4/2/2026
Neural Collapse Dynamics: Depth, Activation, Regularisation, and Feature Norm Threshold
Analysis of neural collapse dynamics identifying critical feature norm threshold for convergence.
4/2/2026
MAC-Attention: a Match-Amend-Complete Scheme for Fast and Accurate Attention Computation
MAC-Attention acceleration technique for LLM long-context decoding that preserves attention computation fidelity without compression.
4/2/2026
Informed Machine Learning with Knowledge Landmarks
Knowledge-Data ML framework integrating numeric data with knowledge for model construction.
4/2/2026
Autonomous Adaptive Solver Selection for Chemistry Integration via Reinforcement Learning
Reinforcement learning framework for autonomous solver selection in chemical kinetics integration.
4/2/2026
SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection
Agent system using RL to select optimal deep generative models for tabular data synthesis.
4/2/2026
Robust Multimodal Safety via Conditional Decoding
Conditional decoding strategy (CASA) for improving safety alignment in multimodal LLMs against cross-modal attacks.
4/2/2026
The Persistent Vulnerability of Aligned AI Systems
AI safety research on vulnerabilities in autonomous agents with filesystem/email access via circuit analysis.
4/2/2026
Is One Token All It Takes? Graph Pooling Tokens for LLM-based GraphQA
Method for encoding graph structure into LLMs via graph pooling tokens for Graph Question Answering tasks.
4/2/2026
Gradient-Based Data Valuation Improves Curriculum Learning for Game-Theoretic Motion Planning
Gradient-based data valuation for curriculum learning in game-theoretic motion planning using TracIn scoring.
4/2/2026
Deep Networks Favor Simple Data
Study showing deep networks assign higher density to simpler out-of-distribution data than in-distribution test data.
4/2/2026
A Cross-graph Tuning-free GNN Prompting Framework
Tuning-free GNN prompting framework for cross-graph adaptation without task-specific parameter updates.
4/2/2026
G-Drift MIA: Membership Inference via Gradient-Induced Feature Drift in LLMs
Membership inference attack on LLMs via gradient-induced feature drift to detect training data exposure.
4/2/2026
Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions
Research on scheduling LLM inference using uncertainty-aware output length predictions instead of point estimates.
4/2/2026
Towards Initialization-dependent and Non-vacuous Generalization Bounds for Overparameterized Shallow Neural Networks
arXiv: Generalization bounds for overparameterized shallow neural networks using initialization-dependent distance norms.
4/2/2026
A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization
arXiv: Decoupled basis-vector-driven generative framework for dynamic multi-objective optimization addressing irregular mutations and cold-start.
4/2/2026
MOON3.0: Reasoning-aware Multimodal Representation Learning for E-commerce Product Understanding
arXiv: MOON3.0 multimodal representation learning framework for fine-grained e-commerce product understanding using reasoning-aware embeddings.
4/2/2026
Lipschitz Dueling Bandits over Continuous Action Spaces
arXiv: First algorithm for Lipschitz dueling bandits over continuous action spaces using adaptive reference arms.
4/2/2026
MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference
arXiv: Multi-format quantization-aware training enables single model robustness across multiple numeric precisions for elastic inference.
4/2/2026
Learning Shared Representations for Multi-Task Linear Bandits
arXiv: Multi-task representation learning in linear bandits with shared latent representations for knowledge transfer.
4/2/2026
Learning from Many and Adapting to the Unknown in Open-set Test Streams
arXiv: Test-time adaptation for LLMs under continual distribution shift and open-set tasks, preserving source knowledge.
4/2/2026
HabitatAgent: An End-to-End Multi-Agent System for Housing Consultation
arXiv: HabitatAgent multi-agent LLM system for housing consultation with transparent reasoning and factuality guarantees.
4/2/2026
Representation choice shapes the interpretation of protein conformational dynamics
arXiv: Study on how representation choice affects interpretation of protein conformational dynamics from molecular dynamics simulations.
4/2/2026
Predicting Dynamics of Ultra-Large Complex Systems by Inferring Governing Equations
arXiv: Sparse Identification Graph Neural Network for discovering interpretable governing equations in ultra-large complex systems.
4/2/2026
A Survey of On-Policy Distillation for Large Language Models
arXiv survey: On-policy distillation transfers reasoning from frontier LLMs to smaller models, addressing exposure bias in knowledge distillation.
4/2/2026
Chameleons do not Forget: Prompt-Based Online Continual Learning for Next Activity Prediction
arXiv: Prompt-based online continual learning for next activity prediction in dynamic processes using catastrophic forgetting mitigation.
4/2/2026
Embedded Variational Neural Stochastic Differential Equations for Learning Heterogeneous Dynamics
arXiv: Variational Neural Stochastic Differential Equations model complex socioeconomic time-series data with heterogeneous dynamics.
4/2/2026
Full-Gradient Successor Feature Representations
arXiv: Full-gradient successor feature representations improve convergence guarantees for transfer learning in RL with non-linear function approximation.
4/2/2026
Performance of Neural and Polynomial Operator Surrogates
arXiv: Empirical comparison of neural operator surrogates including Fourier neural operators vs polynomial methods for parametric PDEs.
4/2/2026
Learning to Hint for Reinforcement Learning
arXiv: Group Relative Policy Optimization for RL addresses advantage collapse in reinforcement learning with verifiable rewards using hints.
4/2/2026
Exploring Silent Data Corruption as a Reliability Challenge in LLM Training
Analysis of silent data corruption during LLM training on hardware, studying gradient corruption impacts and detection mechanisms.
4/2/2026
Spectral Compact Training: Pre-Training Large Language Models via Permanent Truncated SVD and Stiefel QR Retraction
Spectral Compact Training method reduces LLM training memory footprint by replacing dense weight matrices with truncated SVD factors.
4/2/2026
ActivityNarrated: An Open-Ended Narrative Paradigm for Wearable Human Activity Understanding
Open-ended narrative framework for wearable human activity recognition using compositional, unscripted activities instead of closed-set classification.
4/2/2026
Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning
ThoughtSteer backdoor attack exploiting continuous reasoning in language models that operate silently in hidden states without token output.
4/2/2026
Using predefined vector systems to speed up neural network multimillion class classification
Method to reduce neural network multi-class classification complexity from O(n) to O(1) by leveraging known latent space geometry properties.
4/2/2026
Scalable Pretraining of Large Mixture of Experts Language Models on Aurora Super Computer
Optimus training library for pretraining mixture-of-experts LLMs at exascale on Aurora supercomputer, demonstrating 1000s GPU tile scaling.
4/2/2026
Cost-Penalized Fitness in FMA-Orchestrated Mixture of Experts: Experimental Evidence for Molecular Memory in Domain Adaptation
Experimental evaluation of Free-Market Algorithm orchestrated Mixture-of-Experts with cost-penalized fitness for domain adaptation.
4/2/2026
Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation
Optimal decomposition technique for low-rank approximation of LLM weights enabling efficient fine-tuning and inference.
4/2/2026
Learning to Learn-at-Test-Time: Language Agents with Learnable Adaptation Policies
Method for language agents to optimize test-time adaptation policies through iterative refinement during inference.
4/2/2026
Policy Improvement Reinforcement Learning
Reinforcement learning approach with verification for iteratively improving LLM policies based on actual performance gains.
4/2/2026
Fatigue-Aware Learning to Defer via Constrained Optimisation
Framework for human-AI cooperation that models fatigue-induced performance degradation in learning-to-defer systems.
4/2/2026
WARP: Guaranteed Inner-Layer Repair of NLP Transformers
Method for verifiable repair of transformer vulnerabilities to adversarial perturbations with inner-layer guarantees.
4/2/2026
EmbedPart: Embedding-Driven Graph Partitioning for Scalable Graph Neural Network Training
Graph partitioning technique using embeddings to enable scalable distributed training of graph neural networks.
4/2/2026
Transfer learning for nonparametric Bayesian networks
Transfer learning methodologies for Bayesian network structure learning with scarce data.
4/2/2026
Model-Based Learning of Near-Optimal Finite-Window Policies in POMDPs
Model-based learning approach for finite-window policies in partially observable Markov decision processes.
4/2/2026
Fast and Accurate Probing of In-Training LLMs' Downstream Performances
Method for efficiently evaluating LLM downstream performance during training without expensive full inference.
4/2/2026
Reconsidering Dependency Networks from an Information Geometry Perspective
Theoretical analysis of dependency networks using information geometry perspective for modeling complex systems.
4/2/2026
Reasoning Shift: How Context Silently Shortens LLM Reasoning
Analysis showing how irrelevant context degrades LLM reasoning performance despite test-time scaling capabilities.
4/2/2026
Bridging the Simulation-to-Experiment Gap with Generative Models using Adversarial Distribution Alignment
Generative model approach using adversarial distribution alignment to bridge simulation-to-experiment gap in scientific domains.
4/2/2026
Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning
ORCA framework calibrating LLM sampling through conformal prediction to improve test-time reasoning efficiency and generalization.
4/2/2026
Screening Is Enough
Multiscreen mechanism for language models enabling absolute query-key relevance assessment beyond relative attention redistribution.