International Conference on Machine Learning Β· 722 papers
POROver: Improving Safety and Reducing Overrefusal in Large Language Models with Overgeneration and Preference Optimization
Batuhan K. Karaman (Cornell University), Xia Song (Microsoft)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the POROver (Preference Optimization for Reducing Overrefusal) strategy by using more advanced teacher models (such as GPT-4o) for overgeneration of general and toxic instructions, aiming to reduce the overrefusal rate of large language models while maintaining high safety.
π― What it does: A diffusion model called PPDiff is proposed for the joint design of protein ligand sequences and structures, aimed at generating high-affinity binding proteins for any protein target.
Preconditioned Riemannian Gradient Descent Algorithm for Low-Multilinear-Rank Tensor Completion
Yuanwei Zhang (Shanghai Jiao Tong University), Jian-Feng Cai (Hong Kong University of Science and Technology)
CodeOptimizationVideo
π― What it does: This paper addresses the low multilinear rank tensor completion problem and proposes a Preconditioned Riemannian Gradient Descent (PRGD) algorithm, which utilizes a data-driven Riemannian metric to accelerate convergence while maintaining approximately optimal sampling complexity.
Prediction models that learn to avoid missing values
Lena Stempfle (Chalmers University of Technology), Fredrik D. Johansson (Chalmers University of Technology)
CodeTabularAlzheimer's Disease
π― What it does: A missing value avoidance (MA) machine learning framework is proposed, which actively encourages decision trees, sparse linear models, and tree ensemble models to minimize reliance on missing features during prediction through a penalty term, while maintaining or improving predictive performance.
Amr Alkhatib (Orebro University), Michalis Vazirgiannis (Ecole Polytechnique)
CodeExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImageTabular
π― What it does: A method called ViaSHAP is proposed, which embeds the calculation of Shapley values into model training, allowing predictions to be directly obtained by summing Shapley values, while no additional post-processing is required during inference.
Predictive Data Selection: The Data That Predicts Is the Data That Teaches
KaShun SHUM, Junxian He (Hong Kong University of Science and Technology)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
π― What it does: This paper proposes a data selection method called PRESELECT, which selects the most helpful data for learning by evaluating the correlation between the loss of text on different pre-trained models and the ranking of downstream task performance.
Preference Learning for AI Alignment: a Causal Perspective
Kasia Kobalczyk, Mihaela van der Schaar (University of Cambridge)
CodeRecommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: Proposes to view preference learning as a causal inference problem, exploring causal assumptions such as user goal confounding and potential treatment variables in LLM alignment;
Privacy-Shielded Image Compression: Defending Against Exploitation from Vision-Language Pretrained Models
Xuelin Shen (Shenzhen University), Wenhan Yang (Pengcheng Laboratory)
CodeCompressionOptimizationSafty and PrivacyAdversarial AttackVision Language ModelAuto EncoderImageText
π― What it does: The framework PSIC implements privacy protection during the image compression phase, generating a bitstream that can be decoded on demand into an encrypted version and a full version.
Private Federated Learning using Preference-Optimized Synthetic Data
Charlie Hou (Carnegie Mellon University), Giulia Fanti (Carnegie Mellon University)
CodeOptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: A private federated learning method based on policy optimization, POPri, is proposed, which utilizes LLM to fine-tune synthetic data generation based on client feedback as rewards.
π― What it does: A method called ProDiff is proposed, which can complete trajectory missing point interpolation using only endpoint information. It utilizes prototype learning and diffusion models to achieve trajectory reconstruction.
π― What it does: This paper proposes a sampler that combines Parallel Tempering with diffusion models, called the Progressive Tempering Sampler with Diffusion (PTSD), which generates samples from a high-temperature model to a low-temperature model through temperature guidance.
Evan Frick (University of California), Ion Stoica (University of California)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper proposes the P2L method, which trains a large language model to directly output Bradley-Terry coefficients after receiving natural language prompts, generating a model ranking for each prompt to achieve fine-grained evaluation.
ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis
Haoxiong Liu (Tsinghua University), Andrew C Yao (Shanghai Qi Zhi Institute)
CodeLarge Language ModelPrompt Engineering
π― What it does: A fine-grained proof structure analysis method named ProofAug is proposed, which utilizes complete proofs generated by LLM and extracts repairable proof structures through Maximum Compatible Semi-Proofs (MCSP), combining ATP and built-in proof methods to achieve efficient neural theorem proving.
Propagate and Inject: Revisiting Propagation-Based Feature Imputation for Graphs with Partially Observed Features
Daeho Um (Samsung Electronics), Seulki Park (University of Michigan)
CodeGraph Neural NetworkGraph
π― What it does: This study investigates the problem of propagative imputation of missing features in graph data and proposes a new method, FISF, which addresses performance degradation caused by low-variance channels by injecting synthetic features.
Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing
Yuan Xin (CISPA Helmholtz Center for Information Security), Xiao Zhang (CISPA Helmholtz Center for Information Security)
CodeOptimizationAdversarial AttackConvolutional Neural NetworkImageBiomedical Data
π― What it does: A provably cost-sensitive adversarial robustness defense framework based on randomized smoothing is proposed, providing a cost-sensitive certified radius and training method.
π― What it does: Proposes the FedGO algorithm, which uses a GAN discriminator to achieve approximately optimal federated ensemble distillation, addressing the client heterogeneity issue.
Puzzle: Distillation-Based NAS for Inference-Optimized LLMs
Akhiad Bercovich, Ran El-Yaniv
CodeOptimizationComputational EfficiencyKnowledge DistillationNeural Architecture SearchTransformerLarge Language ModelText
π― What it does: This paper studies a distillation-based NAS framework called Puzzle, which can generate sub-models that efficiently infer on a single NVIDIA H100 GPU while maintaining nearly the same performance as the original large model (e.g., Llama-70B).
π― What it does: A PyTDC platform has been constructed, providing multimodal single-cell data retrieval, training, evaluation, and inference, and for the first time offering benchmarks and tools for single-cell drug-target naming tasks;
π― What it does: This paper proposes a quantization framework Q-VDiT specifically for video generation diffusion Transformers (DiT), addressing the issues of quantization information loss and spatiotemporal consistency loss.
π― What it does: A framework for image quality assessment based on Mamba, QMamba, is proposed, modeling task-specific, general, and transferable IQA.
QoS-Efficient Serving of Multiple Mixture-of-Expert LLMs Using Partial Runtime Reconfiguration
HamidReza Imani (George Washington University), Tarek El-Ghazawi (George Washington University)
CodeTransformerLarge Language ModelMixture of ExpertsText
π― What it does: A MoE LLM system has been built to efficiently serve multiple models on a single GPU, utilizing expert similarity to merge shared memory and dynamically replace non-expert layers at runtime to reduce memory usage.
QuEST: Stable Training of LLMs with 1-Bit Weights and Activations
Andrei Panferov (International Scientific and Technological Academy), Dan Alistarh (International Scientific and Technological Academy)
CodeTransformerLarge Language ModelText
π― What it does: A quantization-aware training method named QuEST is proposed, which can stably train large language models with weights and activations ranging from 1-bit to 4-bit.
π― What it does: Proposes the RAGGED framework for systematic evaluation of retrieval-augmented generation systems under different retrieval depths, reader models, retrievers, and noise conditions.
Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing
Han Jiang (Tongji University), Xing Xie (Microsoft Research Asia)
CodeGenerationTransformerLarge Language ModelText
π― What it does: This study proposes an evaluation framework based on Generative Evolutionary Testing (GETA) for dynamically measuring the value and ethical alignment levels of large language models (LLMs).
π― What it does: The researchers view the training of Generative Flow Networks (GFlowNets) as flow iteration in dynamic programming, discovering that it is equivalent to the value evaluation of random policies in tree structures or DAGs that satisfy path invariance. They propose the Rectified Random Policy Evaluation (RPE) algorithm based on random policy evaluation, achieving the same reward matching effect as GFlowNets.
π― What it does: In cross-domain few-shot learning, the authors found that using random registers instead of learnable prompts can significantly improve the cross-domain generalization performance of Vision Transformers. They proposed the REAP method, which enhances attention perturbation by randomly replacing clustering blocks in the image semantic region, resulting in a flatter loss landscape.
π― What it does: A Ranked Entropy Minimization (REM) method is proposed, which achieves efficient and stable updates for Continual Test-Time Adaptation (CTTA) using an advanced masking chain and dual losses (mask consistency loss + entropy ranking loss).
RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding
Guanzheng Chen (National University of Singapore), Michael Qizhe Shieh (National University of Singapore)
CodeGenerationRetrievalComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes the RAPID method, which combines retrieval-augmented speculative decoding to accelerate long-context LLM inference and improve generation quality.
RATE: Causal Explainability of Reward Models with Imperfect Counterfactuals
David Reber (University of Chicago), Victor Veitch (University of Chicago)
CodeRecommendation SystemExplainability and InterpretabilityLarge Language ModelText
π― What it does: The RATE (Rewrite-based Attribute Treatment Estimator) method is proposed, which uses dual LLM rewriting to estimate the causal effects of reward models on high-level attributes such as sentiment, helpfulness, length, etc.
Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger
Qi Yang (Alibaba Cloud Computing), Shiming Xiang (University of Chinese Academy of Sciences)
CodeGenerationRetrievalTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: A multi-modal retrieval-enhanced generation framework RCTS is proposed, which constructs a knowledge base using reasoning context and improves the VQA performance of large visual language models through tree search reordering.
Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures
Yingzhao Jian (Zhejiang University), Yi Yang (Zhejiang University)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes the Reaction Graph (RG) - a unified representation of chemical reaction graphs that can simultaneously capture the molecular structures of reactants and products, the reaction edges generated by atom mapping, and three-dimensional geometric information (edge lengths and angle edges), thereby directly learning the transformation features of reactions during the message passing process of graph neural networks.
Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment
Cheryl Li (Independent Researcher), Steven Y. Guo
CodeTransformerLarge Language ModelTextChain-of-Thought
π― What it does: This paper proposes a framework RaLU that constructs reliable reasoning paths by extracting and aligning program logic units during inference, aiming to reduce the reasoning hallucinations produced by LLMs.
Rectifying Conformity Scores for Better Conditional Coverage
Vincent Plassier (Lagrange Mathematics and Computing Research Center), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
CodeTabularBenchmark
π― What it does: A new method is proposed to generate confidence sets through trainable transformations to improve conditional coverage while ensuring the accuracy of marginal coverage.
π― What it does: A novel task called R3DGS is proposed for object segmentation in 3D Gaussian Splatting based on natural language descriptions, along with the construction of the corresponding dataset Ref-LERF.
ReFocus: Visual Editing as a Chain of Thought for Structured Image Understanding
Xingyu Fu (University of Pennsylvania), Cha Zhang (Microsoft)
CodeRecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityTabularChain-of-Thought
π― What it does: By allowing a multimodal LLM to generate Python code for visual editing of input images (such as drawing boxes, masking, and highlighting), a visual thinking chain is realized, enhancing selective attention and multi-hop visual reasoning.
π― What it does: This paper reveals the contradiction between the theoretical roots and practical implementation of existing conditional diffusion model guidance methods by reconstructing the theoretical framework of joint distribution scale. Based on this framework, a new guidance strategy called Rectified Gradient Guidance (REG) is proposed to enhance the generation quality of existing guidance methods.
Regress, Don't Guess: A Regression-like Loss on Number Tokens for Language Models
Jonas Zausinger (Technical University of Munich), Jannis Born (IBM Research Europe)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed and implemented Number Token Loss (NTL), allowing the language model to perform regression loss directly on numeric tokens during training, rather than using the traditional cross-entropy.
Regression for the Mean: Auto-Evaluation and Inference with Few Labels through Post-hoc Regression
Benjamin Eyre (Columbia University), David Madras (Google DeepMind)
CodeTabular
π― What it does: This study investigates the use of the Prediction Powered Inference (PPI) framework for mean estimation in scenarios with very few labels, and proposes regression-based improvements (Ridge-PPI and Sigmoid-PPI) to reduce estimation variance.
π― What it does: Proposes Regularized Langevin Dynamics (RLD) and implements two CO solvers based on it: RLSA (simulated annealing version) and RLNN (neural network version).
Zherui Li (Beijing University of Posts and Telecommunications), Xiang Wang (University of Science and Technology of China)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes the RLEdit method, which models the lifelong editing problem as a reinforcement learning task, utilizing a supernetwork for offline training across the entire knowledge sequence to achieve efficient and stable continuous knowledge updates.
Reinforcement Learning for Quantum Control under Physical Constraints
Jan Ole Ernst (University of Oxford), Axel Kuhn (University of Oxford)
CodeOptimizationReinforcement LearningPhysics Related
π― What it does: Using physics-constrained reinforcement learning, a control method for open quantum systems is proposed, capable of generating high-fidelity, realizable pulse sequences under noise and experimental constraints.
π― What it does: This paper proposes AdaReMo, which constructs a fast reward model for reinforcement learning systems with high reward evaluation costs, achieving decoupling of online decision-making and offline evaluation.
π― What it does: A design algorithm configuration selection method based on predictive empowered inference and density ratio weighting is proposed, which can ensure a high probability of success within a multiple testing framework;
ResearchTown: Simulator of Human Research Community
Haofei Yu (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
CodeGraph Neural NetworkLarge Language ModelAgentic AITextBenchmark
π― What it does: This paper presents RESEARCHTOWN, a multi-agent LLM framework based on TextGNN, designed to simulate collaborative activities such as reading, writing, and reviewing papers in human research communities.
Residual Matrix Transformers: Scaling the Size of the Residual Stream
Brian Mak (University of California Santa Cruz), Jeffrey Flanigan (University of California Santa Cruz)
CodeTransformerText
π― What it does: A new variant of Transformer called Residual Matrix Transformer (RMT) is proposed, which replaces the residual flow with an outer product memory matrix to achieve scalable residual flow.
π― What it does: This paper proposes Residual TPP, a lightweight hybrid model that integrates traditional Hawkes processes with neural TPP through Residual Event Decomposition (RED), capable of capturing the periodicity, self-excitement, and the difficult-to-interpret residual components of event streams.
ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals
Utkarsh Saxena (Purdue University), Xin Wang (d-Matrix)
CodeTransformerLarge Language ModelTextMultimodality
π― What it does: ResQ achieves efficient post-training quantization by performing low-rank PCA projection on weights, activations, and KV caches, retaining high-variance subspaces at 8-bit precision while quantizing the remaining parts to 4-bit.
Rethinking Addressing in Language Models via Contextualized Equivariant Positional Encoding
Jiajun Zhu (University of Texas at Austin), Zhangyang Wang (Georgia Tech)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A Contextualized Equivariant Positional Encoding (TAPE) is proposed, which enhances the model's representation of positional information and long sequence reasoning ability by gradually contextualizing positional encoding with the sequence content between Transformer layers.
Rethinking Causal Ranking: A Balanced Perspective on Uplift Model Evaluation
Minqin Zhu (Zhejiang University), Kun Kuang (Zhejiang University)
CodeRecommendation SystemOptimizationTabular
π― What it does: This paper proposes the Principled Uplift Curve (PUC) for a fairer and unbiased evaluation of uplift models, and based on this curve, introduces the Principled Treatment and Outcome Network (PTONet) uplift model;
Rethinking Chain-of-Thought from the Perspective of Self-Training
Zongqian Wu (University of Electronic Science and Technology of China), Lei Feng (Southeast University)
CodeLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Revisiting Chain-of-Thought reasoning from the perspective of self-training, a new CoT framework is proposed, consisting of task-specific prompts and adaptive reasoning iterations, aimed at improving the accuracy of large language models in complex reasoning tasks.
π― What it does: A pseudo-labeling confidence learning framework based on error tolerance is designed, replacing traditional threshold and confidence selection;
Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation
Jian Bi (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
CodeClassificationSegmentationPoint Cloud
π― What it does: The SinPoint method is proposed, which utilizes sine transformations and homotopy to maintain the topological consistency of point clouds for data augmentation.
π― What it does: A new method called Unified Distillation Sampling (UDS) is proposed, aimed at simultaneously supporting the generation and editing of 3D assets.
π― What it does: Proposes the Dual-Arch framework, which utilizes two networks with different architectures (a wide shallow stable network and a deep thin flexible network) to jointly complete continual learning tasks, leveraging knowledge distillation for new knowledge transfer.
π― What it does: This paper proposes HC-SMoE, a hierarchical clustering expert merging framework that does not require retraining and is task-agnostic, aimed at compressing sparse mixture of experts models.
π― What it does: A retrieval-enhanced time series forecasting method called RAFT is proposed, which retrieves historical segments from the training set that are similar to the current input and uses their subsequent values in conjunction with the model input to predict the future.
Wenbin Wang (Wuhan University), Dacheng Tao (Nanyang Technological University)
CodeRecognitionRetrievalComputational EfficiencyLarge Language ModelImageMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: A training-free Retrieval-Augmented Perception (RAP) framework is proposed, which enhances the perception ability of multimodal large language models (MLLM) for high-resolution (HR) images by retrieving and fusing high-resolution image patches related to the query.
Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces
Henry Moss, Tom Diethe (AstraZeneca)
CodeOptimizationDrug DiscoveryAuto EncoderTabular
π― What it does: A Bayesian optimization framework called COWBOYS is proposed, which completely separates the Variational Autoencoder (VAE) from the Gaussian Process (GP). It utilizes GP to directly predict the target in the structural space and restricts sampling to areas where the VAE prior probability is high and the GP predicted values are also high through Bayesian updating.
Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning
Shuai Yi (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
CodeDomain AdaptationTransformerImageAgriculture Related
π― What it does: This paper studies the impact of disrupting the continuity of image tokens in Vision Transformers on cross-domain few-shot learning and proposes the ReCIT method, which reduces domain gaps and enhances model generalization by first performing spatial layer block shuffling followed by frequency domain amplitude balancing shuffling.
Reward Translation via Reward Machine in Semi-Alignable MDPs
Yun Hua (Shanghai Jiao Tong University), Xiangfeng Wang (East China Normal University)
CodeLarge Language ModelReinforcement LearningSequential
π― What it does: A Neural Reward Translation (NRT) framework is proposed in the context of semi-aligned MDP environments, utilizing reward machines and graph matching to achieve cross-domain reward transfer.
CodeOptimizationDrug DiscoveryReinforcement LearningDiffusion modelBiomedical Data
π― What it does: A new framework is proposed for reward-guided iterative optimization during the testing phase of diffusion models, specifically applied to protein and DNA design.
Reward-Guided Speculative Decoding for Efficient LLM Reasoning
Baohao Liao (Language Technology Lab, University of Amsterdam), Caiming Xiong (Salesforce AI Research)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: A Reward-Guided Speculative Decoding (RSD) framework is proposed, which dynamically mixes a lightweight draft model with a powerful target model and evaluates the quality of each step through a process reward model to enhance the inference efficiency of large language models.
Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
Seongho Son (University College London), Ilija Bogunovic (University College London)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper studies the algorithm NS-DPO, which directly optimizes preferences for time drift in large language models (LLMs), and provides theoretical analysis and experimental validation.
RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation
Jingxiang Qu (Stony Brook University), Yi Liu (Stony Brook University)
CodeExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkGraph
π― What it does: A subgraph extraction method based on the influence radius of each atom, RISE, is proposed to explain the predictive decisions of 3D molecular graph neural networks.
Risk and cross validation in ridge regression with correlated samples
Alexander Atanasov (Harvard University), Cengiz Pehlevan (Harvard University)
CodeTime Series
π― What it does: The study investigates the risk of high-dimensional ridge regression in the presence of correlated samples and provides precise asymptotic analysis.
Robust Conformal Outlier Detection under Contaminated Reference Data
Meshi Bashari (Technion Israel Institute of Technology), Yaniv Romano (Technion Israel Institute of Technology)
CodeAnomaly DetectionTabular
π― What it does: The study investigates the robustness of conformal prediction in anomaly detection tasks when there are a small number of outlier samples in the reference data, and proposes an active data cleaning method (Label-Trim) based on a limited labeling budget to improve detection rates.
π― What it does: A theoretical and experimental framework is proposed to improve the robustness of black-box fairness auditing using auditors' prior knowledge.
π― What it does: Research and improve the global pooling method of Transformer outputs, proposing Adaptive Pooling (AdaPool) to better extract useful signals in noisy environments.
Robust Offline Reinforcement Learning with Linearly Structured $f$-Divergence Regularization
Cheng Tang (University of Illinois Urbana-Champaign), Pan Xu (Duke University)
CodeReinforcement LearningTabularFinance Related
π― What it does: In the offline robust reinforcement learning task, a d-rectangular linear RRMDP framework is proposed, and the R2PVI algorithm is designed to achieve robustness against dynamic shifts through f-divergence regularization.
Robust Secure Swap: Responsible Face Swap With Persons of Interest Redaction and Provenance Traceability
Yunshu Dai (Sun Yat-sen University), Chip Hong Chang (Nanyang Technological University)
CodeRecognitionImage TranslationSafty and PrivacyKnowledge DistillationGenerative Adversarial NetworkImage
π― What it does: Proposes the Secure Swap model, which combines the ID Passport layer to achieve person identity (POI) recognition and occlusion, and embeds an invisible watermark in the generated non-POI images for identity tracing and abuse prevention.
Robust Spatio-Temporal Centralized Interaction for OOD Learning
Jiaming Ma (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeGraph Neural NetworkTime Series
π― What it does: A spatiotemporal graph neural network (STOP) based on centralized message passing and message perturbation mechanisms is proposed to address the spatiotemporal OOD learning problem.
rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking
Xinyu Guan (Microsoft Research Asia), Mao Yang (Microsoft Research Asia)
CodeTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: The deep thinking framework rStar-Math, based on a small LLM, achieves high-level mathematical reasoning through self-evolution training strategies and reward models.
RULEBREAKERS: Challenging LLMs at the Crossroads between Formal Logic and Human-like Reasoning
Jason Chan (University of Sheffield), Zhixue Zhao (University of Sheffield)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper systematically evaluates the reasoning abilities of large language models in the context of logical rule violations (rulebreakers) and normal reasoning (non-rulebreakers) by constructing a dataset called RULEBREAKERS.
π― What it does: This paper proposes a training-free, stability criterion-based adaptive acceleration framework SADA, designed to accelerate the sampling process of diffusion models (Diffusion and Flow-matching) in ODE form.
Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets
Ning Lu (Southern University of Science and Technology), Ke Tang (Southern University of Science and Technology)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes Safe Delta, a post-training defense method that is safety-aware for incremental parameters after fine-tuning LLMs, aimed at maintaining model safety and enhancing task performance.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelImageVideoText
π― What it does: To address the quantization bottleneck in attention computation, SageAttention2 is proposed, utilizing INT4 quantization for Q and K, FP8 quantization for ΛP and V, and enhancing accuracy through Q and K smoothing, thread-level quantization, and a dual-layer accumulation method, significantly accelerating attention operations while maintaining unchanged end-to-end metrics.
π― What it does: Proposes SAH-Drive, which integrates rule-based planning PDM-Closed with diffusion learning planning to form a context-aware hybrid trajectory planner;
Sample Efficient Demonstration Selection for In-Context Learning
Kiran Purohit (Indian Institute of Technology), Avishek Anand (Delft University of Technology)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A sample-efficient demonstration subset selection algorithm named CASE is proposed for selecting the optimal example set in in-context learning (ICL) of large language models.
π― What it does: This paper proposes a score-adaptive noise injection framework based on diffusion, called SSNI, which can dynamically adjust the noise level for each input sample, thereby better removing adversarial noise while preserving image semantics.
π― What it does: Designed and implemented SASSHA, an optimizer that enhances generalization performance through stable Hessian approximation and sharpness reduction mechanisms within a second-order optimization framework.
π― What it does: Proposed a scalable relative and additive approximation algorithm that uses multiple HSTs and a dynamic weighted bi-colored nearest point (BCP) data structure to compute the p-Wasserstein distance and its (p,k)-RPW variant;
Scalable First-order Method for Certifying Optimal k-Sparse GLMs
Jiachang Liu (Cornell University), Andrea Lodi (Cornell University)
CodeOptimizationComputational EfficiencyTabular
π― What it does: This paper studies the optimality proof problem of sparse generalized linear models (GLMs) and proposes a method based on a first-order proximal gradient algorithm to address the perspective relaxation issue within the branch-and-bound (BnB) framework.
Scalable Gaussian Processes with Latent Kronecker Structure
Jihao Andreas Lin (Meta), Eytan Bakshy (Meta)
CodeOptimizationComputational EfficiencyTabular
π― What it does: A Gaussian Process method utilizing the latent Kronecker structure is proposed, allowing for efficient and accurate inference even in the presence of missing observations.
π― What it does: By learning molecular adaptive rotation and training non-equivariant diffusion models in aligned latent space, efficient 3D molecular generation is achieved with quality comparable to equivariant models.
π― What it does: Systematically train artificial neural networks of different architectures and data scales, evaluating their alignment with the primate ventral visual pathway (V1βIT) and behavior;
π― What it does: This paper proposes an unsupervised scene graph retrieval framework (SCENIR) based on graph autoencoders, using Graph Edit Distance (GED) as a deterministic criterion for assessing retrieval similarity.
π― What it does: Designed the SDMG (Smooth Diffusion Model for Graphs) framework, which utilizes a diffusion probability model for unsupervised representation learning on graph data, and enhances representation quality through a low-frequency encoder and multi-scale smooth loss.
π― What it does: Proposes the SDP-CROWN framework, which combines SDP and linear bound propagation to achieve scalable robustness verification of neural networks under ββ error;
SecEmb: Sparsity-Aware Secure Federated Learning of On-Device Recommender System with Large Embedding
Peihua Mai (National University of Singapore), Yan Pang (National University of Singapore)
CodeRecommendation SystemFederated LearningSafty and PrivacyComputational EfficiencyTabularSequential
π― What it does: A sparse encrypted federated recommendation system protocol named SecEmb is proposed, which utilizes point functions and function secret sharing (FSS) to ensure that users only download item embeddings they have interacted with, while securely aggregating sparse gradients on the server side, ensuring that the server cannot obtain any user's rated item indices or gradient information.
SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning
Jinpeng Chen (City University of Hong Kong), Sam Kwong (Lingnan University)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningMultimodality
π― What it does: A forgetting elimination framework SEFE for multimodal continual instruction tuning is proposed, which separates and simultaneously suppresses superficial forgetting and essential forgetting.
CodeGenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This study investigates the phenomenon of attention decay in the code generation process of large language models and proposes the Selective Prompt Anchoring (SPA) method, which enhances the impact of keywords in user prompts on model attention to improve generation quality.
π― What it does: A universal self-guided adaptive (SPA) framework is proposed, which can perform model adaptation on image, object, and pixel-level tasks without source data or changes in the training process.
π― What it does: Proposes the Self-Organizing Prototypes (SOP) method, which constructs dynamic prototypes using non-parametric Support Embedding (SE) for unsupervised visual feature learning, and designs the SOP-Masked Image Modeling (SOP-MIM) task.
π― What it does: This paper proposes a method that utilizes Singular Value Decomposition (SVD) to project model gradients, selecting the most important weight subspace and fine-tuning only less than 1% of the parameters, thereby achieving efficient forgetting of specific data or concepts.
SERENA: A Unified Stochastic Recursive Variance Reduced Gradient Framework for Riemannian Non-Convex Optimization
Yan Liu (Nankai University), Ruxin Wang (Shenzhen Institutes of Advanced Technology)
CodeOptimizationTabular
π― What it does: This paper proposes the SRVRG estimator, the SRGE unified estimator, and the SERENA framework, extending variance reduction methods to Riemannian geometric non-convex optimization, and provides theoretical convergence analysis and experimental validation.
Ruitao Pu (Sichuan University), Yuan Sun (Sichuan University)
CodeRetrievalMultimodality
π― What it does: A Streaming-media Hashing Retrieval (SHE) method is proposed, supporting parallel learning and retrieval of streaming multimodal data.
π― What it does: This paper proposes an optimized sampler named SITCOM, which utilizes diffusion models to achieve more efficient reconstruction in inverse problem solving through three consistency constraints.