2/20/2026
Self-Improving Skill Learning for Robust Skill-based Meta-Reinforcement Learning
Hierarchical meta-reinforcement learning approach using self-improving skills to handle noisy offline demonstrations.
2/20/2026
Rex: A Family of Reversible Exponential (Stochastic) Runge-Kutta Solvers
Reversible Runge-Kutta solvers for neural differential equations in generative models with improved numerical stability.
2/20/2026
Unlocking [CLS] Features for Continual Post-Training
Continual learning approach for foundation models addressing stability-plasticity trade-off during post-training on new classes/domains.
2/20/2026
Efficient Orthogonal Fine-Tuning with Principal Subspace Adaptation
Parameter-efficient fine-tuning method using orthogonal adaptation on principal subspaces for adapting large models efficiently.
2/20/2026
Oversmoothing, Oversquashing, Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning
Survey demystifying common beliefs about oversmoothing, oversquashing, heterophily and long-range tasks in graph neural networks.
2/20/2026
On the Design of KL-Regularized Policy Gradient Algorithms for LLM Reasoning
Analyzes KL-regularization design choices in policy gradient algorithms for LLM reasoning, comparing forward/reverse KL variants.
2/20/2026
Strict Subgoal Execution: Reliable Long-Horizon Planning in Hierarchical Reinforcement Learning
Strict Subgoal Execution method for hierarchical RL that improves long-horizon planning by validating subgoal feasibility.
2/20/2026
AXLearn: Modular, Hardware-Agnostic Large Model Training
AXLearn production system for scalable hardware-agnostic training of large models with modular software architecture.
2/20/2026
Continuous-Time Value Iteration for Multi-Agent Reinforcement Learning
Extends reinforcement learning to continuous-time systems using Hamilton-Jacobi-Bellman equations for irregular interaction frequencies.
2/20/2026
Watermarking Diffusion Language Models
First watermarking method for diffusion language models that generate tokens non-sequentially, addressing unique DLM challenges.
2/20/2026
Entropy After $\langle \texttt{/Think} \rangle$ for reasoning model early exiting
Analyzes overthinking in reasoning LLMs and proposes early exit using entropy after thinking tags to improve efficiency.
2/20/2026
Multimodal Prompt Optimization: Why Not Leverage Multiple Modalities for MLLMs
Extends prompt optimization to multimodal LLMs, proposing methods to optimize across text, images, video and other modalities.
2/20/2026
Contrastive Diffusion Alignment: Learning Structured Latents for Controllable Generation
Introduces ConDA, a contrastive learning layer for organizing diffusion model latent spaces to enable controllable generation.
2/20/2026
pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation
Proposes pi-Flow, a policy-based approach to improve few-step diffusion model distillation by predicting network-free policies.
2/20/2026
Accelerating Mobile Inference through Fine-Grained CPU-GPU Co-Execution
System for accelerating neural network inference on mobile devices through fine-grained CPU-GPU co-execution and synchronization optimization.
2/20/2026
LRT-Diffusion: Calibrated Risk-Aware Guidance for Diffusion Policies
LRT-Diffusion applies risk-aware sequential hypothesis testing to improve diffusion policy guidance for offline reinforcement learning.
2/20/2026
Semi-Supervised Preference Optimization with Limited Feedback
Semi-supervised preference optimization framework for LLM alignment using limited labeled paired feedback data.
2/20/2026
Efficient Reinforcement Learning for Large Language Models with Intrinsic Exploration
PREPO method improving data efficiency for LLM reinforcement learning with verifiable rewards using intrinsic exploration signals.
2/20/2026
Beyond Linear Surrogates: High-Fidelity Local Explanations for Black-Box Models
Model-agnostic local explanation method using MARS and N-ball sampling for high-fidelity black-box model interpretability.
2/20/2026
Online Robust Reinforcement Learning with General Function Approximation
Distributionally robust reinforcement learning approach using general function approximation for policy robustness under environment shift.
2/20/2026
Learning PDE Solvers with Physics and Data: A Unifying View of Physics-Informed Neural Networks and Neural Operators
Unified framework connecting physics-informed neural networks and neural operators for learning PDE solvers.
2/20/2026
Temporal Graph Pattern Machine
Temporal graph pattern machine for learning transferable representations in dynamic networks without restrictive assumptions.
2/20/2026
The Chicken and Egg Dilemma: Co-optimizing Data and Model Configurations for LLMs
Method for jointly optimizing data mixture and model architecture configurations during LLM training to avoid suboptimal individual choices.
2/20/2026
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Automated black-box pipeline for detecting unverbalized biases in LLM reasoning traces and chain-of-thought explanations.
2/20/2026
Logit Distance Bounds Representational Similarity
Analysis showing logit distance bounds representational similarity in discriminative models including autoregressive language models.
2/20/2026
HPMixer: Hierarchical Patching for Multivariate Time Series Forecasting
HPMixer model for long-term multivariate time series forecasting using hierarchical patching to capture periodic patterns and residuals.
2/20/2026
Goal Inference from Open-Ended Dialog
Framework for embodied AI agents to infer user goals from open-ended dialog using LLMs for efficient task accomplishment.
2/20/2026
Multi-View 3D Reconstruction using Knowledge Distillation
Knowledge distillation pipeline to compress Dust3r foundation model for efficient 3D reconstruction and visual localization.
2/20/2026
Efficient Context Propagating Perceiver Architectures for Auto-Regressive Language Modeling
Perceiver architecture for auto-regressive language modeling reducing attention complexity from quadratic to semi-linear.
2/20/2026
Enhancing Multilingual LLM Pretraining with Model-Based Data Selection
Model-based data filtering framework for multilingual LLM pretraining that identifies diverse, high-quality training samples.
2/20/2026
Simple Self Organizing Map with Vision Transformers
Combining Self-Organizing Maps with Vision Transformers to improve performance on smaller datasets through explicit inductive biases.
2/20/2026
Capturing Individual Human Preferences with Reward Features
Learning user-specialized reward models for reinforcement learning from human feedback to capture individual preference disagreement.
2/20/2026
A Unifying Framework for Robust and Efficient Inference with Unstructured Data
Framework for handling unstructured data feature extraction with neural networks while accounting for measurement bias in economic analysis.
2/20/2026
ReplaceMe: Network Simplification via Depth Pruning and Transformer Block Linearization
ReplaceMe: training-free depth pruning method replacing transformer blocks with linear operations for efficient model compression.
2/20/2026
LLM Fingerprinting via Semantically Conditioned Watermarks
LLM fingerprinting via semantically conditioned watermarks that survive finetuning and quantization without being easily detected.
2/20/2026
Beyond Needle(s) in the Embodied Haystack: Environment, Architecture, and Training Considerations for Long Context Reasoning
∞-THOR framework for long-horizon embodied AI tasks with Needle(s) in Embodied Haystack benchmark for testing long-context reasoning in agents.
2/20/2026
Persona-driven Simulation of Voting Behavior in the European Parliament with Large Language Models
Using persona-driven prompting to simulate European Parliament voting behavior with LLMs, addressing political bias in model responses.
2/20/2026
$\texttt{SPECS}$: Faster Test-Time Scaling through Speculative Drafts
SPECS: method for faster test-time scaling in LLMs through speculative drafts, balancing reasoning accuracy with user-facing latency.
2/20/2026
Bongard-RWR+: Real-World Representations of Fine-Grained Concepts in Bongard Problems
Bongard Problems benchmark using real-world images to test abstract visual reasoning and fine-grained concept identification in models.
2/20/2026
Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction
Proposes inference-time search algorithm that guides diffusion model sampling with side information for improved image reconstruction in inverse problems.
2/20/2026
LayerSync: Self-aligning Intermediate Layers
LayerSync regularizes diffusion models using their own intermediate layer representations to improve generation quality and training efficiency.
2/20/2026
Toward LLM-Supported Automated Assessment of Critical Thinking Subskills
Uses LLMs for automated assessment of critical thinking skills in educational contexts, addressing evaluation of evidence and claim reliability.
2/20/2026
Multimodal Wireless Foundation Models
Extends wireless foundation models to accept multiple input modalities for improved task performance and adaptation across varying conditions.
2/20/2026
Block-Recurrent Dynamics in Vision Transformers
Introduces Block-Recurrent Hypothesis explaining Vision Transformer depth as block-recurrent computational flow for mechanistic interpretation.
2/20/2026
Puzzle it Out: Local-to-Global World Model for Offline Multi-Agent Reinforcement Learning
Proposes world model approach for offline multi-agent RL using local-to-global puzzle solving to overcome conservative policies and improve generalization.
2/20/2026
Beyond Predictive Uncertainty: Reliable Representation Learning with Structural Constraints
Framework for treating representation reliability as a first-class property in machine learning, beyond traditional predictive uncertainty quantification.
2/20/2026
Laser interferometry as a robust neuromorphic platform for machine learning
Proposes implementing optical neural networks using linear optical resources and phase-shift encoding for neuromorphic machine learning hardware.
2/20/2026
Beyond In-Domain Detection: SpikeScore for Cross-Domain Hallucination Detection
SpikeScore method for detecting LLM hallucinations that generalizes across domains, addressing the gap in cross-domain hallucination detection for real-world deployment.
2/20/2026
Accelerating Large-Scale Dataset Distillation via Exploration-Exploitation Optimization
Proposes exploration-exploitation optimization for dataset distillation to compress large datasets into synthetic versions while maintaining model performance.
2/20/2026
Building Safe and Deployable Clinical Natural Language Processing under Temporal Leakage Constraints
Addresses temporal leakage in clinical NLP models for discharge planning, proposing methods to prevent overconfident predictions from deployment artifacts.