These 550 ICML 2024 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICML 2024 paper, free trial on arXivSub.
${\rm E}(3)$-Equivariant Actor-Critic Methods for Cooperative Multi-Agent Reinforcement Learning
Dingyang Chen (University of South Carolina), Qi Zhang (University of South Carolina)
π― What it does: A multi-agent reinforcement learning framework utilizing E(3) Euclidean symmetry is proposed, designing an actor-critic architecture based on E(3) equivariant message passing networks (SEGNN);
$\texttt{MoE-RBench}$: Towards Building Reliable Language Models with Sparse Mixture-of-Experts
Guanjie Chen (Shanghai Artificial Intelligence Laboratory), Yu Cheng (The Chinese University of Hong Kong)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelMixture of ExpertsTextBenchmark
π― What it does: This paper proposes MoE-RBench, a comprehensive benchmark for evaluating the multi-dimensional reliability metrics of sparse expert models (MoE) in terms of safety, hallucination, adversarial robustness, and out-of-distribution (OOD) robustness.
A Computational Framework for Solving Wasserstein Lagrangian Flows
Kirill Neklyudov (University of Montreal), Alireza Makhzani (Vector Institute)
CodeOptimizationFlow-based ModelBiomedical Data
π― What it does: A unified modeling of various variants (such as OT, SchrΓΆdinger bridge, unbalanced OT, etc.) is proposed through the minimization of the Lagrangian action in probability density space, along with a deep learning solving framework;
A General Online Algorithm for Optimizing Complex Performance Metrics
Wojciech Kotlowski, Krzysztof Dembczynski (Poznan University of Technology)
CodeClassificationOptimizationTabular
π― What it does: A general online algorithm OMMA is proposed to directly maximize complex non-decomposable performance metrics (such as F-measure, G-mean, etc.) in binary classification, multi-class classification, and multi-label problems, along with a theoretical regret upper bound.
π― What it does: This paper proposes a novel generative method called the Coupled Counterfactual Generative Adversarial Model (C2 GAM), which reinterprets confounding bias as an out-of-distribution (OOD) problem in discrete environments. It utilizes a small amount of unbiased representative data along with a large amount of biased observational data to jointly generate missing S=0 samples and missing S labels, thereby eliminating confounding bias and improving the accuracy of causal effect estimation.
A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity
Andrew Lee (University of Michigan), Rada Mihalcea (University of Michigan)
CodeOptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study investigates the mechanism of the alignment algorithm (DPO) in reducing toxic behavior in large language models, finding that toxic information is represented through specific MLP value vectors. It demonstrates that DPO avoids these toxic trigger areas by slight weight shifts (GPT2) or gating mechanisms (Llama2); it also shows that toxicity can be easily restored by amplifying key vectors or opening gates.
A New Branch-and-Bound Pruning Framework for $\ell_0$-Regularized Problems
Theo Guyard, Ayse-Nur Arslan
CodeOptimizationTabular
π― What it does: A new Branch-and-Bound (BnB) pruning framework is proposed for solving optimization problems with β_0 regularization terms, which can efficiently evaluate multiple subregions for the presence of optimal solutions without solving convex relaxations.
A New Computationally Efficient Algorithm to solve Feature Selection for Functional Data Classification in High-dimensional Spaces
Tobia Boschi (IBM Research Europe), Jonathan P Epperlein
CodeClassificationComputational EfficiencyTabularTime SeriesElectronic Health Records
π― What it does: An efficient algorithm for simultaneous feature selection and functional classification (FSFC) is proposed and applied to multivariate longitudinal data.
π― What it does: This paper proposes a probabilistic method to learn and explain the adjustable degree of equivariance in differentiable Steerable CNNs.
π― What it does: A distance-aware bottleneck (DAB) model based on information bottleneck and rate-distortion theory is proposed, which estimates uncertainty by learning a codebook to compress training samples and measuring the statistical distance between the input and the codebook.
A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
CodeGenerationTransformerLarge Language ModelText
π― What it does: A distribution-preserving, easily retrievable, and robust LLM watermarking framework called DiPmark is proposed, which can mark generated content without affecting text quality.
A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules?
Agustinus Kristiadi (Vector Institute), Geoff Pleiss (University of British Columbia)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper explores the use of large language models (LLMs) for Bayesian optimization (BO) in molecular space and evaluates their effectiveness in practical chemistry tasks.
A Space Group Symmetry Informed Network for O(3) Equivariant Crystal Tensor Prediction
Keqiang Yan (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeGraph Neural NetworkTabularPhysics Related
π― What it does: GMTNet is proposed, a graph neural network based on crystal space group symmetry and O(3) equivariance, for predicting the dielectric, piezoelectric, and elastic tensors of crystals.
A Sparsity Principle for Partially Observable Causal Representation Learning
Danru Xu (University of Amsterdam), Sara Magliacane
CodeRepresentation LearningTabular
π― What it does: This paper studies how to recover latent causal variables from high-dimensional observations using unpaired, instance-dependent biased observational data in the context of partially observable causal representation learning (CRL) problems.
A Unified Adaptive Testing System Enabled by Hierarchical Structure Search
Junhao Yu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeReinforcement LearningTabular
π― What it does: A unified adaptive testing system framework is proposed, treating CAT and MST as hierarchical search problems, and achieving automatic generation of optimal problem sequences, eliminating the need for manual design.
A3S: A General Active Clustering Method with Pairwise Constraints
Xun Deng (University of Science and Technology of China), Zheng Wang (Alibaba Group)
CodeOptimizationImage
π― What it does: An adaptive active clustering framework A3S is proposed, which utilizes human pairwise constraints to aggregate and split the initial clustering results, significantly improving clustering quality.
π― What it does: A new Distributed Mean Estimation (DME) method called QUIC-FL is proposed, which significantly reduces client encoding time and server decoding time while ensuring optimal O(1/n) NMSE.
CodeRetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: This paper proposes the RaLMSpec framework, which accelerates the iterative retrieval-enhanced language model service using techniques such as speculative retrieval with batch validation, cache prefetching, optimal speculation step scheduling, and asynchronous validation.
Accelerating Look-ahead in Bayesian Optimization: Multilevel Monte Carlo is All you Need
Shangda Yang (University of Manchester), Kody J. H. Law (Meta Platforms, Inc.)
CodeOptimizationTabular
π― What it does: A framework is proposed that applies the Multi-Layer Monte Carlo (MLMC) method to multi-step Bayesian optimization, significantly reducing the computational complexity of estimating the acquisition function in a forward-looking manner.
Accurate LoRA-Finetuning Quantization of LLMs via Information Retention
Haotong Qin (ETH Zurich), Michele Magno (ETH Zurich)
CodeCompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A framework called IR-QLoRA is proposed for LoRA fine-tuning on low-bit quantized LLMs, significantly improving the accuracy of the quantized model.
π― What it does: An ACM-MILP framework based on adaptive constraint modification and community detection is proposed to generate mixed-integer linear programming (MILP) instances that maintain the same level of difficulty as the original instances.
Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations
Jiaqi Zhai (Meta AI), Yu Shi (Meta AI)
CodeRecommendation SystemTransformerSequential
π― What it does: This paper proposes a generative recommender (GRs) framework that treats the recommendation task as a sequence transfer task, and designs an efficient Hierarchical Sequence Transfer Unit (HSTU) to achieve a scalable recommendation model.
Active Label Correction for Semantic Segmentation with Foundation Models
Hoyoung Kim (POSTECH), Jungseul Ok (POSTECH)
CodeSegmentationAutonomous DrivingImage
π― What it does: Utilize a base model to generate pseudo-labels and superpixels, and correct errors in pixel-level semantic segmentation datasets through an active labeling correction framework.
Adapting Pretrained ViTs with Convolution Injector for Visuo-Motor Control
Dongyoon Hwang (KAIST), Jaegul Choo (KAIST)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningImage
π― What it does: This paper proposes a lightweight module called CoIn, which injects the spatial locality and translational equivariance bias of convolutional layers into the pre-trained Vision Transformer (ViT), enabling ViT to better adapt to visual-motor control tasks.
π― What it does: This paper proposes an Advantage-guided Adaptive Policy Regularization method (A2PR) to address the issues of excessive conservativeness and out-of-distribution (OOD) overestimation in offline reinforcement learning.
Adaptive Hierarchical Certification for Segmentation using Randomized Smoothing
Alaa Anani (Max Planck Institute for Informatics), Mario Fritz (Max Planck Institute for Informatics)
CodeSegmentationAutonomous DrivingImage
π― What it does: This paper proposes an adaptive hierarchical randomized smoothing certification method, aimed at pixel-level classification of semantic segmentation models, to provide robustness proofs at multiple semantic levels, thereby reducing rejection rates and enhancing information gain.
Adaptive Observation Cost Control for Variational Quantum Eigensolvers
Christopher J. Anders (Berlin Institute for the Foundations of Learning and Data), Shinichi Nakajima (RIKEN Center for Advanced Intelligence Project)
CodeOptimizationTabularPhysics Related
π― What it does: An adaptive observation cost control method (SubsCoRe) is proposed, which reduces the total measurement overhead by dynamically allocating the number of quantum measurements during the subspace minimum optimization (SMO) process of VQE.
π― What it does: A robust learning algorithm RLVI based on latent Bernoulli variables is proposed, which automatically infers the data contamination ratio and performs parameter estimation within a maximum likelihood framework.
π― What it does: This paper proposes a context-based Las Vegas Bandit algorithm UCR, which can dynamically learn and rank K items from N candidate items to maximize user satisfaction.
Agent Smith: A Single Image Can Jailbreak One Million Multimodal LLM Agents Exponentially Fast
Xiangming Gu (Sea AI Lab), Min Lin (Sea AI Lab)
CodeAdversarial AttackTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
π― What it does: In a multi-agent environment composed of multimodal large language model agents (MLLM), an attack method called 'infectious jailbreak' was studied, demonstrating that a single adversarial image can infect nearly a hundred million agents and induce harmful behavior within only O(log N) rounds of dialogue.
Agent-Specific Effects: A Causal Effect Propagation Analysis in Multi-Agent MDPs
Stelios Triantafyllou (Max Planck Institute for Software Systems), Goran Radanovic (Max Planck Institute for Software Systems)
CodeGraph Neural NetworkReinforcement LearningAgentic AIGraphBiomedical Data
π― What it does: This paper introduces the concept of Agent-Specific Effect (ASE) to quantify the causal effect of a certain agent's behavior on outcomes, which is propagated only through other agents in a Multi-Agent Markov Decision Process (MMDP).
Aligned Objective for Soft-Pseudo-Label Generation in Supervised Learning
Ning Xu (Southeast University), Xin Geng (Southeast University)
CodeMeta LearningImage
π― What it does: The SEAL framework is proposed, which dynamically optimizes the soft pseudo-label generator through a learnable meta-network and alternately trains with the prediction model to achieve adaptive generation and utilization of soft pseudo-labels in supervised learning.
π― What it does: A novel inductive reasoning learning framework is proposedβAmbiguity-Aware Abductive Learning (A3BL), which addresses the issue of traditional ABL easily falling into erroneous pseudo-label problems when dealing with ambiguous reasoning results by probabilistically weighting all possible reasoning candidate sets and optimizing using the EM algorithm.
π― What it does: This paper studies the equation discovery problem in hybrid dynamical systems and proposes an end-to-end AMORE framework that integrates mode classification, equation learning, and switching behavior, extending it to multi-object scenarios (AMORE-MIO).
Yanbo Wang (Peking University), Muhan Zhang (Peking University)
CodeGraph Neural NetworkContrastive LearningGraph
π― What it does: This paper systematically evaluates the performance of 23 types of GNNs in practical expressiveness by constructing a novel expressive dataset BREC and proposing the RPC evaluation framework.
π― What it does: This study investigates a maximum mean discrepancy (MMD)-based independence regularization loss to enhance the token quality of multi-codebook audio segmenters (auto-encoders) in music generation language models.
An Iterative Min-Min Optimization Method for Sparse Bayesian Learning
Yasen Wang (Huazhong University of Science and Technology), ye yuan
CodeOptimizationTabularTime Series
π― What it does: An iterative Min-Min optimization method based on CCCP is proposed to maximize the marginal likelihood function of Sparse Bayesian Learning (SBL), and a global convergence proof is provided.
Sehoon Kim (University of California Berkeley), Amir Gholami (University of California Berkeley)
CodeRecommendation SystemComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper presents LLMCompiler, a framework based on compiler principles that automatically decomposes LLM tasks into a function call graph with dependencies and implements parallel execution, significantly reducing inference latency and cost.
π― What it does: The ME-DOL algorithm is proposed, studying the online optimization perspective of distributed non-convex non-smooth stochastic optimization;
AND: Audio Network Dissection for Interpreting Deep Acoustic Models
Tung-Yu Wu (National Taiwan University), Tsui-Wei Weng (University of California San Diego)
CodeRecognitionExplainability and InterpretabilityLarge Language ModelAudio
π― What it does: This paper proposes AND (Audio Network Dissection), an automated framework based on LLM for generating natural language descriptions of neurons in deep audio models and for concept identification.
AnyTool: Self-Reflective, Hierarchical Agents for Large-Scale API Calls
Yu Du (Tsinghua University), Hongyang Zhang (University of Waterloo)
CodeTransformerLarge Language ModelAgentic AIBenchmarkChain-of-Thought
π― What it does: A self-reflective hierarchical agent system named AnyTool has been developed, utilizing over 16,000 RapidAPI interfaces to address user queries without the need for additional training, employing GPT-4 function calls to implement API retrieval, problem-solving, and self-reflection loops.
APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference
Bowen Zhao (University of Washington), Qingqing Cao (Apple)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The Adaptive Pruning and Tuning (APT) method is proposed for fine-tuning and inference processes of pre-trained language models, achieving adaptive pruning and tuning parameter adjustments during training, significantly enhancing training and inference efficiency.
AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA
Weitao Feng (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
CodeGenerationData SynthesisSafty and PrivacySupervised Fine-TuningDiffusion modelImageText
π― What it does: Proposes AquaLoRA, which utilizes the LoRA module to embed watermarks into the U-Net structure of Stable Diffusion, achieving white-box protection;
ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL
Yifei Zhou (University of California), Aviral Kumar (Google Deepmind)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: A hierarchical multi-round reinforcement learning framework called ArCHer is designed to train large language model agents for efficient RL in multi-round decision-making tasks.
π― What it does: An AI-Hybrid numerical solver based on an attention mechanism (AttNS) is designed to enhance the generalization and robustness of ODE solving in limited data scenarios.
π― What it does: In value-based reinforcement learning, the authors found that early training stages have low utilization and are sensitive to critic bias. They subsequently proposed incorporating a 'hypothesis' representation (weak-model and adaptive rollout) into the actor/critic framework and built the ALH algorithm on top of TD3 to address these issues.
π― What it does: This paper proposes PPS-VAE, a variational autoencoding framework that adaptively infers a context pixel set, utilizing CNP for sparse observation and reconstruction of images.
Logan Murphy (University of Toronto), Xujie Si (University of Toronto)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper proposes a neural-symbolic framework that combines large language models (LLMs) with SMT solvers to automatically convert natural language proofs of Euclidean geometry into Lean proofs and assess their semantic correctness automatically.
Automated Evaluation of Retrieval-Augmented Language Models with Task-Specific Exam Generation
Gauthier Guinet (Amazon), Laurent Callot (Amazon)
CodeRetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Automatically generate multiple-choice exams and use them to assess the task-specific accuracy of retrieval-augmented language models (RAG).
Averaging $n$-step Returns Reduces Variance in Reinforcement Learning
Brett Daley (University of Alberta), Marlos C. Machado (University of Alberta)
CodeReinforcement LearningSequential
π― What it does: This paper studies and proves that composite returns (such as Ξ» returns) have lower variance under the same contraction modulus, and proposes a low-cost Piecewise Ξ» return (Pilar) to approximate Ξ» returns; it also provides a finite sample analysis under linear function approximation and experimentally validates its improvement in sample efficiency.
π― What it does: A unified black-box adversarial patch attack framework is proposed for pixel-level regression tasks (such as monocular depth estimation and optical flow estimation), which can deploy a unified patch on any image through query optimization;
Batch and match: black-box variational inference with a score-based divergence
Diana Cai (Flatiron Institute), Lawrence K. Saul (Columbia University)
CodeOptimizationScore-based ModelImage
π― What it does: A black-box variational inference method based on score matching, called Batch and Match (BaM), is proposed, achieving efficient optimization of the complete covariance Gaussian variational family through closed-form updates and a stochastic proximal point algorithm.
Bayesian Design Principles for Offline-to-Online Reinforcement Learning
Hao Hu (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
CodeReinforcement LearningTabular
π― What it does: This paper proposes an offline-to-online reinforcement learning framework based on Bayesian probability matching (BOORL). It achieves a balance between exploration and exploitation by using bootstrapped ensemble training to bias conservative policies during the offline phase, and then constructing posterior distributions and sampling actions based on softened Q-values during the online phase.
Bayesian Program Learning by Decompiling Amortized Knowledge
Alessandro B. Palmarini (Santa Fe Institute), Siddharth N
CodeNeural Architecture SearchBenchmark
π― What it does: Developed the DREAMDECOMPILER (Dream Decompiling) method, which uses the 'compiled' knowledge from a neural search strategy to guide the selection of library functions, thereby reducing both the breadth and depth of the search.
BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models
Haotian Sun (Georgia Tech), Bo Dai (Georgia Tech)
CodeDomain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: A lightweight adapter BBOX-ADAPTER is proposed for black-box large language models (such as GPT-3.5, Mixtral) to enable adaptive improvements for specific tasks without model parameters or output probabilities.
Best Arm Identification for Stochastic Rising Bandits
Marco Mussi (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
Code
π― What it does: This paper studies the fixed budget best arm identification (BAI) problem in the Stochastic Rising Bandit (SRB) environment, proposing two algorithms R-UCBE (UCB-based) and R-SR (Successive Reject) and providing theoretical upper bounds for error probability and simple reward.
Better & Faster Large Language Models via Multi-token Prediction
Fabian Gloeckle (Meta), Gabriel Synnaeve (Meta)
CodeGenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelText
π― What it does: Improvements have been made to the training method of large language models by adopting multi-step prediction (predicting multiple future tokens at once) instead of the traditional single-step prediction, thereby enhancing sample efficiency and model performance.
Better Locally Private Sparse Estimation Given Multiple Samples Per User
Yuheng Ma (Renmin University of China), Hanfang Yang (Renmin University of China)
CodeOptimizationSafty and PrivacyTabular
π― What it does: This paper studies sparse linear regression under user-level local differential privacy (ULDP) and proposes a new framework that eliminates the linear dependence on dimension d by selecting candidate variables and estimating in a low-dimensional space.
Beyond Individual Input for Deep Anomaly Detection on Tabular Data
Hugo Thimonier (University Paris-Saclay), Bich-LiΓͺn DOAN
CodeAnomaly DetectionTransformerTabular
π― What it does: A deep anomaly detection method based on Non-Parametric Transformer (NPT) is proposed, which generates anomaly scores by reconstructing the features of normal samples through masking in the training set.
π― What it does: This paper proposes a pseudo-likelihood estimation method based on Score Matching, called SMASH, for learning labeled spatiotemporal point processes. It can provide confidence intervals/regions for event times and locations, as well as calibrated probabilities for event labels.
Beyond Sole Strength: Customized Ensembles for Generalized Vision-Language Models
Zhihe Lu (National University of Singapore), Xinchao Wang (National University of Singapore)
CodeClassificationRecognitionDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: The paper proposes three ensemble strategies based on pre-trained vision-language models (CLIP), utilizing the collaboration of weak and strong models to enhance the zero-shot, few-shot, and cross-dataset generalization performance.
Bidirectional Reciprocative Information Communication for Few-Shot Semantic Segmentation
Yuanwei Liu (Northwestern Polytechnical University), Fahad Shahbaz Khan (Linkoping University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: The IFRNet framework is proposed, which uses bidirectional recursive information transmission to address the intra-class diversity problem in few-shot semantic segmentation.
Bifurcated Attention for Single-Context Large-Batch Sampling
Ben Athiwaratkun (Together.ai), Bing Xiang (Goldman Sachs)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: A context-aware bifurcated attention method is proposed for the incremental decoding phase in single-context large batch sampling, significantly reducing KV cache memory I/O and thereby decreasing inference latency.
Boosting Reinforcement Learning with Strongly Delayed Feedback Through Auxiliary Short Delays
Qingyuan Wu (University of Liverpool), Chao Huang (University of Southampton)
CodeReinforcement LearningSequential
π― What it does: This paper proposes an Auxiliary-Delayed Reinforcement Learning (AD-RL) framework that utilizes short-delay auxiliary tasks to accelerate the learning of the original long-delay tasks, while ensuring performance and reducing sample complexity.
Bounded and Uniform Energy-based Out-of-distribution Detection for Graphs
Shenzhi Yang (Soochow University), Xiaofang Zhang (Soochow University)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: The NODESAFE method is proposed, which introduces two constraints (L_bound and L_uniform) during training to bound and uniform the negative energy scores of GNNs, thereby enhancing the OOD detection capability at the node level.
π― What it does: This paper proposes a breadth-first exploration method based on adaptive grids (BEAG) for systematically managing achieved and unachieved sub-goals in sparse long-horizon goal-conditioned reinforcement learning, significantly improving sample efficiency.
Breaking the Barrier: Enhanced Utility and Robustness in Smoothed DRL Agents
Chung-En Sun (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
CodeReinforcement LearningVideo
π― What it does: A stochastic smooth robust training algorithm for reinforcement learning, S-DQN and S-PPO, is proposed to enhance clean rewards and robust performance.
π― What it does: A few-shot transfer learning method based on diffusion models, DPMs-ANT, is proposed, which significantly improves the quality and diversity of few-shot image generation by combining similarity-guided training and adversarial noise selection.
π― What it does: A general and efficient data subset selection method called Best Window Selection (BWS) is proposed, which achieves data pruning by sorting based on difficulty scores and selecting the best subset within a window.
CaM: Cache Merging for Memory-efficient LLMs Inference
Yuxin Zhang (Xiamen University), Rongrong Ji (Xiamen University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A cache merging (CaM) method is proposed, which utilizes attention weights to merge evicted KV caches into retained caches, reducing memory usage and output disturbance.
π― What it does: This paper systematically evaluates the contextual learning ability of non-attention state space models (especially Mamba) in various artificially constructed reasoning tasks and compares it with Transformers. It then proposes the MambaFormer hybrid model that integrates both architectures, significantly improving multi-task performance.
Causal Discovery via Conditional Independence Testing with Proxy Variables
Mingzhou Liu (Peking University), Yizhou Wang (Peking University)
CodeBiomedical DataElectronic Health Records
π― What it does: A non-parametric hypothesis testing method based on proxy variables is proposed, which tests causal relationships between continuous variables using discretization techniques, avoiding biases caused by unobserved confounding in traditional conditional independence tests.
Causal Discovery with Fewer Conditional Independence Tests
Kirankumar Shiragur (Broad Institute), Caroline Uhler (Massachusetts Institute of Technology)
CodeGraph Neural NetworkGraphTabular
π― What it does: A causal graph representation that can be recovered with only polynomially many conditional independence tests is proposedβCausal Consistent Partition Graph (CCPG), along with an efficient corresponding algorithm; under identifiable graphs or situations with sufficient interventions, a complete DAG can be further recovered.
Causal Effect Identification in LiNGAM Models with Latent Confounders
Daniele Tramontano (Technical University of Munich), Negar Kiyavash (Ecole Polytechnique Federale de Lausanne)
CodeGraph
π― What it does: The study provides general identifiability conditions for causal effects between observable variables in linear non-Gaussian causal models (LiNGAM) that include potential confounding variables, and proposes corresponding judgment algorithms and estimation methods.
π― What it does: A causal representation learning (CRL) model is proposed that utilizes the grouping structure of observed variables to achieve identifiability, along with a corresponding self-supervised learning framework G-CaRL;
Causality Based Front-door Defense Against Backdoor Attack on Language Models
Yiran Liu (Tsinghua University), Yang Yu (Tsinghua University)
CodeClassificationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A front-door adjustment defense framework (FABE) based on causal inference has been designed and implemented to counter backdoor attacks on language models.
Centralized Selection with Preferences in the Presence of Biases
L. Elisa Celis (Yale University), Andrew Xu (Yale University)
CodeTabular
π― What it does: The study investigates how to maximize utility while considering candidate preference fairness and group representativeness fairness in centralized selection problems with systemic bias.
π― What it does: Research and optimize the training process of Physics-Informed Neural Networks (PINNs), focusing on the condition number issue of the loss landscape;
Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference
Wei-Lin Chiang (University of California Berkeley), Ion Stoica (University of California Berkeley)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
π― What it does: An open, human-preference-based real-time large language model (LLM) evaluation platform called Chatbot Arena has been constructed, and approximately 100K pairs of model comparison preference data have been made public.
CHEMREASONER: Heuristic Search over a Large Language Modelβs Knowledge Space using Quantum-Chemical Feedback
Henry W. Sprueill (Pacific Northwest National Laboratory), Sutanay Choudhury (Pacific Northwest National Laboratory)
CodeDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningGraphChain-of-Thought
π― What it does: A heuristic search framework named CHEMREASONER has been developed, combining large language models (LLM) with quantum chemistry (GNN) feedback to achieve automated catalyst discovery.
Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference
Luca Masserano (Carnegie Mellon University), Ann B. Lee (Carnegie Mellon University)
CodeClassificationBiomedical Data
π― What it does: In the presence of noise parameters and label distribution shift (GLS), a reliable classification and uncertainty quantification method is proposed, which can provide a prediction set that satisfies conditional and marginal coverage without retraining the model;
Guy Horowitz (Technion Israel Institute of Technology), Nir Rosenfeld (Technion Israel Institute of Technology)
CodeClassificationOptimizationTabular
π― What it does: The study examines the self-selection behavior of candidates based on whether they apply for a learned classifier in selective machine learning, and proposes a differentiable learning framework to simultaneously optimize prediction accuracy and self-selection distribution.
CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks
Yulong Huang (Hong Kong University of Science and Technology), Bojun Cheng (Hong Kong University of Science and Technology)
CodeSpiking Neural NetworkImage
π― What it does: A Complementary Leaky Integrate-and-Fire (CLIF) neuron is proposed and implemented to address the vanishing gradient problem encountered by traditional LIF neurons during SNN training, allowing for direct replacement of LIF in training across various network architectures.
CodeConvolutional Neural NetworkTime SeriesSequentialPhysics Related
π― What it does: This paper proposes the Clifford-Steerable CNN (CS-CNN) framework, which is capable of maintaining equivariance in convolutional neural networks under isometric transformations E(p,q) in pseudo-Euclidean space R^{p,q}, handling multi-vector fields.
π― What it does: Developed CLIPZyme, a virtual enzyme screening framework based on contrastive learning, which ranks enzyme sequences given a chemical reaction to identify the catalytic activity of natural enzymes.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: In response to the sequential bottleneck of LLM inference, a new Consistency Large Language Models (CLLMs) is proposed, which fine-tunes the pre-trained model to predict multiple extra tokens at once during the Jacobi parallel decoding process, thereby achieving efficient parallel inference.
π― What it does: This paper proposes a clustering-based group federated learning algorithm, CFL-GP, which can automatically partition clients into several clusters and train a shared model for each cluster without sharing the original data.
π― What it does: A method called Progressive Tensor Training (PuTT) is proposed for learning compact and high-quality representations of visual data.
CogBench: a large language model walks into a psychology lab
Julian Coda-Forno (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: This paper proposes the CogBench benchmark, which systematically evaluates 40 large language models (LLMs) using ten behavioral metrics derived from seven cognitive psychology experiments, and investigates the impact of factors such as RLHF, model size, code training, and prompt-engineering on LLM behavior.
COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability
Xingang Guo (University of Illinois), Bin Hu (University of Illinois)
CodeAdversarial AttackLarge Language ModelText
π― What it does: This paper proposes a new framework called COLD-Attack for generating controllable attacks on large language models (LLMs), aimed at enhancing the stealth and controllability of the attacks.
π― What it does: We propose CA-TRIDE, a framework for adversarial training that combines collapse awareness and triplet decoupling to enhance the robustness of image retrieval models.
π― What it does: The first collective provable robustness scheme against graph injection attacks is proposed, which transforms the problem into a binary quadratic constrained linear programming (BQCLP) and further linearizes it to obtain a solvable LP, ensuring prediction consistency for multiple nodes on a single perturbed graph.
Combining Experimental and Historical Data for Policy Evaluation
Ting Li (Shanghai University of Finance and Economics), Hongtu Zhu (University of North Carolina at Chapel Hill)
CodeReinforcement LearningTabularBiomedical Data
π― What it does: The study investigates how to integrate experimental data with historical control group data to evaluate the average treatment effect of new strategies, and proposes non-pessimistic and pessimistic weighted estimators.
Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks
Rahul Ramesh (University of Pennsylvania), Hidenori Tanaka (NTT Research)
CodeData SynthesisExplainability and InterpretabilityTransformerPrompt EngineeringTextSequentialChain-of-Thought
π― What it does: This study investigates the compositional capabilities of autoregressive Transformers on synthetic interpretability tasks, exploring how the model learns to apply predefined function sequences to input strings and generate correct outputs.
π― What it does: This work studies and proposes a method to compute the upper bound of the second derivative (curvature) of deep differentiable networks, and based on this upper bound, provides closed-form robustness and attack certificates, while using curvature as a differentiable regularization term in training to enhance the network's adversarial robustness.