ICML 2026 Papers — Page 45
International Conference on Machine Learning · 6554 papers
PPDL: LLM-Based Flows as Probabilistic Programs
Louis Mandel (IBM), Martin Hirzel (IBM)
Federated LearningExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes PPDL, a language that combines prompt programming with probabilistic programming, enabling explicit tracking of uncertainty and inference scaling within large language model (LLM) flows.
PPI Candidate Ranking: Large-Scale Evaluation of a Domain Knowledge–Guided Pipeline
Maria Emilia Russo (Sapienza University), Roberto Capobianco (Sony AI)
Explainability and InterpretabilityRepresentation LearningDrug DiscoveryTransformerLarge Language ModelContrastive LearningGraphBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose an explainability-driven protein interaction candidate ranking framework based on embedding activation regions derived from known interactions, and further enhance candidate prioritization by incorporating multi-source re-ranking (interaction scores, structural feasibility, semantic similarity, and LLM).
PPT-Eval: A Benchmark for Computer-Use Agents on PowerPoint Tasks
Apurva Gandhi (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Large Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Created the PPT-EVAL benchmark, covering more than 120 PowerPoint tasks of varying difficulty levels, and designed a hierarchical evaluation rubric with a natural language feedback mechanism.
PRAC: Principal-Random Subspace for LLM Activation Compression and Memory-Efficient Training
Yanyi Li (Peking University), Cong Fang (Peking University)
CompressionComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose a compression method called PRAC for activation compression, achieving memory compression during LLM training by decomposing activations into two parts: a principal subspace and a random subspace.
Practical and Optimal Algorithm for Linear Contextual Bandits with Rare Parameter Updates
Sanghoon Yu (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement LearningAuto EncoderContrastive LearningTabularTime Series
🎯 What it does: This paper addresses the online linear contextual bandit problem, proposing two practical algorithms, BLCE-G and BLCE, in a scenario where only sparse parameter updates (O(log log T) times) are allowed rather than updates after each round. The latter further eliminates the dependence on G-optimal design and provides an extension called BGLE for generalized linear contextual bandits.
Practical and Scalable Hamiltonian Monte Carlo Without the Metropolis Test
Jakob Robnik (University of California), Uros Seljak (Lawrence Berkeley National Laboratory)
OptimizationFederated LearningComputational EfficiencyHyperparameter SearchData-Centric LearningScore-based ModelTabularTime SeriesSequentialBiomedical DataBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes an automatic step size adjustment scheme, enabling Hamiltonian Monte Carlo, Langevin Monte Carlo, and MCLMC without Metropolis-Hastings correction to function as black-box samplers without manual parameter tuning.
Practical Mechanism for Fault-Tolerant Spiking Neural Networks via Simple Input Control Based on Learnable Fragmentation
Hyun-Jong Lee (Kwangwoon University), Jae-Han Lim (Kwangwoon University)
ClassificationComputational EfficiencyRepresentation LearningSpiking Neural NetworkAuto EncoderContrastive LearningImageAudio
🎯 What it does: Propose a mechanism to improve the fault tolerance performance of hardware-implemented spiking neural networks (SNNs) by using only a learnable input fragmentation control.
PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts
Qinfeng Li (Zhejiang University), Xuhong Zhang (Zhejiang University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: In untrustworthy deployment environments, system prompts of LLM agents are vulnerable to being stolen and reused on other models, leading to IP loss. The paper proposes PragLocker, a black-box prompt protection scheme, which achieves non-transferable encryption of prompts through a two-phase process, preventing their reuse after being stolen.
Precision-Induced Miscalibration: Understanding and Correcting Confidence Distortion in Quantized Neural Networks
Jiawei Gu (Great Bay University), Yanpeng Sun (Singapore University of Technology and Design)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerScore-based ModelContrastive LearningImageText
🎯 What it does: Systematically study the impact of low-precision arithmetic (FP16, BF16, INT8, etc.) on the confidence of neural network predictions, and propose an adaptive temperature scaling method (PACS) based on hardware precision constants to achieve confidence calibration.
Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences
Neehal Tumma (MIT), Daniela Rus (MIT)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningTextSequential
🎯 What it does: This paper introduces preconditioned DeltaNet (PDN) and its derivative models (PGDN, PKDA) by adding preconditioning to linear recurrent layers such as DeltaNet, achieving more accurate least squares approximation and improvements in long context sequence modeling tasks.
Preconditioning Neural Tangent Kernel for Adaptive Optimization
Xiyuan Yang (University of Illinois Urbana Champaign), Jingrui He (University of Illinois Urbana Champaign)
OptimizationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose Optimizer Aware Kernel (OAK), which statically estimates the preconditioners of adaptive optimizers such as Adam, incorporates their effects into the NTK framework, and theoretically analyzes the kernel degradation during the fine-tuning process.
Predictable Compression Failures: Order Sensitivity and Information Budgeting for Evidence-Grounded Binary Adjudication
Leon Chlon (University of Oxford), MarcAntonio Awada (Harvard University)
ClassificationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: This paper studies the prediction variance caused by the sensitivity of Transformer to evidence order in binary adjudication tasks with exchangeable evidence, and proposes an answer/abstain decision mechanism based on an information budget.
Predicting evolutionary rate as a pretraining task improves genome language model representations
Micaela Elisa Consens (University of Toronto), Alex Xijie Lu
Representation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningBiomedical DataBenchmark
🎯 What it does: Two novel pre-training tasks for predicting evolutionary rates were constructed, and they were used to train the Gamba series of genomic language models.
Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction
Ziyao Tang (Fudan University), Jingjing Chen (Fudan University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a KV cache eviction method based on long-term utility (LU-KV), achieving head-level cache budget allocation through global combinatorial optimization.
Predicting Large Model Test Losses with a Noisy Quadratic System
Chuning Li (University of Toronto), Chris J. Maddison (University of Toronto)
OptimizationLarge Language ModelTabular
🎯 What it does: A predictive model is proposed that can estimate the pre-training loss of large models based on model size (N), batch size (B), and the number of weight update steps (K). This is the first loss prediction model capable of handling varying batch sizes.
Predicting the Emergence of Induction Heads in Language Model Pretraining
Tatsuya Aoyama (Georgetown University), Nathan Schneider (Georgetown University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Studied the timing and driving factors of the emergence of inductive heads (IH) during the pre-training process of language models, focusing on the impact of batch size, context size, and the frequency and reliability of bigrams on IH formation.
Predicting the Order of Upcoming Tokens Improves Language Modeling
Zayd Muhammad Kawakibi Zuhri (Mohamed Bin Zayed University of Artificial Intelligence), Alham Fikri Aji (Mohamed Bin Zayed University of Artificial Intelligence)
Computational EfficiencyRepresentation LearningData-Centric LearningAI Code AssistantTransformerLarge Language ModelContrastive LearningTextSequential
🎯 What it does: Proposed a Token Order Prediction (TOP) auxiliary training objective, training the model to sort future tokens in sequence rather than making exact predictions
Predicting What Matters: Robust Generalist Robot Policy Learning via Future Semantic Mask
Yunfan Lou (National University of Singapore), Shanghang Zhang (Beijing Academy of Artificial Intelligence)
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelFlow-based ModelAuto EncoderWorld ModelImageVideo
🎯 What it does: A Mask World Model (MWM) is studied, which improves the robustness of robot control by predicting future semantic masks instead of RGB videos.
Prediction-Powered Adaptive Inference with Pretrained AI Models for Contextual Bandits
Gabriel Sargent (University of North Carolina at Chapel Hill), Yufeng Liu (University of Michigan)
Recommendation SystemReinforcement LearningTabularTime Series
🎯 What it does: Propose the Prediction‑Powered Adaptive Inference (PPAI) estimator, which integrates pre-trained AI predictions, unlabeled data, and adaptive experimental data to achieve statistical inference.
Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts
Guangyi Zhang (Zhejiang University), Osvaldo Simeone (Northeastern University London)
Domain AdaptationAnomaly DetectionTransformerReinforcement LearningContrastive LearningImageText
🎯 What it does: Proposed a semi-supervised risk monitoring method called PPRM, which combines synthetic labels generated by prediction with a small number of real labels to achieve real-time monitoring of the performance of deployed models.
Predictive Prefetching for Retrieval-Augmented Generation
Wuyang Zhang (University of Massachusetts Boston), Shichao Pei (University of Massachusetts Boston)
GenerationRetrievalComputational EfficiencyTransformerReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose an asynchronous prefetching framework that predicts and prefetches required information in advance during the generation process, using a retrieval predictor, context monitor, and query generator to reduce retrieval blocking.
Predictive variational inference: Learn the predictively optimal posterior distribution
Jinlin Lai (University of Massachusetts Amherst), Yuling Yao (University of Texas Austin)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningScore-based ModelImageTextTabularBenchmarkPhysics Related
🎯 What it does: A Predictive Variational Inference (PVI) framework is proposed, aiming to sample from the optimal posterior distribution such that the posterior predictive distribution is as close as possible to the true data-generating process.
Preference Goal Tuning: Post-Training as Latent Control for Frozen Policies
Guangyu Zhao (Peking University), Yitao Liang (Peking University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringContrastive LearningTextSequentialBenchmark
🎯 What it does: Achieve fine-tuning in later stages by optimizing the potential goal vector rather than network parameters on pre-trained goal-conditioned policies
Preference-based Antibody Expression Ranking: Scaling with Large-scale Weak Supervision
Josh Qixuan Sun (Kisoji Biotechnology Inc), David Young (Kisoji Biotechnology Inc)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBiomedical Data
🎯 What it does: Propose a unified antibody expression ranking framework based on preference learning, combining scarce expression labels with a large number of weak positive samples generated by immune responses.
Preference-Calibrated Optimization with Score-Level Distribution Alignment for Text-to-Image Diffusion Model Unlearning
Xiuyuan Wang (Zhejiang University), Xiaolin Zheng (Zhejiang University)
GenerationOptimizationSafty and PrivacyReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Propose a new text-to-image diffusion model unlearning framework called POSDA, which aligns the model's strategy directly with the forgetting objective through reinforcement learning, while using score-level distribution alignment to ensure the generation quality of retained concepts.
Preference-Enhanced Reinforcement Learning for Pluralistic Image Inpainting
Peng Zhou (Fudan University), Yun Xiong (Fudan University)
RestorationGenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes an online reinforcement learning framework that optimizes policies for text-guided image inpainting through GRPO and DPO, achieving diversity and aesthetic alignment.
Preference-Modulated Structural Attention for Multi-Objective Combinatorial Optimization
Rongsheng Jia (Nanjing University of Science and Technology), Jian Cheng (Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningGraphBenchmark
🎯 What it does: This paper proposes a lightweight neural solver, PMSA, which integrates node and edge features to address multi-objective combinatorial optimization problems.
Prefix-Cache-Aware Data Reordering for LLM-Augmented Database Analytics
Yingze Li (Harbin Institute of Technology), Bingsheng He (National University of Singapore)
OptimizationData-Centric LearningLarge Language ModelPrompt EngineeringTabularFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose a relation data reordering scheme based on prefix caching (SOLO), which adjusts row and column order to maximize prefix sharing in LLMs, thereby improving prefill throughput;
Prescriptive Scaling Reveals the Evolution of Language Model Capabilities
Hanlin Zhang (Harvard University), Sham M. Kakade (Stanford University)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Investigated the upper limit of downstream task performance (capability boundary) achievable after standard fine-tuning, given a pre-trained computational budget, and provided a method to predict achievable performance based on the budget.
Preserve-Then-Quantize: Balancing Rank Budgets for Quantization Error Reconstruction in LLMs
Yoonjun Cho (Yonsei University), Albert No (Yonsei University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose a structured residual reconstruction (SRR) framework with a Preserve-Then-Quantize approach, designed to balance preserving low-rank structures and compensating quantization errors under a fixed rank budget;
Preserving Plasticity in Continual Learning via Dynamical Isometry
Andries Rosseau (Vrije Universiteit Brussel), Ann Nowe (Vrije Universiteit Brussel)
OptimizationRepresentation LearningMeta LearningSupervised Fine-TuningReinforcement LearningContrastive LearningImageVideo
🎯 What it does: Propose to maintain plasticity in continual learning by dynamic isometry (keeping the hierarchical Jacobian eigenvalues close to 1), and present isometry regularization along with the AdamO optimizer;
Pressure Reveals Character: Behavioural Alignment Evaluation at Depth
Nora Petrova (Prolific AI Research), John Burden (Prolific AI Research)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Conducted multi-round alignment evaluation of large-scale language models in realistic scenarios, constructing 904 evaluation scenarios involving conflicting instructions, tool access, and escalation adversarial situations.
Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning
Huihan Liu (University of Texas at Austin), Yuke Zhu
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageVideoTextMultimodality
🎯 What it does: The study investigates the forgetting phenomenon in pre-trained vision-language-action (VLA) models during continual learning, and explores their robustness to experience replay through experiments.
PretrainZero: Reinforcement Active Learning on Pretraining Data
Xingrun Xing (Institute of Automation, Chinese Academy of Sciences), Debing Zhang (Xiaohongshu Inc.)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextChain-of-Thought
🎯 What it does: Propose a reinforcement learning-based active learning framework called PretrainZero, which utilizes self-supervised reward signals to perform full-process pretraining and subsequent RL fine-tuning of LLMs on large-scale pretraining corpora (Wikipedia).
PRIM：Cooperative Dynamic Token Compression for Efficient Large Multimodal Models
Song Li (Beijing University of Posts and Telecommunications), yongping xiong
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Propose the PRIM framework, which achieves collaborative dynamic token compression for large multimodal models (LMMs), significantly reducing the number of audiovisual tokens and computational load without sacrificing inference accuracy.
Primal-Spectral Generative Modeling: Fast Analytical Generation via Pseudoinverse Lévy Inversion
Zhiyuan Ouyang (East China Normal University), Junchi Yan (Shanghai Jiao Tong University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Developed a primitive-spectral generation framework called PriSpecNet, which maps probability distributions to continuous feature functions and achieves one-time analytical sampling by utilizing the pseudo-inverse Levy inverse solution (PiLI);
Principle-Evolvable Scientific Discovery via Uncertainty Minimization
Yingming Pu (Zhejiang University), Hongyu Chen (Westlake University)
OptimizationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringGaussian SplattingTextTabularBenchmarkPhysics Related
🎯 What it does: Propose the PIEVO framework, which treats scientific discovery as a process of Bayesian optimization over an evolvable space of scientific principles, enabling LLM-driven scientific agents to actively evolve principles and guide hypothesis search.
Principled Confidence Estimation for Deep Computed Tomography
Matteo Gätzner, Johannes Kirschner (ETH Zürich)
Mixture of ExpertsDiffusion modelScore-based ModelBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a confidence interval construction method with theoretical coverage guarantees in deep learning CT reconstruction based on the sequential likelihood mixture framework.
Principled RL for Flow Matching Emerges from the Chunk-level Policy Optimization
Yifu Luo (Tsinghua University), Xueqian Wang (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes Group Chunking Policy Optimization (GCPO), a chunk-level reinforcement learning method for post-training text-to-image flow matching, which addresses the inaccurate advantage attribution problem in traditional GRPO.
Principled SVD-based Delta Compression via Quantization Error Minimization
Boya Xiong (Shanghai University of Finance and Economics), Yun Chen (Shanghai University of Finance and Economics)
CompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Provide a mixed-precision SVD compression scheme for dense high-dimensional incremental weights generated by customized fine-tuning of LLMs in multi-tenant scenarios, which can minimize quantization error while meeting a given bit budget.
Principled Synthetic Data Enables the First Scaling Laws for LLMs in Recommendation
Benyu Zhang (Meta), Hong Yan (Meta)
Data SynthesisRecommendation SystemTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphSequentialRetrieval-Augmented Generation
🎯 What it does: Propose a hierarchical synthetic data framework to generate high-quality recommendation data without position and exposure bias, used for LLM continual pre-training, achieving predictable scaling laws in the recommendation domain.
Prior Diffusiveness and Regret in the Linear-Gaussian Bandit
Yifan Zhu (Stanford University), Benjamin Van Roy (Stanford University)
OptimizationReinforcement LearningContrastive LearningGaussian SplattingStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a new Bayesian burn-in and long-term uniform error separated upper bound for posterior sampling (Thompson sampling) in linear Gaussian Bayesian bandits, proving that the upper bound can reach ~O(σ√T+dr√Tr(Σ)), and provides the corresponding lower bound to show that this separation is unavoidable.
Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped Language Models
Jonathan Williams (Princeton University), Esin Tureci (Princeton University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the RLTT (Reward Latent Thought Trajectories) framework, which assigns rewards across the entire latent reasoning trajectory in the Loop language model, rather than only rewarding the terminal state;
Prioritized Model Experience Replay
Muxi Tao (Tsinghua University), Yuxing Han (Tsinghua University)
Reinforcement LearningWorld ModelTabularTime SeriesSequential
🎯 What it does: Propose and implement a Prioritized Model Experience Replay (PMER) mechanism, which prioritizes replaying transitions with high prediction errors during the model learning phase, thereby enhancing the model's fitting capability in regions relevant to the current policy.
Priority-Aware Shapley Value
Kiljae Lee (Ohio State University), Yuan Zhang (Ohio State University)
OptimizationExplainability and InterpretabilityImageTabular
🎯 What it does: Proposes Priority-Aware Shapley Value (PASV), which simultaneously considers hard priority constraints and soft priority weights in data value assessment and feature attribution tasks, forming a unified attribution framework.
Prism-MoE: Efficient Dense-to-MoE Conversion for Visual Autoregressive Generation
Ying Li (Westlake University), Huan Wang (Westlake University)
GenerationTransformerMixture of ExpertsAuto EncoderContrastive LearningImage
🎯 What it does: Efficiently convert a pre-trained dense visual autoregressive model into a sparse MoE model, significantly reducing inference costs;
PRISM: Demystifying Retention and Interaction in Mid-Training
Bharat Runwal (IBM Research), Rameswar Panda (IBM Research)
Computational EfficiencyRepresentation LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextSequentialBenchmark
🎯 What it does: Studies the design and effectiveness of mid-training in large language models, systematically evaluating the impact of different data mixing, evaluation criteria, domain interaction, and subsequent RL on model reasoning and long-term performance.
PRISM: Distribution-free Adaptive Computation of Matrix Functions for Accelerating Neural Network Training
Shenghao Yang (International Computer Science Institute), Michael W. Mahoney (International Computer Science Institute)
OptimizationComputational EfficiencyImageText
🎯 What it does: Propose the PRISM framework, which accelerates the iterative computation of matrix functions (such as square roots, inverse roots, and polarization decomposition) through adaptive polynomial approximation and randomized projection, thereby improving the training efficiency of neural networks.
PRISM: Gauge-Invariant Tangent-Space Differentially Private LoRA
Shihao Wang (Ohio State University), Xueru Zhang (Ohio State University)
OptimizationSafty and PrivacyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a differential privacy mechanism called PRISM for LoRA fine-tuning, addressing the problem of noise amplification caused by the decomposition being non-identifiable.
PRISM: Learning Realistic Depth via Physics-Grounded Noise Disentanglement with Semantic-Geometric Collaboration
Xiujian Liang (Fudan University), Jianhua Sun (Shanghai JiaoTong University)
Data SynthesisDepth EstimationRobotic IntelligenceConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImagePoint CloudPhysics Related
🎯 What it does: Propose the PRISM framework, which decomposes depth noise into two modes—perception failure and measurement inaccuracy—through physical mechanisms, and generates more realistic noise by leveraging 3D vision foundation models for semantic-geometric collaboration;
PRISM: Parallel Residual Iterative Sequence Model
Jie Jiang (Tencent), Zhouchen Lin (Peking University)
Recommendation SystemComputational EfficiencyTransformerMixture of ExpertsContrastive LearningSequential
🎯 What it does: Propose the PRISM architecture to achieve parallel Rank-L state updates, balancing efficiency and expressiveness.
PRISM: Perception Reasoning Interleaved for Sequential Decision Making.
Mohamed Salim AISSI, Nicolas THOME
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose the PRISM framework, which achieves tight coupling between perception and reasoning through dynamic question-answering between Vision-Language Models (VLM) and Large Language Models (LLM), significantly enhancing decision-making performance in multimodal tasks.
Prism: Spectral-Aware Block-Sparse Attention
Xinghao Wang (Fudan University), Xipeng Qiu (Fudan University)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsVideoTextRetrieval-Augmented Generation
🎯 What it does: Propose Prism, a training-agnostic dual-spectral block sparse attention framework;
PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization
Ying Tang (Huazhong University of Science and Technology), Wei Yang (Huazhong University of Science and Technology)
Knowledge DistillationRepresentation LearningTransformerMixture of ExpertsContrastive LearningImage
🎯 What it does: Propose a dual-stream conditional Mixture-of-Experts framework PRISM for knowledge distillation in multi-teacher visual foundation models, achieving model fusion through self-organized expert specialization.
PRISM: Training-Free Video Anomaly Detection via Intrinsic Statistical Modeling
YUANTONG CHEN, YanFeng Shang
Anomaly DetectionTransformerVision Language ModelAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes a training-free, real-time video anomaly detection framework called PRISM, which is based on a pre-trained multi-modal embedder. It statistically whitens textual anomaly descriptions to suppress common-mode noise, then aligns video features with the denoised anomaly semantic axis to generate anomaly scores.
PrivAct: Internalizing Contextual Privacy Preservation via Multi-Agent Preference Training
Yuhan Cheng (Duke University), Yiran Chen (Duke University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Studies how to internalize context-aware privacy consciousness into multi-agent LLM systems, achieving privacy-compliant generation through multi-agent preference training.
Privacy Amplification in Differentially Private Zeroth-Order Optimization with Hidden States
Eli Chien (National Taiwan University), Pan Li (Georgia Institute of Technology)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In the fine-tuning process of large language models, the study proposes a zeroth-order differential privacy optimization (ZOGD) framework and provides convergence privacy guarantees under hidden states.
Privacy-Aware Data Integration for Enhanced Quantile Inference under Heterogeneity
Leheng Cai (Tsinghua University), Shuyuan Wu (Shanghai University of Finance and Economics)
Federated LearningSafty and PrivacyReinforcement LearningContrastive LearningTabularBenchmark
🎯 What it does: This paper proposes a framework for quantile estimation and inference using multi-source heterogeneous data under the condition of local differential privacy (LDP);
Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection
Lei Wang (Griffith University), Yongsheng Gao (Griffith University)
Anomaly DetectionSafty and PrivacyConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: Propose two lightweight differentiable projection layers, Orthogonal Projection Layer (OPL) and Guided OPL (G-OPL), embedded into video anomaly detection models to remove features irrelevant to anomalies and sensitive information such as faces under weakly supervised settings.
Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch
Hyunwoo Kim (Nvidia), Yejin Choi (Nvidia)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringAuto EncoderGenerative Adversarial NetworkTextTabular
🎯 What it does: Create a fully synthetic private dataset of millions of records called PRIVASIS, and build a corpus called PRIVASIS‑SANITIZATION that can be instructed to de-identify text while preserving its usability. Subsequently, train a lightweight (≤4B) de-identification model.
Private and Stable Test-time Adaptation with Differential Privacy
Zefeng Li (University of British Columbia), Evan Shelhamer (University of British Columbia)
Domain AdaptationSafty and PrivacyConvolutional Neural NetworkTransformerContrastive LearningGaussian SplattingImage
🎯 What it does: Propose a differential privacy version of test-time adaptation (TTA) methods, utilizing per-sample gradient clipping and Gaussian noise to protect the privacy of test data, and investigate its impact on various TTA methods (Tent, EATA, SAR, DeYO, COME).
Private Learning with Public Feature Conditioning
Shuli Jiang (AWS Agentic AI), Nicolas Mayoraz (Google Research)
OptimizationSafty and PrivacyContrastive LearningTabularTime SeriesSequential
🎯 What it does: Propose a training method called Cond‑DP (and its improved version Switch‑Cond‑DP) that utilizes public features for conditioning in differential privacy regression tasks, and provide theoretical convergence guarantees.
Privately Fine-Tuned LLMs Preserve Temporal Dynamics in Tabular Data
Lucas Rosenblatt (Williams College), Natalia Ponomareva (Google Research)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Propose the PATH framework, which treats the entire user table as a privacy unit, and synthesizes longitudinal tabular data using a differentially private fine-tuned large language model.
PrivCode++: Latent-Conditioned Differentially Private Code Generation for Comprehensive Guarantees
Zheng Liu (Institute Of Automation Chinese Academy Of Sciences), Xiaochen Li (University Of North Carolina At Greensboro)
GenerationSafty and PrivacyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: Propose PrivCode++, achieving differential privacy code generation that is sensitive to both code and prompts;
PrivGate: Steering Contextual Integrity in LLMs via Latent Space Geometry
Runshan Hu (Huawei Technologies Co., Ltd.), Tieyan Li (Huawei Technologies Co., Ltd.)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the PrivGate framework, which achieves compute-on-demand inference path switching while ensuring context integrity (CI) by discovering privacy manifolds in the residual flow of LLMs and utilizing a linear detector for low-cost risk assessment of inputs.
Privileged Information Distillation for Language Models
Emiliano Penaloza (ServiceNow AI Research), Massimo Caccia (ServiceNow AI Research)
Federated LearningSafty and PrivacyComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes two methods for utilizing privileged information (PI) during training for language model distillation: π‑Distill (joint teacher-student objective) and On‑Policy Self‑Distillation (OPSD), achieving the retention of advantages brought by PI even when PI is not available during testing;
PRM-PBE: Process Reward Model for Reinforcement Learning in Programming-by-Example
Yue Fang (Peking University), Naijun Zhan (Peking University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a reinforcement learning framework based on a process reward model, utilizing process-level feedback to enhance the program synthesis capability of large language models in programming-by-example tasks.
Proact-VL: A Proactive VideoLLM for Real-Time AI Companions
Weicai Yan (Zhejiang University), Jianxun Lian (Microsoft Research Asia)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Proposed and implemented Proact-VL, a real-time proactive video large language model (VideoLLM), for game commentary (solo, co-commentary) and user guidance, capable of adaptively deciding when to speak and generating short and coherent sentences in continuous video streams.
ProAct: A Benchmark and Multimodal Framework for Structure-Aware Proactive Response
Xiaomeng Zhu (Hong Kong University of Science and Technology), Xuantang Xiong (Tencent)
Autonomous DrivingOptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the ProAct-75 visual proactive response benchmark (75 tasks, 91,581 step-level annotations and task graphs), and build the ProAct-Helper reference framework based on multimodal large language models to achieve structure-aware proactive action selection.
Proactive Defense Benchmark against Deepfake Generation
Joonhyuk Baek (Ulsan National Institute of Science and Technology), Hoki Kim (Chung Ang University)
GenerationData SynthesisOptimizationAdversarial AttackTransformerDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Proposed the first unified evaluation benchmark for systematically assessing the interference, robustness, and transferability of active defense against deepfake generation.
ProactiveLLM: Learning Active Interaction for Streaming Large Language Models
Junlong Tong (Shanghai Jiao Tong University), Xiaoyu Shen (Eastern Institute of Technology)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationAudio
🎯 What it does: Propose the ProactiveLLM framework, enabling large language models to proactively determine when to generate output in streaming input, using masked streaming modeling and synchronous privileged self-distillation to learn internal semantic sufficiency.
Probabilistic Bisection Algorithm Provably Achieves Exponential Convergence
Ganghua Wang (University of Chicago), Haifeng Xu (University of Chicago)
OptimizationContrastive LearningTabularBenchmarkStochastic Differential Equation
🎯 What it does: This paper provides a detailed analysis of the querying behavior of the probabilistic binary algorithm (PBA), and for the first time, proves that PBA exhibits an exponential convergence rate for the one-dimensional root-finding problem under a fixed true value, with a convergence speed comparable to that of the classical binary search in the noise-free case.
Probabilistic Modeling of Latent Agentic Substructures in Deep Neural Networks
Su Hyeong Lee (University of Chicago), Richard Ngo (Independent)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackAgentic AIMixture of Experts
🎯 What it does: Proposes a probabilistic model of internal sub-agents in neural networks, defining the reward and stability of composite agents through logarithmic pooling.
Probabilistic Performance Guarantees for Multi-Task Reinforcement Learning
Yannik Schnitzer (University of Oxford), David Parker (University of Oxford)
Explainability and InterpretabilityReinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper proposes a general post-training safety certification method that, given any multi-task reinforcement learning (MTRL) policy, provides high-confidence performance lower bounds under unknown task distributions.
Probabilistic Retrofitting of Learned Simulators
Cristiana Diaconu (Polymathic AI Collboration), Payel Mukhopadhyay (Polymathic AI Collboration)
Computational EfficiencyRepresentation LearningScore-based ModelContrastive LearningTabularTime SeriesPhysics Related
🎯 What it does: Propose a training-efficient probabilistic post-processing method to convert pre-trained deterministic PDE models into probabilistic models.
Probabilistic Robustness Certificates against Adversarial Attacks
Sara Taheri (LMU Munich), Majid Zamani (CU Boulder)
OptimizationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImageStochastic Differential Equation
🎯 What it does: This paper proposes a robustness certification framework based on probabilistic safety verification, utilizing barrier certificates (Barrier Certificates) to defend against adversarial attacks during training and testing in the stochastic gradient descent process;
Probabilistic Salient Object Ranking
Rongjin Guo (City University of Hong Kong), Rynson W. H. Lau (City University of Hong Kong)
RecognitionImage TranslationOptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose the ProbSOR framework, which transfers the significant object ranking problem from deterministic sequences to probability distributions, using Attention Shift Probability Tree (ASPT) to characterize the diversity of human gaze transitions.
Probabilistically-routed Bayesian Additive Spanning Trees for Learning on Constrained Domains
Abhisek Chakraborty (Eli Lilly and Company), Anirban Chakraborty (Medical University of South Carolina)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkMixture of ExpertsGaussian SplattingGraphTabularTime SeriesBenchmark
🎯 What it does: Propose Probabilistically-routed Bayesian Additive Spanning Trees (PR-BAST) on constrained domains, using probabilistic routing instead of traditional hard splits to achieve smooth and continuous regression functions.
Probability of Matching for Batch Multi-Objective Bayesian Optimization
Mingqian Li (Texas A&M University), Xiaoning Qian (Texas A&M University)
OptimizationTabularBenchmark
🎯 What it does: Propose a probability matching (Probability of Matching, POM) acquisition strategy for batch multi-objective Bayesian optimization, which directly maximizes the probability of matching the batch with the true Pareto set.
Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning
Wenhao Yu (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposes the RankTuner mechanism based on probability-entropy calibration, which adaptively reweights the importance of each token using relative ranking metrics to enhance fine-tuning performance.
Probably Approximately Correct Labels
Emmanuel Candes (Stanford University), Tijana Zrnic (Stanford University)
Recommendation SystemAnomaly DetectionData-Centric LearningProtein Structure PredictionConvolutional Neural NetworkTransformerLarge Language ModelMixture of ExpertsDiffusion modelScore-based ModelContrastive LearningImageTextBiomedical DataAudio
🎯 What it does: Propose a method to automatically generate a set of labels with guaranteed error upper bounds (PAC labeling) by utilizing AI predictions and uncertainty information, without making any assumptions about the AI model, and further reduce the cost of expert annotation through uncertainty calibration and multi-model routing.
ProbeLLM: Automating Principled Diagnosis of LLM Failures
Yue Huang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a framework called PROBELLM, which uses hierarchical Monte Carlo Tree Search (MCTS) to actively probe large language models (LLMs), discovering and aggregating interpretable failure modes.
Probing Cross-modal Information Hubs in Audio-Visual LLMs
Jihoo Jung (Korea Advanced Institute of Science and Technology), Joon Son Chung (Korea Advanced Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodalityAudio
🎯 What it does: This paper studies the internal mechanisms of audio-visual LLMs through causal tracing and a unimodal dominance framework, revealing that cross-modal information is mainly concentrated in the sink token, and proposes a training-free hallucination suppression method based on this finding.
Probing How Scalable Table Data Enhances General Long-Context Reasoning
Huaibing Xie (Tencent), Pluto Zhou (Tencent)
Data-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper studies and demonstrates the enhancing effect of structured tabular data on the long-context reasoning ability of large language models (LLMs), and proposes an expandable and verifiable pipeline for constructing tabular data (TableLong) used for RL fine-tuning.
Probing RLVR Training Instability through the Lens of Objective-Level Hacking
Yiming Dong (School Of Physics Peking University), Zheng Wang (Tongyi Lab)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: This study proposes the 'objective-level hacking' framework to explain the abnormal growth of the training-inference gap in large models MoE (Mixture-of-Experts) during long-term RLVR (reinforcement learning with verifiable rewards) training.
Probing the Geometry of Diffusion Models with the String Method
Elio Moreau (Capital Fund Management), Eric Vanden-Eijnden (Capital Fund Management)
GenerationProtein Structure PredictionDiffusion modelScore-based ModelRectified FlowImagePoint CloudMeshBiomedical DataStochastic Differential Equation
🎯 What it does: Compute continuous paths on pre-trained diffusion models using string methods to explore the geometry of distributions.
Probing the Inductive Bias of Neural Networks through Learning Random Cellular Automata
Jan Disselhoff (Johannes Gutenberg University Mainz), Michael Wand (Johannes Gutenberg University Mainz)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImagePhysics Related
🎯 What it does: This paper investigates the inductive biases that neural networks rely on for generalization over natural data by training neural networks on two-dimensional random cellular automata (CA).
Probing the Knowledge Boundary: An Interactive Agentic Framework for Deep Knowledge Extraction
Yuheng Yang (University of Illinois at Urbana-Champaign), Jiaxuan You (University of Illinois at Urbana-Champaign)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose an interactive agent framework that quantifies the knowledge boundaries of LLMs through multi-strategy exploration and a three-stage knowledge processing pipeline.
Problem Distributions as Tasks: Repurposing Meta Learning for Generative Combinatorial Optimization towards Multi-task Pretraining and Adaptation
Wenzheng Pan (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationMeta LearningGraph Neural NetworkMixture of ExpertsDiffusion modelGraph
🎯 What it does: This paper proposes a multi-task meta-learning framework called M GenCO, which treats different types of graph combinatorial optimization problems as meta-learning tasks, and utilizes a diffusion-based generative model to solve various graph COPs.
ProbMoE: Differentiable Probabilistic Routing for Mixture-of-Experts
Heng Zhao (University of Virginia), Zhe Zeng (University of Virginia)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: A probability-based Mixture-of-Experts routing framework called ProbMoE was constructed, which supports both fixed number of experts (Exactk) and dynamic number of experts (Dynamick), and achieves differentiable training through SAMPLE-based sampling and SIMPLE gradient estimation.
Procedural Generation Of Algorithm Discovery Tasks in Machine Learning
Alexander David Goldie (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
Data-Centric LearningMeta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelImageVideoTextTabularTime SeriesSequentialBenchmarkChain-of-Thought
🎯 What it does: Propose DiscoGen, a framework for discovering tasks through procedural generation algorithms, and build DiscoBench based on it to evaluate automatic algorithm discovery agents (ADA).
Procedural Pretraining: Warming Up Language Models with Abstract Data
Liangze Jiang (École Polytechnique Fédérale De Lausanne), Damien Teney (Idiap Research Institute)
Computational EfficiencyRepresentation LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequential
🎯 What it does: This paper proposes using abstract procedural data for a 'pre-pre-training' phase before the pre-training of large language models, aiming to enhance the model's algorithmic reasoning ability;
Process Reward Agents for Steering Knowledge-Intensive Reasoning
Jiwoong Sohn (ETH Zuerich), Michael Moor (ETH Zuerich)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Process Reward Agents (PRA), a framework that gradually rewards frozen medical reasoning models during inference, leveraging retrieval and reward evaluation for each step to improve the model's reasoning accuracy.
ProConMV: Provenance-Enabled Conceptual Framework for Interpretable Multi-View Diabetic Retinopathy Diagnosis
Xiaoling Luo (Shenzhen University), Chengliang Liu (University of Macau)
ClassificationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataChain-of-Thought
🎯 What it does: Proposes a multi-view diabetic retinopathy diagnosis framework named ProConMV, which achieves diagnosis by fusing lesion masks, clinical text, and multi-view images, and utilizing interpretable conceptual reasoning chains.
ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation
Yizheng Huang (Google DeepMind), Zi Wang (Google DeepMind)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelGaussian SplattingTextMultimodalityBenchmark
🎯 What it does: Propose the PROEVAL framework, which utilizes transfer learning to construct an efficient Gaussian Process prior, achieving both active failure detection and sample-efficient performance estimation.
Profiling the Irrational Agent: Cognitive Modeling of LLM Behaviors in Sequential Jailbreaks
Xikang Yang (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextSequentialBenchmark
🎯 What it does: This paper models the sequential jailbreak behavior of large language models (LLMs) as a multi-round decision-making process, proposes a contextualized Iowa Gambling Task (C-IGT) for behavioral induction, and uses the Generalized Rescorla-Wagner (GRW) framework to provide interpretable parameterization of the LLM's learning, perception, and decision-making processes.
Progressive Cramming: Reliable Token Compression and What It Reveals
Dmitrii Tarasov (FusionBrain Lab), Andrey Kuznetsov (FusionBrain Lab)
CompressionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: Proposes a Progressive Cramming method, which optimizes individual embeddings token-by-token to achieve perfect autoregressive reconstruction, and analyzes the optimization path through low-dimensional projection;
Progressive Graph Structure Adjustment for Homophily Shift Adaptation
Hongwei Wen (University of Sydney), Minglong Lei (Beijing University of Technology)
Domain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a homogeneity shift adaptation method based on progressive graph structure adjustment (PSAHS), which enhances the homogeneity of low-homogeneity nodes in the source graph in one go, predicts consistent pseudo-labels for structural refinement in the target graph through GNN and MLP, and achieves cross-domain representation alignment via adversarial training;
Projection-Free Algorithms for Minimax Problems
Khanh-Hung Giang-Tran (Cornell University), Nam Ho-Nguyen (University of Sydney)
OptimizationTextTabular
🎯 What it does: This paper proposes a unified dual dynamic smoothing framework, and designs three projection-free single-loop algorithms for solving convex-concave, non-convex-concave, and non-convex-strongly concave extremum problems.
ProjQ: Project-and-Quantize for Adapter-Aware LLM Compression
Wenya Yu (Central South University), Merouane Abdelkader DEBBAH
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a framework called ProjQ that combines post-training quantization (PTQ) with low-rank adaptation (LoRA), which can confine quantization noise within a low-rank subspace, making subsequent LoRA easier to correct;