2/25/2026
Position-Aware Sequential Attention for Accurate Next Item Recommendations
Improves sequential attention for recommendation systems by integrating position information into attention mechanism beyond additive embeddings.
2/25/2026
Motivation is Something You Need
Proposes dual-model training framework inspired by neuroscience motivation states with alternating base and larger model activation.
2/25/2026
An Enhanced Projection Pursuit Tree Classifier with Visual Methods for Assessing Algorithmic Improvements
Extends projection pursuit tree classifier with visual diagnostic methods for high-dimensional multi-class classification problems.
2/25/2026
A Benchmark for Deep Information Synthesis
DEEPSYNTH benchmark evaluates LLM-based agents on complex multi-source information synthesis tasks beyond fact retrieval.
2/25/2026
Scaling State-Space Models on Multiple GPUs with Tensor Parallelism
Implements tensor parallelism for selective state-space models on multi-GPU systems to scale SSM inference beyond single GPU limits.
2/25/2026
Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions
Decomposes epistemic uncertainty in Bayesian deep learning into per-class contributions for safety-critical classification asymmetric costs.
2/25/2026
Aletheia tackles FirstProof autonomously
Aletheia, a mathematics research AI agent powered by Gemini 3 Deep Think, autonomously solves 6 of 10 FirstProof challenge problems.
2/25/2026
Squint: Fast Visual Reinforcement Learning for Sim-to-Real Robotics
Combines off-policy and on-policy reinforcement learning for fast visual sim-to-real robotics training with minimal sample waste.
2/25/2026
Towards Attributions of Input Variables in a Coalition
Proposes coalition-based partitioning approach for Shapley value attribution methods in explainable AI to resolve attribution conflicts.
2/25/2026
Rethinking Disentanglement under Dependent Factors of Variation
Theoretical framework for disentangled representation learning when factors of variation are dependent rather than independent.
2/25/2026
Minibatch Optimal Transport and Perplexity Bound Estimation in Discrete Flow Matching
Proposes dynamic optimal transport minimization for discrete flow matching on categorical data with Kantorovich formulation.
2/25/2026
Characterizing LLM Inference Energy-Performance Tradeoffs across Workloads and GPU Scaling
Measurement-driven analysis of LLM inference energy and performance tradeoffs across workloads using GPU DVFS on 1B-32B parameter models.
2/25/2026
Universality of Benign Overfitting in Binary Linear Classification
Theoretical investigation of benign overfitting phenomenon in binary linear classification across over-parametrized neural networks.
2/25/2026
A Statistical Learning Perspective on Semi-dual Adversarial Neural Optimal Transport Solvers
Theoretical analysis of neural network-based optimal transport solvers using semi-dual adversarial formulations for generative modeling.
2/25/2026
Improving the Convergence of Private Shuffled Gradient Methods with Public Data
Theoretical analysis of differentially private shuffled gradient methods for convex ERM, addressing privacy-accuracy tradeoffs compared to standard DP-SGD.
2/25/2026
Using the Path of Least Resistance to Explain Deep Networks
Analyzes flaws in Integrated Gradients attribution method and proposes alternative path-based approach using model-induced geometry for better feature importance explanations.
2/25/2026
Distributional Vision-Language Alignment by Cauchy-Schwarz Divergence
Proposes Cauchy-Schwarz divergence for vision-language alignment to address distributional differences and alignment-uniformity conflicts in multimodal models.
2/25/2026
Semantic Parallelism: Redefining Efficient MoE Inference via Model-Data Co-Scheduling
Semantic Parallelism optimizes MoE LLM inference by co-scheduling model device placement and request routing, reducing communication overhead in multi-device serving.
2/25/2026
A Survey on Federated Fine-tuning of Large Language Models
Comprehensive survey of Federated Learning combined with LLM fine-tuning (FedLLM), covering privacy-preserving collaborative model adaptation methods.
2/25/2026
Towards Trustworthy GUI Agents: A Survey
Survey of trustworthy GUI agents built on LLMs, identifying execution gap challenges in real-world digital environment automation with irreversible actions.
2/25/2026
RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models
RefLoRA improves LoRA fine-tuning of large models by identifying optimal low-rank factorizations to address convergence and performance degradation issues.
2/25/2026
Performance Asymmetry in Model-Based Reinforcement Learning
Analysis of performance asymmetry in Model-Based RL agents on Atari100k, showing dramatic variance across task types despite high average performance.
2/25/2026
Wasserstein Barycenter Soft Actor-Critic
Wasserstein Barycenter Soft Actor-Critic algorithm improves sample efficiency in off-policy reinforcement learning via directed exploration.
2/25/2026
Foundation Models for Causal Inference via Prior-Data Fitted Networks
CausalFM framework trains Prior-Data Fitted Networks as foundation models for causal inference via in-context learning on tabular data.
2/25/2026
FREQuency ATTribution: benchmarking frequency-based occlusion for time series data
Frequency-domain occlusion method for interpreting time series neural networks, benchmarking frequency-based attribution approaches.
2/25/2026
Complexity-aware fine-tuning
Efficient fine-tuning method for LLMs using entropy-based complexity detection to apply chain-of-thought reasoning selectively on difficult examples.
2/25/2026
Transfer Learning in Infinite Width Feature Learning Networks
Theoretical analysis of transfer learning in infinitely wide neural networks under gradient flow, quantifying pretraining benefits.
2/25/2026
One-Step Flow Q-Learning: Addressing the Diffusion Policy Bottleneck in Offline Reinforcement Learning
One-Step Flow Q-Learning accelerates Diffusion Q-Learning for offline reinforcement learning by enabling single-step denoising without auxiliary modules.
2/25/2026
Uncertainty Propagation Networks for Neural Ordinary Differential Equations
Uncertainty Propagation Networks extend neural ODEs to model both state trajectories and uncertainty quantification in continuous-time systems.
2/25/2026
Monte Carlo Tree Diffusion with Multiple Experts for Protein Design
MCTD-ME combines masked diffusion models with Monte Carlo Tree Search for protein design, addressing long-range dependencies and search space challenges.
2/25/2026
From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting
Probabilistic Scenarios paradigm for time series forecasting generates finite scenario sets instead of samples to address computational and coverage limitations.
2/25/2026
RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility
RHYTHM framework uses LLMs as spatio-temporal predictors with hierarchical temporal tokenization for human mobility prediction.
2/25/2026
Polychromic Objectives for Reinforcement Learning
Polychromic objectives framework for reinforcement learning fine-tuning preserves policy diversity during RLFT to prevent mode collapse.
2/25/2026
Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models
Recursive Self-Aggregation (RSA) test-time scaling method combines parallel and sequential inference to improve LLM reasoning capabilities.
2/25/2026
Cautious Weight Decay
Cautious Weight Decay (CWD) optimizer modification applies weight decay only to parameters aligned with optimizer updates.
2/25/2026
TeamFormer: Shallow Parallel Transformers with Progressive Approximation
TeamFormer proposes shallow parallel Transformer architecture with progressive approximation for efficient training and inference.
2/25/2026
Latent-Augmented Discrete Diffusion Models
Latent-Augmented Discrete Diffusion (LADD) improves discrete diffusion models for fast language generation by modeling cross-token dependencies.
2/25/2026
Towards Scalable Oversight via Partitioned Human Supervision
Framework for scalable AI oversight by partitioning complex multi-domain evaluation tasks among domain-specific human experts.
2/25/2026
ContextPilot: Fast Long-Context Inference via Context Reuse
ContextPilot accelerates long-context LLM inference by enabling context reuse via KV-cache optimization for RAG and agent memory layers.
2/25/2026
SAGE-5GC: Security-Aware Guidelines for Evaluating Anomaly Detection in the 5G Core Network
Framework for evaluating anomaly detection in 5G networks accounting for non-IID data and adaptive attackers.
2/25/2026
Towards Robust Scaling Laws for Optimizers
Research on scaling laws for LLM pretraining with different optimizers beyond AdamW, examining new optimizers like Muon, Shampoo, and SOAP.
2/25/2026
Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs
Research on optimization algorithms for LLM reinforcement learning, comparing SGD vs Adam optimizers and their effectiveness in RL training phases.
2/25/2026
AceGRPO: Adaptive Curriculum Enhanced Group Relative Policy Optimization for Autonomous Machine Learning Engineering
AceGRPO method for autonomous ML engineering agents using adaptive curriculum and group relative policy optimization to overcome behavioral stagnation.
2/25/2026
VESPO: Variational Sequence-Level Soft Policy Optimization for Stable Off-Policy LLM Training
VESPO algorithm for stable off-policy LLM training via importance sampling with variance reduction to prevent policy divergence and collapse.
2/25/2026
KBVQ-MoE: KLT-guided SVD with Bias-Corrected Vector Quantization for MoE Large Language Models
Vector quantization compression for MoE LLMs using KLT-guided SVD and bias-corrected quantization for ultra-low-bit model deployment.
2/25/2026
MUSE: Multi-Tenant Model Serving With Seamless Model Updates
Multi-tenant ML serving system handling seamless model updates while maintaining decision thresholds across clients with distribution shifts.
2/25/2026
Pawsterior: Variational Flow Matching for Structured Simulation-Based Inference
Pawsterior framework for simulation-based inference using variational flow matching with structured domain constraints for bounded parameters.
2/25/2026
Silent Inconsistency in Data-Parallel Full Fine-Tuning: Diagnosing Worker-Level Optimization Misalignment
Analysis of worker-level optimization misalignment in data-parallel LLM fine-tuning despite parameter synchronization, termed silent inconsistency.
2/25/2026
Divine Benevolence is an $x^2$: GLUs scale asymptotically faster than MLPs
Numerical analysis showing GLU variants scale asymptotically faster than MLPs, explaining architectural dominance in frontier LLMs.
2/25/2026
\"UberWeb: Insights from Multilingual Curation for a 20-Trillion-Token Dataset
Study of multilingual data curation across 13 languages identifying interference patterns and optimal training strategies for 20-trillion-token dataset.