2/20/2026
Learning under noisy supervision is governed by a feedback-truth gap
Two-timescale analysis showing feedback-truth gap governs learning under noisy supervision across neural networks and human studies.
2/20/2026
VAM: Verbalized Action Masking for Controllable Exploration in RL Post-Training -- A Chess Case Study
Verbalized Action Masking method for controllable exploration in RL post-training of LLMs with chess case study.
2/20/2026
A Residual-Aware Theory of Position Bias in Transformers
Residual-aware theoretical analysis explaining position bias in transformer attention mechanisms and cumulative attention rollout.
2/20/2026
Training Large Reasoning Models Efficiently via Progressive Thought Encoding
Progressive Thought Encoding method for efficient training of large reasoning models via parameter-efficient RL fine-tuning.
2/20/2026
On the Mechanism and Dynamics of Modular Addition: Fourier Features, Lottery Ticket, and Grokking
Mechanistic analysis of how two-layer networks learn Fourier features for modular addition with theoretical training dynamics explanation.
2/20/2026
ML-driven detection and reduction of ballast information in multi-modal datasets
Framework for detecting and reducing ballast information across structured, semi-structured, and unstructured multimodal datasets.
2/20/2026
Construction of a classification model for dementia among Brazilian adults aged 50 and over
Dementia classification model for Brazilian adults using variable selection and multivariable analysis from ELSI-Brazil dataset.
2/20/2026
Exact Certification of Data-Poisoning Attacks Using Mixed-Integer Programming
Mixed-integer programming framework providing sound and complete certification guarantees for worst-case data poisoning attacks.
2/20/2026
Beyond Message Passing: A Symbolic Alternative for Expressive and Interpretable Graph Learning
Symbolic alternative to GNNs with improved expressivity beyond 1-Weisfeiler-Lehman barrier and fine-grained interpretability.
2/20/2026
Neural Proposals, Symbolic Guarantees: Neuro-Symbolic Graph Generation with Hard Constraints
NSGGM: neuro-symbolic framework for molecule generation combining neural proposals with symbolic constraints for controllability.
2/20/2026
Multi-Agent Lipschitz Bandits
Communication-free decentralized multi-agent bandit protocol with Lipschitz-structured action spaces and hard collision constraints.
2/20/2026
A Unified Framework for Locality in Scalable MARL
Unified framework for exploiting locality in scalable multi-agent RL with relaxed conditions on exponential decay property.
2/20/2026
Early-Warning Signals of Grokking via Loss-Landscape Geometry
Studies grokking transition via loss-landscape geometry on sequence-learning tasks SCAN and Dyck-1 using commutator defect metrics.
2/20/2026
Fail-Closed Alignment for Large Language Models
Fail-closed alignment design principle for robust LLM safety through redundant refusal mechanisms across latent features.
2/20/2026
Discovering Universal Activation Directions for PII Leakage in Language Models
UniLeak: mechanistic interpretability framework identifying universal activation directions that trigger PII leakage in language models.
2/20/2026
Dynamic Delayed Tree Expansion For Improved Multi-Path Speculative Decoding
Dynamic Delayed Tree Expansion improves multi-path speculative decoding for faster LLM token sampling verification.
2/20/2026
Action-Graph Policies: Learning Action Co-dependencies in Multi-Agent Reinforcement Learning
Action Graph Policies: framework for learning action dependencies in multi-agent reinforcement learning to coordinate agent behavior.
2/20/2026
WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning
WS-GRPO: weakly-supervised training method for language models on complex reasoning with improved rollout efficiency.
2/20/2026
Transforming Behavioral Neuroscience Discovery with In-Context Learning and AI-Enhanced Tensor Methods
Using in-context learning and AI-enhanced tensor methods to automate behavioral neuroscience discovery pipelines.
2/20/2026
Forecasting Anomaly Precursors via Uncertainty-Aware Time-Series Ensembles
Uncertainty-aware time-series ensemble method for proactive anomaly prediction providing early warning signals before anomalies occur.
2/20/2026
Spatio-temporal dual-stage hypergraph MARL for human-centric multimodal corridor traffic signal control
Multi-agent reinforcement learning framework using spatio-temporal hypergraphs for human-centric multimodal traffic signal control.
2/20/2026
AdvSynGNN: Structure-Adaptive Graph Neural Nets via Adversarial Synthesis and Self-Corrective Propagation
Graph neural network architecture with adversarial synthesis and contrastive learning for resilient node classification under structural noise.
2/20/2026
Adam Improves Muon: Adaptive Moment Estimation with Orthogonalized Momentum
NAMO optimizer combining Adam's adaptive moments with Muon's orthogonalized momentum for efficient large language model training.
2/20/2026
MeGU: Machine-Guided Unlearning with Target Feature Disentanglement
Machine unlearning approach using target feature disentanglement to balance privacy and model utility for right to be forgotten compliance.
2/20/2026
A Locality Radius Framework for Understanding Relational Inductive Bias in Database Learning
Formal framework defining locality radius to determine when multi-hop reasoning is necessary for foreign key discovery in relational databases.
2/20/2026
FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment
Federated fine-tuning approach for LLMs using low-rank Gram matrices and Procrustes alignment to enable collaborative adaptation without data sharing.
2/20/2026
Operationalization of Machine Learning with Serverless Architecture: An Industrial Operationalization of Machine Learning with Serverless Architecture: An Industrial Implementation for Harmonized System Code Prediction
Serverless MLOps framework orchestrating complete ML lifecycle with event-driven pipelines for model-agnostic inference and rapid deployment.
2/20/2026
Online Learning with Improving Agents: Multiclass, Budgeted Agents and Bandit Learners
Online learning framework for multiclass settings where agents can improve features to achieve better labels with budgeted costs.
2/20/2026
VP-VAE: Rethinking Vector Quantization via Adaptive Vector Perturbation
Vector Perturbation VAE decouples representation learning from discretization by eliminating explicit codebook dependency in generative modeling.
2/20/2026
When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer
Analysis of underfitting challenges in multi-expert learning to defer systems where classifiers abstain and defer to experts.
2/20/2026
Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization
Gradient-free zeroth-order optimization method for memory-efficient fine-tuning of large-scale models via subspace gradient orthogonalization.
2/20/2026
In-Context Learning in Linear vs. Quadratic Attention Models: An Empirical Study on Regression Tasks
Empirical study comparing how transformers and linear attention models perform in-context learning on regression tasks, analyzing MSE, convergence, and generalization.
2/20/2026
Continual uncertainty learning
Deep reinforcement learning with domain randomization for robust control of mechanical systems with multiple uncertainties and nonlinear dynamics.
2/20/2026
SoftDTW-CUDA-Torch: Memory-Efficient GPU-Accelerated Soft Dynamic Time Warping for PyTorch
Open-source PyTorch library implementing GPU-accelerated Soft Dynamic Time Warping with improved memory efficiency and numerical stability.
2/20/2026
CounterFlowNet: From Minimal Changes to Meaningful Counterfactual Explanations
ArXiv paper on CounterFlowNet for generating multiple minimal counterfactual explanations for tabular data with heterogeneous features.
2/20/2026
Structured Prototype-Guided Adaptation for EEG Foundation Models
ArXiv research on adapting EEG foundation models with limited supervision using prototype-guided fine-tuning for clinical settings.
2/20/2026
Learning a Latent Pulse Shape Interface for Photoinjector Laser Systems
ArXiv paper applying Wasserstein Autoencoders to control laser pulse shapes in Free-Electron Lasers via differentiable latent interface.
2/20/2026
Unified Latents (UL): How to train your latents
ArXiv research on Unified Latents framework combining diffusion priors with diffusion decoders for efficient latent representation learning.
2/20/2026
RLGT: A reinforcement learning framework for extremal graph theory
ArXiv paper proposing RL framework for extremal graph theory problems, extending Deep Cross-Entropy methods to combinatorial optimization.
2/20/2026
LexiSafe: Offline Safe Reinforcement Learning with Lexicographic Safety-Reward Hierarchy
LexiSafe: Offline safe reinforcement learning method using lexicographic hierarchy to prevent safety violations in cyber-physical systems.
2/20/2026
Flickering Multi-Armed Bandits
Research paper introducing Flickering Multi-Armed Bandits framework where available actions change dynamically, modeled via random graph processes.
2/20/2026
The Sound of Death: Deep Learning Reveals Vascular Damage from Carotid Ultrasound
Machine learning framework using deep learning on carotid ultrasound videos to detect vascular damage for cardiovascular disease risk assessment.
2/20/2026
SubQuad: Near-Quadratic-Free Structure Inference with Distribution-Balanced Objectives in Adaptive Receptor framework
Subquadratic structure inference pipeline for immune repertoire analysis combining retrieval, affinity kernels, and multimodal fusion.
2/20/2026
From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection
Test-time training method using prompts for Graph Neural Network out-of-distribution detection without accessing training data.
2/20/2026
Shortcut learning in geometric knot classification
Study of shortcut learning in neural networks applied to knot topology classification with applications to protein folding and polymer physics.
2/20/2026
A feature-stable and explainable machine learning framework for trustworthy decision-making under incomplete clinical data
ML framework for biomedical data emphasizing feature stability and interpretability under incomplete/missing data for trustworthy clinical decision-making.
2/20/2026
MDP Planning as Policy Inference
Formulates MDP planning as Bayesian policy inference with return-based optimality prior for discrete action domains.
2/20/2026
Retrospective In-Context Learning for Temporal Credit Assignment with Large Language Models
Uses LLMs for temporal credit assignment in self-evolving agents via retrospective in-context learning from sparse feedback.
2/20/2026
LORA-CRAFT: Cross-layer Rank Adaptation via Frozen Tucker Decomposition of Pre-trained Attention Weights
Introduces CRAFT, parameter-efficient fine-tuning method using Tucker decomposition on frozen pre-trained transformer attention weights.
2/20/2026
The Anxiety of Influence: Bloom Filters in Transformer Attention Heads
Identifies transformer attention heads functioning as membership testers/bloom filters across GPT-2 and Pythia models.