4/6/2026
AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation
Low-precision training method for LLMs using adaptive Hadamard transforms based on outlier patterns in weights, activations, and gradients.
4/6/2026
Jump Start or False Start? A Theoretical and Empirical Evaluation of LLM-initialized Bandits
Research on using LLM-generated synthetic data to warm-start contextual bandits, examining alignment between LLM choices and actual user preferences.
4/6/2026
A Spectral Framework for Multi-Scale Nonlinear Dimensionality Reduction
Spectral framework for multi-scale nonlinear dimensionality reduction balancing global-local structure preservation and expressiveness-transparency.
4/6/2026
Fast NF4 Dequantization Kernels for Large Language Model Inference
Optimized NF4 dequantization kernels for fast LLM inference on NVIDIA GPUs, addressing FP16 conversion bottleneck.
4/6/2026
Communication-Efficient Distributed Learning with Differential Privacy
Communication-efficient distributed learning algorithm with differential privacy using local training and gradient clipping.
4/6/2026
VoxelCodeBench: Benchmarking 3D World Modeling Through Code Generation
VoxelCodeBench platform benchmarking code generation models for 3D spatial reasoning with execution in Unreal Engine.
4/6/2026
Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens
Analysis of function vectors in LLMs showing they steer behavior beyond logit lens interpretability across 4,032 cross-template transfer pairs.
4/6/2026
Complex-Valued GNNs for Distributed Basis-Invariant Control of Planar Systems
Complex-valued GNNs for distributed control of networked systems with basis-invariance for GPS and compass-denied environments.
4/6/2026
Analytic Drift Resister for Non-Exemplar Continual Graph Learning
Continual graph learning without storing exemplars, using class-level prototypes and analytic continual learning to address catastrophic forgetting.
4/6/2026
AXELRAM: Quantize Once, Never Dequantize
AXELRAM smart SRAM architecture computing attention scores from quantized KV cache without dequantization using orthogonal-transform quantization.
4/6/2026
Communication-free Sampling and 4D Hybrid Parallelism for Scalable Mini-batch GNN Training
Distributed GNN training with communication-free sampling and 4D hybrid parallelism for scaling mini-batch learning on large graphs.
4/6/2026
Generalization Limits of Reinforcement Learning Alignment
Study of generalization limits in RLHF alignment, proposing compound jailbreaks targeting LLM safety through redistribution of existing capabilities.
4/6/2026
Product-Stability: Provable Convergence for Gradient Descent on the Edge of Stability
Theoretical analysis of gradient descent training at edge of stability with product-stability property for convergence guarantees.
4/6/2026
Low-Rank Compression of Pretrained Models via Randomized Subspace Iteration
Low-rank compression of pretrained models using randomized subspace iteration for efficient SVD-based model reduction.
4/6/2026
Finding Belief Geometries with Sparse Autoencoders
Mechanistic interpretability research on whether LLMs encode belief geometries like transformers trained on hidden Markov models.
4/6/2026
Beyond Semantic Manipulation: Token-Space Attacks on Reward Models
Token-space attacks on reward models used in RLHF, introducing TOMPA framework for adversarial optimization beyond semantic manipulation.
4/6/2026
FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving
FluxMoE system decouples expert residency to improve inference serving throughput in Mixture-of-Experts LLMs.
4/6/2026
Generative Frontiers: Why Evaluation Matters for Diffusion Language Models
Technical analysis of evaluation methodology challenges for diffusion language models at scale.
4/6/2026
Towards Realistic Class-Incremental Learning with Free-Flow Increments
Free-flow class-incremental learning framework handling realistic variable-sized task streams without fixed task schedules.
4/6/2026
Random Is Hard to Beat: Active Selection in online DPO with Modern LLMs
Study comparing active preference learning vs random sampling for online DPO data selection in modern LLMs.
4/6/2026
Toward an Operational GNN-Based Multimesh Surrogate for Fast Flood Forecasting
GNN-based surrogate model for accelerating hydraulic flood forecasting simulations.
4/6/2026
Towards Near-Real-Time Telemetry-Aware Routing with Neural Routing Algorithms
Neural routing algorithms for network traffic optimization using live telemetry data with near-real-time latency.
4/6/2026
Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
Addresses reward hacking in RLHF by using advantage sign robustness to prevent policy degradation.
4/6/2026
FedSQ: Optimized Weight Averaging via Fixed Gating
Federated learning optimization via fixed gating for weight averaging under statistical heterogeneity.
4/6/2026
Co-Evolution of Policy and Internal Reward for Language Agents
Self-Guide framework for LLM agents using co-evolved internal rewards to address sparse reward problem in long-horizon tasks.
4/6/2026
Self-Distilled RLVR
On-policy self-distillation training paradigm for LLMs combining dense signals from larger teacher models.
4/6/2026
Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models
arXiv paper analyzes hallucination's role in RL post-training of multimodal LLMs, investigating whether models genuinely learn visual information under RL.
4/6/2026
PRISM: LLM-Guided Semantic Clustering for High-Precision Topics
arXiv paper proposes PRISM, combining LLM-guided labeling with semantic clustering for interpretable topic modeling with low computational cost.
4/6/2026
Reflective Context Learning: Studying the Optimization Primitives of Context Space
arXiv paper studies context space optimization for generally capable agents, examining credit assignment, forgetting, and overfitting in context learning.
4/6/2026
Enhancing Robustness of Federated Learning via Server Learning
arXiv paper exploring server learning to enhance federated learning robustness against malicious attacks with non-IID client data.
4/6/2026
MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites
arXiv paper on federated class-incremental learning for LEO satellites addressing non-IID data, memory, and communication constraints.
4/6/2026
Fighting AI with AI: AI-Agent Augmented DNS Blocking of LLM Services during Student Evaluations
arXiv paper proposes AI-agent-augmented DNS blocking to prevent LLM access during student evaluations, addressing academic integrity concerns.
4/6/2026
Backdoor Attacks on Decentralised Post-Training
arXiv paper analyzing backdoor attacks on decentralized LLM post-training via pipeline parallelism, examining vulnerabilities from malicious participants.
4/6/2026
Photonic convolutional neural network with pre-trained in-situ training
arXiv paper proposing fully photonic convolutional neural network with pre-trained in-situ training to overcome CMOS energy bottlenecks.
4/6/2026
PlayGen-MoG: Framework for Diverse Multi-Agent Play Generation via Mixture-of-Gaussians Trajectory Prediction
PlayGen-MoG: Mixture-of-Gaussians framework for diverse multi-agent trajectory generation in team sports with spatial coordination.
4/6/2026
Guideline2Graph: Profile-Aware Multimodal Parsing for Executable Clinical Decision Graphs
Guideline2Graph: LLM/VLM-based parser converting multimodal clinical guidelines into executable decision graphs with cross-page control flow.
4/6/2026
Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models
Examination of confirmation bias in LLMs through rule-discovery tasks, evaluating and mitigating systematic reasoning failures.
4/6/2026
Optimal Projection-Free Adaptive SGD for Matrix Optimization
Projection-free adaptive SGD for matrix optimization achieving dimension-independent convergence without costly quadratic projections.
4/6/2026
Reinforcement Learning from Human Feedback: A Statistical Perspective
Statistical survey of RLHF for LLM alignment, examining noisy subjective feedback for reward models and policy optimization.
4/6/2026
Neural posterior estimation for scalable and accurate inverse parameter inference in Li-ion batteries
Neural posterior estimation for Bayesian parameter inference in Li-ion batteries using probabilistic calibration.
4/6/2026
AQVolt26: High-Temperature r$^2$SCAN Halide Dataset for Universal ML Potentials and Solid-State Batteries
AQVolt26: High-temperature halide dataset with r²SCAN calculations for training universal ML potentials for solid-state batteries.
4/6/2026
Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions?
Feature Attribution Stability Suite: Evaluates post-hoc attribution method stability under realistic input perturbations with improved metrics.
4/6/2026
Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization
Synapse: LLM-guided two-phase retrieval and genetic algorithm optimization for job-person matching under information imbalance.
4/6/2026
Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation
Empirical study of confidence calibration and hallucination in vision-language models for medical VQA, spanning three model families.
4/6/2026
Contrastive Language-Colored Pointmap Pretraining for Unified 3D Scene Understanding
UniScene3D: Transformer-based 3D scene encoder jointly modeling appearance and geometry from multi-view colored pointmaps with CLIP alignment.
4/6/2026
WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models
WSVD: Weighted SVD technique for efficient low-precision execution of vision-language models with reduced computational burden.
4/6/2026
Understanding the Effects of Safety Unalignment on Large Language Models
Study examining safety alignment bypass methods in LLMs via jailbreak-tuning and weight orthogonalization, analyzing disabled safety guardrails.
4/6/2026
Learning interacting particle systems from unlabeled data
Trajectory-free self-test loss function for learning interacting particle system potentials from unlabeled discrete time-point data.
4/6/2026
Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport
Gromov-Wasserstein optimal transport methods for geometry-aware multi-view data embedding under heterogeneous geometries and nonlinear distortions.
4/6/2026
AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models
AutoVerifier: LLM-based agentic framework automating end-to-end verification of technical claims across scientific literature without domain expertise required.