International Conference on Learning Representations · 1064 papers
Be Aware of the Neighborhood Effect: Modeling Selection Bias under Interference
Haoxuan Li (Peking University), Xiangnan He (University of Science and Technology of China)
CodeRecommendation SystemTabular
🎯 What it does: This paper studies the issue of selection bias in recommendation systems when neighborhood effects are present, proposing a neighborhood intervention representation framework based on causal inference and providing the corresponding ideal loss function.
Be Careful What You Smooth For: Label Smoothing Can Be a Privacy Shield but Also a Catalyst for Model Inversion Attacks
Lukas Struppek (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
CodeSafty and PrivacyAdversarial AttackImage
🎯 What it does: This study investigates the impact of label smoothing (positive and negative factors) on model inversion attacks (MIA) in deep learning models and proposes negative label smoothing as a defense strategy.
Beam Enumeration: Probabilistic Explainability For Sample Efficient Self-conditioned Molecular Design
Jeff Guo (Institut des Sciences et Ingénierie Chimiques National Centre of Competence in Research Catalysis), Philippe Schwaller (Institut des Sciences et Ingénierie Chimiques National Centre of Competence in Research Catalysis)
CodeExplainability and InterpretabilityDrug DiscoveryLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: This paper proposes the Beam Enumeration method, which utilizes high-probability token subsequence enumeration generated by a language model to extract molecular substructures and perform self-conditioning generation, thereby significantly improving sample efficiency and providing interpretability.
🎯 What it does: This paper proposes a Bayesian-enhanced pessimistic Q-learning method that uniformly uses worst-case states during both training and testing phases, thereby improving resistance to robust state perturbation attacks.
🎯 What it does: For the pre-trained flow matching generative model, we propose the BOSS (Bellman Optimal Stepsize Straightening) method, which first uses dynamic programming to obtain the optimal sampling step size, and then straightens the velocity network to achieve low NFE high-quality image generation.
🎯 What it does: A complete benchmark framework for Federated Domain Generalization is proposed, along with a new controllable heterogeneous data partitioning method, systematically evaluating the performance of 14 methods on 7 multi-domain datasets.
Benchmarking and Improving Generator-Validator Consistency of Language Models
Xiang Lisa Li (Stanford University), Percy Liang (Stanford University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper designs a Generative-Validation Consistency (GV-consistency) metric to systematically evaluate the consistency of language models in both generative and validation queries, and proposes a label-free consistency fine-tuning method to enhance consistency as well as the quality of generation and validation.
🎯 What it does: This paper proposes the BEND benchmark, which systematically evaluates the performance of various DNA language models on seven biologically significant tasks.
BENO: Boundary-embedded Neural Operators for Elliptic PDEs
Haixin Wang (National Engineering Research Center for Software Engineering Peking University), Tailin Wu (Westlake University)
CodeGraph Neural NetworkTransformerMeshPhysics Related
🎯 What it does: A new neural operator architecture called BENO is proposed to solve elliptic partial differential equations (Poisson/Laplace) with complex geometries and inhomogeneous boundary conditions.
Better Neural PDE Solvers Through Data-Free Mesh Movers
Peiyan Hu (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Zhi-Ming Ma (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
🎯 What it does: An unsupervised 'Data-Driven Mesh Mover (DMM)' is proposed to learn adaptive meshes and build a Moving Mesh Neural PDE Solver (MM-PDE) based on it, achieving high-precision simulations of dynamic systems.
Beyond Accuracy: Evaluating Self-Consistency of Code Large Language Models with IdentityChain
Marcus J. Min (Columbia University), Baishakhi Ray (Columbia University)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a new framework called IdentityChain, which simultaneously evaluates the accuracy and self-consistency of code large language models (Code LLM) in natural language to program (NL-to-PL) and program to natural language (PL-to-NL) tasks, and quantifies self-consistency through a new metric TOM (Test Output Match);
Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment
Geyang Guo (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method called FIGA is proposed, which aligns large language models directly under conditions without reinforcement learning by utilizing fine-grained comparative signals between initial low-quality responses and their improved versions.
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
Chaoqi Wang (University of Chicago), Yuxin Chen (University of Chicago)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a generalized direct preference optimization framework f-DPO, which allows for alignment fine-tuning of LLMs under various f-divergence constraints;
Beyond Spatio-Temporal Representations: Evolving Fourier Transform for Temporal Graphs
Anson Bastos (HERE Technologies), Toyotaro Suzumura (University of Tokyo)
CodeRecommendation SystemComputational EfficiencyGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: This paper proposes a reversible spectral transformation called the Evolving Graph Fourier Transform (EFT), which can simultaneously map dynamic graphs to the frequency domain in both the time domain and node domain, capturing the spectral features of the graph structure as it evolves over time.
Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning
Mustafa Shukor (Sorbonne University), Matthieu Cord (Valeo)
CodeExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Evaluate and mitigate the shortcomings of large-scale multimodal models (LMMs), proposing various untrained in-context learning (ICL) variants to enhance the model's capabilities in interpretability, refusal, compositional reasoning, and instruction following.
Beyond Weisfeiler-Lehman: A Quantitative Framework for GNN Expressiveness
Bohang Zhang (Peking University), Liwei Wang (Peking University)
CodeGraph Neural NetworkGraph
🎯 What it does: A quantitative framework for the expressive power of graph neural networks based on isomorphic mapping counting is proposed, defining isomorphic expressiveness and providing a complete description of various mainstream GNN models.
🎯 What it does: The study investigates the robustness of reinforcement learning in adversarial attack environments, proposing the construction of a finite non-dominated policy set during the training phase and adaptive defense through online no-regret learning during the testing phase.
BioBridge: Bridging Biomedical Foundation Models via Knowledge Graphs
Zifeng Wang (University of Illinois Urbana-Champaign), RISHITA ANUBHAI
CodeGenerationRetrievalDrug DiscoveryTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes the BioBRIDGE framework, which learns cross-modal conversion modules through knowledge graphs, achieving multi-modal retrieval and generation without fine-tuning the base foundational models (FMs).
🎯 What it does: This paper proposes a two-stage unsupervised graph anomaly detection framework called GADAM. It first uses MLP for contrastive learning to obtain local anomaly scores based on local inconsistencies, and then generates global anomaly scores through adaptive message passing with mixed attention and global consistency discrimination, ultimately fusing the scores from both stages.
Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models
Sijia Chen (University of Toronto), Di Niu (University of Alberta)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The Boosting of Thoughts (BoT) framework is proposed, utilizing LLM to generate, aggregate, and evaluate tree-like thinking during the iterative process, and continuously incorporating error analysis and suggestions as experience into the prompts, thereby achieving complex mathematical problem solving without human annotations.
🎯 What it does: This paper proposes a method for adversarial training based on OOD detection called GOOD-AT, aimed at enhancing the robustness of Graph Neural Networks (GNNs) against graph structure attacks.
CodeClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes an interpretable image classification framework called Concept-QA+V-IP, which automatically generates semantic queries using the large language model GPT and trains a lightweight Concept-QA network to utilize pseudo-labels from CLIP and GPT to answer these queries, thereby achieving information tracking and interpretation in V-IP without the need for manual annotation.
Bounding Box Stability against Feature Dropout Reflects Detector Generalization across Environments
Yang Yang (Australian National University), Liang Zheng (Australian National University)
CodeObject DetectionAutonomous DrivingImage
🎯 What it does: A detection box stability score (BoS) based on feature map masking is proposed for label-free evaluation of the accuracy of object detection models.
Bounding the Expected Robustness of Graph Neural Networks Subject to Node Feature Attacks
Yassine ABBAHADDOU, Henrik Boström (KTH Stockholm)
CodeAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper studies the robustness of Graph Neural Networks (GNN) against node feature attacks, proposes the concept of expected robustness and provides theoretical upper bounds, designs the GCORN model based on orthogonalized weights to enhance robustness, and presents an attack-independent probability assessment method.
🎯 What it does: This paper studies the representation-induced confounding bias (RICB) that may arise when using low-dimensional representation learning to estimate the conditional average treatment effect (CATE) in causal inference. It proposes a neural rebuttal framework that is independent of specific representation learning methods to estimate the upper and lower bounds of RICB, thereby enhancing the reliability of decision-making.
Breaking Physical and Linguistic Borders: Multilingual Federated Prompt Tuning for Low-Resource Languages
Wanru Zhao (University of Cambridge), Nicholas Donald Lane
CodeFederated LearningSafty and PrivacyTransformerPrompt EngineeringText
🎯 What it does: A multi-language federated prompt tuning framework is proposed, which efficiently fine-tunes parameters for low-resource languages without sharing data.
🎯 What it does: Proposes an unsupervised Transitional Dictionary Learning (TDL) framework that utilizes diffusion decomposition of visual input and online prototype clustering to automatically extract interpretable visual components and their relationships from images, represented in a dictionary form.
Bridging Vision and Language Spaces with Assignment Prediction
Jungin Park (NAVER AI Lab), Kwanghoon Sohn (Korea Institute of Science and Technology)
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Using a single-layer linear mapping, the embedding space of the pre-trained visual model is mapped to the word embedding space of the LLM, and the alignment of visual and language representations is achieved through optimal transport allocation prediction, enabling the frozen LLM to understand visual inputs and generate text.
🎯 What it does: This paper proposes the BTR (Binary Token Representations) technology, which uses a 1-bit vector to precompute token representations for retrieving paragraphs, significantly improving the inference speed and storage efficiency of Retrieval-Augmented Language Models.
🎯 What it does: This paper proposes a context-aware multi-instance learning framework CAMIL that combines neighborhood-constrained attention with Nystromformer for cancer detection and subtype classification in whole slide images (WSI).
Canyu Chen (Illinois Institute of Technology), Kai Shu (Illinois Institute of Technology)
CodeClassificationGenerationTransformerLarge Language ModelText
🎯 What it does: This study investigates the difficulty of detecting misinformation generated by LLMs (such as ChatGPT) for both humans and machine detectors, constructs a five-dimensional classification framework for LLM misinformation, and validates three types of generation methods (Hallucination, Arbitrary, Controllable), generating the LLMFake dataset for empirical evaluation.
Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs
Miao Xiong (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the self-confidence expression methods of black-box LLMs, proposing a three-component framework (prompt, sampling, aggregation) and conducting systematic evaluations across multiple models and tasks.
Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attacks
Vaidehi Patil (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: In pre-trained language models, an attack and defense framework is proposed for the direct deletion and extraction defense of sensitive information, exploring 'complete deletion' through model editing and verifying the deletion effect.
CodeGenerationAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a method to bypass the alignment of open-source LLMs through changes in decoding configurations (removing system prompts, adjusting temperature/TopK/TopP, etc.) and, based on this, introduces a multi-decoding strategy for a generation-aware alignment method for defense.
🎯 What it does: This study investigates the deep information bottleneck for regression tasks and proposes using Cauchy-Schwarz divergence as a substitute for KL/MSE for prediction and compression.
🎯 What it does: A deep learning model named Causal-StoNet is proposed, which jointly estimates the average causal effect, propensity score, and outcome function, and can handle high-dimensional nonlinear data and missing values.
Causality-Inspired Spatial-Temporal Explanations for Dynamic Graph Neural Networks
Kesen Zhao (City University of Hong Kong), Liang Zhang (Shenzhen Research Institute of Big Data)
CodeExplainability and InterpretabilityGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: This paper proposes a causal heuristic spatiotemporal interpreter for dynamic graph neural networks, DyGNNExplainer, which can generate subgraphs that conform to causal relationships without compromising the model.
CellPLM: Pre-training of Cell Language Model Beyond Single Cells
Hongzhi Wen (Michigan State University), Jiliang Tang (Michigan State University)
CodeTransformerBiomedical Data
🎯 What it does: A cell language model called CellPLM based on Transformer is proposed, which can learn intercellular relationships through pre-training and generate high-quality cell embeddings.
Certified Adversarial Robustness for Rate Encoded Spiking Neural Networks
Bhaskar Mukhoty (Mohamed bin Zayed University of Artificial Intelligence), Bin Gu (Jilin University)
CodeAdversarial AttackSpiking Neural NetworkImage
🎯 What it does: This paper studies the robustness of spike neural networks (SNN) using rate coding and provides a provable robustness proof under the l1-norm based on Bernoulli random smoothing.
Chain of Hindsight aligns Language Models with Feedback
Hao Liu (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A novel method called 'Chain of Hindsight (CoH)' is proposed, which fine-tunes large language models by converting all human feedback into natural language sequences and using them as contextual input, enabling the model to learn and align with both positive and negative feedback.
Chain-of-Experts: When LLMs Meet Complex Operations Research Problems
Ziyang Xiao (Zhejiang University), Gang Chen (Huawei Noah's Ark Lab)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextChain-of-Thought
🎯 What it does: The Chain-of-Experts (CoE) framework is proposed, utilizing multiple expert LLMs for collaborative reasoning to automatically complete the modeling and program generation of complex operations research (OR) problems.
Chain-of-Knowledge: Grounding Large Language Models via Dynamic Knowledge Adapting over Heterogeneous Sources
Xingxuan Li (Alibaba Group), Lidong Bing (Alibaba Group)
CodeTransformerLarge Language ModelTextTabularBiomedical DataPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Chain-of-Knowledge (CoK) framework, which enhances the factual accuracy of large language models through a three-stage dynamic knowledge adaptation, utilizing multi-source (structured and unstructured) knowledge.
Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding
Zilong Wang (University of California), Tomas Pfister (Google Cloud AI Research)
CodeTransformerLarge Language ModelPrompt EngineeringTabularBenchmarkChain-of-Thought
🎯 What it does: The CHAIN-OF-TABLE framework is proposed, which utilizes large language models to gradually perform table operations during the reasoning process, forming a table reasoning chain.
Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words
Yujia Bao (Accenture), Theofanis Karaletsos (Chan Zuckerberg Initiative)
CodeClassificationRecognitionTransformerImage
🎯 What it does: Proposes ChannelViT, an improved Vision Transformer to handle multi-channel images, and enhances robustness using hierarchical channel sampling.
Circuit Component Reuse Across Tasks in Transformer Language Models
Jack Merullo (Brown University), Ellie Pavlick (Brown University)
CodeTransformerLarge Language ModelText
🎯 What it does: A study on circuit reuse in the GPT-2 model demonstrates that different tasks (IOI and Colored Objects) share approximately 78% of attention heads, and interventions improve the accuracy of the Colored Objects task.
🎯 What it does: This study investigates seven methods of concept erasure in text-to-image models and designs a Concept Inversion (CI) attack, demonstrating that these methods cannot completely remove the target concept.
CivRealm: A Learning and Reasoning Odyssey in Civilization for Decision-Making Agents
Siyuan Qi (National Key Laboratory of General Artificial Intelligence, BIGAI), Song-Chun Zhu (National Key Laboratory of General Artificial Intelligence, BIGAI)
CodeRecurrent Neural NetworkTransformerLarge Language ModelReinforcement Learning
🎯 What it does: In this paper, the authors propose CivRealm—an interactive environment based on the Civilization game, designed to simultaneously test the learning and reasoning abilities of agents, and provide two types of interfaces: tensor-based RL and LLM-based;
CLAP: Collaborative Adaptation for Patchwork Learning
Sen Cui (Tsinghua University), Fei Wang (Cornell University)
CodeFederated LearningRepresentation LearningAuto EncoderMultimodalityBiomedical Data
🎯 What it does: A Patchwork Learning framework (CLAP) is designed to complete multimodal missing data imputation and representation learning under the condition that different clients have different combinations of modalities and do not share raw data.
Cleanba: A Reproducible and Efficient Distributed Reinforcement Learning Platform
Shengyi Huang (Drexel University), Santiago Ontanon (Hugging Face)
CodeReinforcement LearningImage
🎯 What it does: A distributed reinforcement learning platform called Cleanba is proposed, addressing the reproducibility issues in traditional systems like IMPALA, and achieving reproducible, efficient, and scalable distributed PPO/IMPALA training on 57 Atari games.
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkOrdinary Differential Equation
🎯 What it does: The CLEX method is proposed, which utilizes neural ODE to learn the continuous scale dynamics of position encoding (RoPE), allowing for the use of short sequences during training and supporting ultra-long contexts during testing, enabling the context window of LLMs to be expanded to more than four times the original training length without performance loss; this module is also designed as a lightweight component that can be directly inserted into RoPE.
🎯 What it does: ClimODE is proposed, a climate and weather prediction model based on continuous-time neural ODEs, which directly implements the physical continuity equation, ensuring conservation of quantities and capable of producing uncertainty estimates.
🎯 What it does: A CLIP-guided multi-subject visual neural information semantic decoding method (CLIP-MUSED) is proposed, which utilizes a Transformer to extract global brain signal features and guides token learning through the combination of low-level/high-level subject-specific tokens with Representational Similarity Analysis (RSA), achieving effective aggregation of multi-subject data and modeling of individual differences.
🎯 What it does: This study investigates the dense feature representation of Vision Transformer (ViT) in CLIP and proposes CLIPSelf, which enhances the performance of this dense representation in open vocabulary dense prediction tasks through self-distillation.
Closing the Curious Case of Neural Text Degeneration
Matthew Finlayson (University of Southern California), Ashish Sabharwal (Allen Institute for Artificial Intelligence)
CodeGenerationTransformerLarge Language ModelText
🎯 What it does: This paper explains why truncated sampling (such as nucleus and top-k) is effective in large language models and proposes a basis-aware truncation (BAT) sampling algorithm based on the softmax bottleneck.
CO2: Efficient Distributed Training with Full Communication-Computation Overlap
Weigao Sun (OpenNLPLab Shanghai AI Laboratory), Yiran Zhong (OpenNLPLab Shanghai AI Laboratory)
CodeOptimizationComputational EfficiencyImageText
🎯 What it does: Proposes the CO2 method, achieving complete overlap of communication and computation in distributed training, significantly enhancing the scalability of low-bandwidth clusters.
🎯 What it does: This paper constructs the COCO-Periph dataset, simulating human peripheral vision by applying a Uniform Texture Tiling Model (Uniform TTM) to MS-COCO images, and trains and evaluates various object detection models based on this; it also designs a machine psychophysical experiment to compare the performance of humans and models in peripheral object detection.
CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules
Hung Le (Salesforce Research), Shafiq Joty (Salesforce Research)
CodeGenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the CodeChain framework, which utilizes chain-of-thought (CoT) prompts to generate modular code and enhances code quality through a self-revision loop of representative submodules.
🎯 What it does: This paper studies the multi-round code auto-editing task based on code change history and proposes the Coeditor model for implementation.
COLEP: Certifiably Robust Learning-Reasoning Conformal Prediction via Probabilistic Circuits
Mintong Kang (University of Illinois Urbana-Champaign), Bo Li (University of Chicago)
CodeAutonomous DrivingAdversarial AttackImage
🎯 What it does: This paper proposes the COLEP framework, which combines knowledge-driven logical reasoning (using probabilistic circuits) with split-based conformal prediction. It provides a global coverage proof for finite L₂ error and validates its robustness through experiments.
Seyed Saman Saboksayr (University of Rochester), Mariano Tepper (Intel Labs)
CodeGraph Neural NetworkGraphTabular
🎯 What it does: The CoLiDE framework is proposed, which simultaneously learns the linear DAG structure and the noise variance of each node through continuous optimization.
Combinatorial Bandits for Maximum Value Reward Function under Value-Index Feedback
Yiliu Wang (Allen Institute), Milan Vojnovic (London School of Economics)
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: Proposed a method to learn and maximize rewards in the k-MAX combinatorial multi-armed bandit problem using only feedback from the maximum value and the corresponding arm index;
CompA: Addressing the Gap in Compositional Reasoning in Audio-Language Models
Sreyan Ghosh (University of Maryland), Dinesh Manocha (University of Maryland)
CodeClassificationRetrievalLarge Language ModelContrastive LearningMultimodalityBenchmarkAudio
🎯 What it does: Two expert-annotated benchmarks, CompA-order and CompA-attribute, were proposed to evaluate the capability of audio-language models (ALM) in combinatorial reasoning. Based on this, the CompA-CLAP model was developed, enhancing the combinatorial reasoning ability of ALM through an improved contrastive learning method.
Complete and Efficient Graph Transformers for Crystal Material Property Prediction
Keqiang Yan (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeGraph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: The study investigates the prediction of crystal material properties and proposes a geometrically complete crystal graph representation along with the ComFormer model.
Complex priors and flexible inference in recurrent circuits with dendritic nonlinearities
Benjamin S. H. Lyo (New York University), Cristina Savin (New York University)
CodeSpiking Neural NetworkDiffusion modelImage
🎯 What it does: A sampleable neural network utilizing dendritic branch neurons and global oscillations has been constructed, capable of implicitly representing complex priors and achieving flexible posterior inference in perceptual tasks.
🎯 What it does: A new conditional diffusion model called Compose and Conquer (CnC) is proposed, which can place multiple objects in 3D depth space and inject global semantics from multiple sources into local areas.
🎯 What it does: A multi-granularity uncertainty regularization method is proposed for combining fine-grained one-to-one matching with coarse-grained one-to-many matching in synthetic image retrieval.
Compositional Conservatism: A Transductive Approach in Offline Reinforcement Learning
Yeda Song (Seoul National University), Gunhee Kim (Seoul National University)
CodeReinforcement LearningTabularBenchmark
🎯 What it does: A conservative framework COCOA for offline reinforcement learning is proposed, which decomposes the state into anchor and delta, and utilizes an inverse dynamics model for anchor pathfinding, thereby enhancing the generalization ability of offline RL.
CodeGenerationOptimizationDiffusion modelTime SeriesPhysics Related
🎯 What it does: This paper proposes a reverse design framework based on diffusion models, called CinDM, which utilizes an energy function for joint optimization of design variables and achieves extrapolation of time, object quantity, and geometric shapes by combining different sub-models.
Dongyoung Go (Naver Corporation), Marc Dymetman (Independent Researcher)
CodeRecommendation SystemExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed and experimented with the 'Combination Preference Model (CPM)', which breaks down single preference evaluations into several interpretable features, then uses LM to extract feature scores and combines them into a final preference score through logistic regression;
Compressed Context Memory for Online Language Model Interaction
Jang-Hyun Kim (Seoul National University), Hyun Oh Song (Seoul National University)
CodeCompressionRecommendation SystemTransformerLarge Language ModelText
🎯 What it does: A compressed context memory (CCM) framework is proposed, which supports dynamic context compression during online inference by inserting lightweight conditional LoRA for adaptive compression of keys/values in the Transformer.
🎯 What it does: A regularization method called Least Volume is proposed, which can compress the latent space dimension of the autoencoder into a low-dimensional linear subspace without requiring prior information.
Compressing LLMs: The Truth is Rarely Pure and Never Simple
AJAY KUMAR JAISWAL, Yinfei Yang (Apple)
CodeRetrievalCompressionTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper explores the true performance of compressed large language models in terms of knowledge retention, reasoning, retrieval, summarization, and instruction following by creating a new evaluation benchmark called LLM-KICK.
Aya Abdelsalam Ismail (Genentech), Kyunghyun Cho (New York University)
CodeGenerationExplainability and InterpretabilityDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A universal Concept Bottleneck (CB) layer is proposed, which can be inserted into VAE, GAN, and diffusion models, aligning the internal representations of generative models with human-interpretable concepts, thus achieving interpretable and adjustable generation.
Conditional Information Bottleneck Approach for Time Series Imputation
MinGyu Choi (Massachusetts Institute of Technology), Changhee Lee (Chung-Ang University)
CodeAuto EncoderContrastive LearningImageTime SeriesSequentialBiomedical Data
🎯 What it does: A framework for time series missing value imputation based on Conditional Information Bottleneck (CIB) is proposed and implemented, addressing the shortcomings of traditional information bottleneck in capturing temporal dynamics.
Gabriel Della Maggiora, Artur Yakimovich (Helmholtz Zentrum Dresden Rossendorf)
CodeRestorationGenerationSuper ResolutionDiffusion modelImageBiomedical Data
🎯 What it does: A conditional variational diffusion model (CVDM) is proposed, which can learn noise scheduling during training to achieve high-quality sampling for inverse problems.
Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models
Kyuyoung Kim (KAIST), Kimin Lee (Google DeepMind)
CodeImage TranslationOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelImageTextBenchmark
🎯 What it does: Confidence calibration of reward optimization for text-to-image models to reduce alignment and quality degradation caused by over-optimization.
Victor Quach (Massachusetts Institute of Technology), Regina Barzilay (Massachusetts Institute of Technology)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelTextMagnetic Resonance Imaging
🎯 What it does: Provides a reliable prediction set for generative language models without distribution and with limited sample size, and can further select high-confidence components at the clause or sentence level.
Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers
Qingyan Guo (Tsinghua University), Yujiu Yang (Tsinghua University)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By combining evolutionary algorithms with large language models (LLM), this paper automates the optimization of discrete prompts, generating readable and more effective prompts.
ConR: Contrastive Regularizer for Deep Imbalanced Regression
Mahsa Keramati (Simon Fraser University), R. David Evans (Borealis AI)
CodeDepth EstimationContrastive LearningImage
🎯 What it does: A contrastive regularizer for deep imbalance regression, ConR, is proposed, which utilizes local and global similarities in the continuous label space to regularize the feature space and prevent feature collapse of minority samples.
Consciousness-Inspired Spatio-Temporal Abstractions for Better Generalization in Reinforcement Learning
Harry Zhao, Yoshua Bengio (Mila)
CodeReinforcement Learning
🎯 What it does: Proposes the Skipper framework, which uses spatiotemporal abstraction to automatically generate proxy problems, achieving zero-shot generalization reinforcement learning through checkpoints in decision-making planning.
🎯 What it does: The CURE framework is proposed, achieving a better balance between adversarial robustness and standard generalization through dynamic retention, updating, and revision of weights in adversarial training.
🎯 What it does: The researchers use a consistency model as a policy representation in reinforcement learning, proposing two algorithms: Consistency-BC (Behavior Cloning) and Consistency-AC (Actor-Critic), and evaluate them in offline, offline-to-online, and online RL scenarios.
🎯 What it does: In response to Graph Anomaly Detection (GAD) in scenarios with limited supervision, the CONSISGAD model is proposed, which utilizes consistency training and learnable data augmentation to fully leverage unlabeled data, and enhances detection performance through a GNN structure driven by homophily distribution.
🎯 What it does: This paper proposes the Consistency Trajectory Model (CTM), a generative framework that simultaneously considers score-based models and distillation models, capable of achieving probability flow ODE jumps at arbitrary time intervals in a single forward pass, and supports controllable sampling.
Consistency-guided Prompt Learning for Vision-Language Models
Shuvendu Roy (Queen's University), Ali Etemad (Queen's University)
CodeClassificationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: In the few-shot fine-tuning of visual-language models (such as CLIP), a method called Consistency-Guided Prompt Learning (CoPrompt) is proposed, which enhances the model's generalization ability in downstream tasks and zero-shot tasks by utilizing consistency constraints, input perturbations, and a combination of prompts and adapters.
Consistent algorithms for multi-label classification with macro-at-$k$ metrics
Erik Schultheis (Aalto University), Krzysztof Dembczynski (Yahoo Research)
CodeClassificationOptimizationBenchmark
🎯 What it does: A consistency learning framework for macro-average (macro-atk) metrics that predicts only k labels for each instance in multi-label classification is proposed, and the analytical form of the optimal classifier is provided.
Constrained Bi-Level Optimization: Proximal Lagrangian Value Function Approach and Hessian-free Algorithm
Wei Yao (Southern University of Science and Technology), Jin Zhang (Southern University of Science and Technology)
CodeOptimizationFederated LearningTabular
🎯 What it does: A single-layer smoothing method based on the proximal Lagrangian value function is proposed, and on this basis, a Hessian-free, single-loop gradient optimization algorithm is designed to solve bi-level optimization problems with upper and lower layer coupling constraints.
🎯 What it does: A cross-language label projection method based on constrained decoding (CODEC) is proposed for automatically translating and projecting word/phrase-level labels in low-resource languages.
Constraint-Free Structure Learning with Smooth Acyclic Orientations
Riccardo Massidda (Università di Pisa), Davide Bacciu (Università di Pisa)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: A constraint-free continuous optimization framework COSMO is designed, utilizing smooth priority vectors and temperature annealing to learn DAGs.
Sharut Gupta (Meta AI, Massachusetts Institute of Technology), Kartik Ahuja (Meta AI)
CodeDomain AdaptationTransformerLarge Language ModelImage
🎯 What it does: A new domain generalization method is proposed - In-Context Risk Minimization (ICRM), which treats the environment as context and dynamically focuses on minimizing the risk of the test environment using unlabeled contextual information during inference.
ContextRef: Evaluating Referenceless Metrics for Image Description Generation
Elisa Kreiss (University of California Los Angeles), Nick Haber (Stanford University)
CodeGenerationSupervised Fine-TuningVision Language ModelImageTextBenchmark
🎯 What it does: Proposed the ContextRef benchmark, which evaluates the quality of reference-free image descriptions using human assessment and data augmentation.
Continual Learning in the Presence of Spurious Correlations: Analyses and a Simple Baseline
Donggyu Lee (Sungkyunkwan University), Taesup Moon (Seoul National University)
CodeClassificationData-Centric LearningImage
🎯 What it does: This study investigates the issue of bias transfer caused by dataset bias in continual learning, systematically evaluating its impact on forward and backward transfer, and proposes a general baseline method BGS based on group-class balanced greedy sampling.
🎯 What it does: A continuous momentum filtering (CMF) framework is proposed for online testing time adaptation (OTTA), which alternately updates the model between SGD optimization and parameter denoising using a Kalman filter.
CodeDomain AdaptationAdversarial AttackTabularTime Series
🎯 What it does: A new continuous domain invariance learning framework (CIL) is proposed to address the issue of traditional methods like IRM failing easily in continuous domains.
🎯 What it does: This paper proposes a new image extrapolation method called PQDiff, which can generate image content at any multiple in a single step, addressing two technical challenges that have not been solved in the existing literature: 1) extrapolation at arbitrary continuous multiples, and 2) completing extrapolation in one step.