of a 70B model in 10 hours on a single A100 80G GPU.
Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting
Melanie Sclar (University of Washington), Alane Suhr (University of California)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This study investigates the performance fluctuations of large language models under different equivalent prompt formats and proposes a rapid evaluation method called FORMATSPREAD.
Quantifying the Plausibility of Context Reliance in Neural Machine Translation
Gabriele Sarti (University of Groningen), Arianna Bisazza (University of Groningen)
CodeGenerationExplainability and InterpretabilityTransformerText
π― What it does: Proposes the PECORE framework for identifying context-sensitive words in generated text from neural machine translation and attributing the contextual cues that lead to their generation, thereby quantifying the contextual dependency interpretability of the model.
π― What it does: This paper proposes quasi-Monte Carlo (QMC) slice Wasserstein (QSW) and randomized quasi-Monte Carlo (RQSW) estimation methods based on spherical low-discrepancy point sets to improve the traditional Monte Carlo (MC) slice Wasserstein (SW) distance computation and gradient estimation.
Query-Dependent Prompt Evaluation and Optimization with Offline Inverse RL
Hao Sun (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringText
π― What it does: A Prompt-OIRL method based on offline inverse reinforcement learning is proposed, which utilizes offline prompt demonstration data to train a reward model, achieving query-dependent prompt evaluation and optimization during inference, significantly enhancing arithmetic reasoning performance.
Query-Policy Misalignment in Preference-Based Reinforcement Learning
Xiao Hu (Tsinghua University), Ya-Qin Zhang (Tsinghua University)
CodeReinforcement LearningTime SeriesBenchmark
π― What it does: This paper studies the query-policy misalignment issue in preference-based reinforcement learning (PbRL) and proposes an improved framework based on policy alignment for query and mixed experience replay (QPA) to enhance feedback efficiency and sampling efficiency.
π― What it does: The Least Disagree Metric (LDM) is proposed to measure the uncertainty of deep models on unlabeled samples, and based on this, an active learning algorithm LDM-S is designed.
π― What it does: A method named Quick-Tune has been designed and implemented, which can quickly select the best pre-trained model and its fine-tuning hyperparameters simultaneously from a given model hub.
π― What it does: This paper proposes a region-based masked autoencoding task (RAE) and combines it with MAE to form R-MAE for unsupervised image representation learning and detection/segmentation tasks.
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: This paper proposes a lightweight retrieval-augmented dual instruction tuning (RA-DIT) method, which first fine-tunes a language model to better utilize retrieval information, and then performs LM-supervised retrieval fine-tuning on the retriever to transform any pre-trained LLM into a retrieval-augmented model.
Chengzhi Mao (Columbia University), Junfeng Yang (Columbia University)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Utilizing large language models (LLMs) to perform rewriting tasks on the same text, calculating the edit distance (Levenshtein distance) and bag-of-words similarity metrics between the original and rewritten texts as features for a binary classifier to determine whether the text is AI-generated.
RAIN: Your Language Models Can Align Themselves without Finetuning
Yuhui Li (Peking University), Hongyang Zhang (University of Waterloo)
CodeGenerationTransformerLarge Language ModelReinforcement LearningText
π― What it does: By incorporating a self-assessment and backtracking mechanism during the inference phase, the RAIN method enables frozen large language models to achieve self-alignment without fine-tuning or additional data.
RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
Parth Sarthi (Stanford University), Christopher D Manning
CodeRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: This paper proposes and implements a recursive abstract tree structure retrieval system called RAPTOR, which can first chunk, embed, and cluster long texts, and then recursively generate a summary tree. During retrieval, information can be obtained at different levels of abstraction, thereby better supporting multi-step reasoning and understanding of long texts.
π― What it does: A theoretical framework based on distribution matching is proposed to guide the synthesis of training data, and under this framework, Stable Diffusion is fine-tuned to generate synthetic training sets suitable for image classification.
Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning
LINHAO LUO, Shirui Pan (Griffith University)
CodeExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextGraphRetrieval-Augmented Generation
π― What it does: This paper proposes a reasoning framework called RoG that combines large language models with knowledge graphs, implementing trustworthy and interpretable KG question answering through a three-stage process of planning, retrieval, and reasoning.
π― What it does: A retrieval-reconstruction-based time series contrastive learning method called REBAR is proposed, which constructs positive and negative samples by learning the reconstruction error between sequences.
Reclaiming the Source of Programmatic Policies: Programmatic versus Latent Spaces
Tales Henrique Carvalho, Levi Lelis
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderSequential
π― What it does: This study investigates the effects of using the raw DSL space and the learned latent space for local search in programmatic strategy synthesis.
π― What it does: This paper presents RECOMBINER, a compression method based on Bayesian Implicit Neural Representation (INR), which enhances compression quality through linear reparameterization, learnable positional information encoding, and hierarchical priors.
RECOMP: Improving Retrieval-Augmented LMs with Context Compression and Selective Augmentation
Fangyuan Xu (University of Texas at Austin), Eunsol Choi (University of Washington)
CodeRetrievalCompressionComputational EfficiencyTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
π― What it does: The RECOMP method is proposed, which first compresses the retrieved long texts into brief summaries in the retrieval-augmented language model, and then uses the summary as a prefix input to the language model, thereby improving inference efficiency while maintaining or enhancing task performance.
Reconciling Spatial and Temporal Abstractions for Goal Representation
Mehdi Zadem (Γcole Polytechnique), Sao Mai Nguyen (IMT Atlantique)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: This paper proposes a three-layer hierarchical reinforcement learning algorithm called STAR, which utilizes spatial and temporal abstraction to jointly construct goal representations, thereby improving the learning efficiency of continuous control tasks.
π― What it does: This paper proposes a Recursive Generalized Transformer (RGT) for image super-resolution, aiming to efficiently capture global spatial information at high resolutions and improve reconstruction quality.
REFACTOR: Learning to Extract Theorems from Proofs
Jin Peng Zhou (Cornell University), Roger Baker Grosse
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes a neural network model named REFACTOR, which is used to automatically extract reusable theorems (lemmas) from formal mathematical proof trees, and implements extraction, reconstruction, and proof tasks in the Metamath library.
π― What it does: ReFusion is proposed, a retrieval-enhanced framework for non-knowledge-intensive tasks that directly integrates retrieval representations into the model's hidden layers, using a bi-level optimization approach to find the optimal fusion structure.
Yilong Xu (Renmin University of China), Hao Sun (Renmin University of China)
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper proposes a symbolic regression framework RSRM that combines Monte Carlo Tree Search, Double Q-Learning, and Modular Subtree Discovery to efficiently search for and discover complex mathematical equations with limited data.
π― What it does: A distributed Bayesian optimization algorithm named DuMBO is proposed, which can infer and utilize arbitrarily complex additive decompositions to optimize high-dimensional noisy and expensive black-box functions without restricting the maximum factor size (MFS) of the additive structure.
Relay Diffusion: Unifying diffusion process across resolutions for image synthesis
Jiayan Teng (Tsinghua University), Jie Tang (Tsinghua University)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: The Relay Diffusion Model (RDM) is proposed, which seamlessly extends low-resolution generation results to high resolution through block noise and block-level fuzzy diffusion, addressing the issue of noise scheduling and low-resolution conditions mismatch in traditional cascaded methods.
π― What it does: This paper proposes the InfoCORE method to eliminate batch effects and improve representation quality in multimodal molecular representation learning.
π― What it does: This paper proposes a method called 'Repeated Random Sampling' (RS2), which randomly samples subsets multiple times during each training epoch, and demonstrates that it outperforms existing complex data pruning and data distillation techniques in terms of time-to-accuracy.
Isaac Reid (University of Cambridge), Adrian Weller (University of Cambridge)
CodeGraph Neural NetworkGraph
π― What it does: A repelling random walks mechanism is proposed, which utilizes the mutual repulsion relationship between multiple walk trajectories to enhance the efficiency of random walks on graphs by reducing the estimator variance while keeping the edge transition probabilities unchanged.
Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models
Archiki Prasad (University of North Carolina), Mohit Bansal (University of North Carolina)
CodeRecognitionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: A gradient-free, instance-level framework called REPARE is proposed, which rewrites and supplements visual question answering (VQA) problems to enhance their visual relevance and clarity, thereby improving zero-shot VQA performance.
RepoBench: Benchmarking Repository-Level Code Auto-Completion Systems
Tianyang Liu (University of California San Diego), Julian McAuley (University of California San Diego)
CodeRetrievalAI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes RepoBench, a benchmarking system for code autocompletion in multi-file projects, covering three main tasks: retrieval (RepoBench-R), completion (RepoBench-C), and the complete pipeline (RepoBench-P), supporting both Python and Java languages.
Representation Deficiency in Masked Language Modeling
Yu Meng (University of Illinois Urbana Champaign), Luke Zettlemoyer (Meta AI)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
π― What it does: This study investigates how the MASK symbol in Masked Language Modeling pre-training leads to model dimensions being specifically occupied by MASK, resulting in defects in the representation of real words, and proposes MAE-LM to address this issue by excluding MASK in the encoder.
ResFields: Residual Neural Fields for Spatiotemporal Signals
Marko Mihajlovic (ETH Zurich), Siyu Tang (Microsoft)
CodeNeural Radiance FieldVideo
π― What it does: The study introduces a temporal residual layer (ResFields) into the neural field model to enhance its ability to model long and complex sequences.