ICML 2025

3,384 papers

Top Research Directions

Research directions most represented by ICML 2025 papers.

Shifting from Static Benchmarks to Dynamic Evaluation 37
Inference Time Compute Scaling for Complex Reasoning 30
Clean Accuracy Degradation in Machine Learning Defenses 29
Overcoming the Iterative Inference Latency of Diffusion Models 27
The Performance Cliff in Ultra Low Bit Quantization 27
The Computational Overhead of Generative Data Synthesis 23
Bypassing the Computational Bottleneck of Adversarial Training 23
Outlier Suppression for Low Bit Transformer Quantization 20
The Function Approximation Gap in Reinforcement Learning Theory 20
Understanding Training Data Memorization in Generative Models 18
Resolving Superposition with Sparse Autoencoders 17
Preserving Model Utility During Machine Unlearning 17
Adapting Graph Neural Networks to Evolving Topologies 17
Overcoming the Homophily Assumption in Graph Neural Networks 16
Over-Smoothing in Graph Neural Networks 16

Top Authors

AuthorPapers
Mihaela van der Schaar15
Dacheng Tao12
Guancheng Wan12
Li Shen12
Mang Ye12
Wenke Huang12
Sergey Levine11
Kun Zhang10
Quanquan Gu10
Sanmi Koyejo10
Bo Han9
Feng Liu9
Ion Stoica9
Junchi Yan9
Martin Vechev9
Ruixuan Li9
Shirui Pan9
Yu Cheng9
Jiang Bian8
Jun Zhu8

Impactful Papers

Papers from ICML 2025 that shaped research clusters.

Foundational

Influential

Papers

Title
LEMoN: Label Error Detection using Multimodal Neighbors
LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection
Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective
Let LLM Tell What to Prune and How Much to Prune
LETS Forecast: Learning Embedology for Time Series Forecasting
Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection
Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models
Leveraging Offline Data in Linear Latent Contextual Bandits
Leveraging Online Olympiad-Level Math Problems for LLMs Training and Contamination-Resistant Evaluation
Leveraging Per-Instance Privacy for Machine Unlearning
Leveraging Predictive Equivalence in Decision Trees
Leveraging Randomness in Model and Data Partitioning for Privacy Amplification
Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration
Leveraging Sparsity for Sample-Efficient Preference Learning: A Theoretical Perspective
LEVIS: Large Exact Verifiable Input Spaces for Neural Networks
Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries
LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations
LieRE: Lie Rotational Positional Encodings
LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning
Liger: Linearizing Large Language Models to Gated Recurrent Structures
LightGTS: A Lightweight General Time Series Forecasting Model
LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos
Lightspeed Geometric Dataset Distance via Sliced Optimal Transport
Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty
Lightweight-Mark: Rethinking Deep Learning-Based Watermarking
Lightweight Online Adaption for Time Series Foundation Model Forecasts
Lightweight Protocols for Distributed Private Quantile Estimation
LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning
Limitations of measure-first protocols in quantum machine learning
Linear $Q$-Learning Does Not Diverge in $L^2$: Convergence Rates to a Bounded Set
Linear Bandits with Partially Observable Features
Linear Contextual Bandits With Interference
Linear convergence of Sinkhorn's algorithm for generalized static Schrödinger bridge
Linearization Turns Neural Operators into Function-Valued Gaussian Processes
Linear Mode Connectivity between Multiple Models modulo Permutation Symmetries
Linear Q-Learning Does Not Diverge in L2: Convergence Rates to a Bounded Set
Linear Transformers as VAR Models: Aligning Autoregressive Attention Mechanisms with Autoregressive Forecasting
LineFlow: A Framework to Learn Active Control of Production Lines
LipsNet++: Unifying Filter and Controller into a Policy Network
LIVS: A Pluralistic Alignment Dataset for Inclusive Public Spaces
LlavaGuard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models
LLaVA-ReID: Selective Multi-image Questioner for Interactive Person Re-Identification
LLM Alignment as Retriever Optimization: An Information Retrieval Perspective
LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination
LLM-Augmented Chemical Synthesis and Design Decision Programs
LLM Data Selection and Utilization via Dynamic Bi-level Optimization
LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification
LLMScan: Causal Scan for LLM Misbehavior Detection
LLMs Can Reason Faster Only If We Let Them
LLMs can see and hear without any training

« prev Page 35/68 next »