15d ago
Enhancing Adversarial Transferability through Block Stretch and Shrink
Block Stretch and Shrink method improving adversarial attack transferability through implicit ensemble perspective on model frontend responses.
15d ago
CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning
CUDA-L2: combines LLMs and RL to automatically optimize HGEMM CUDA kernels, outperforming cuBLAS and torch.matmul on matrix multiplication.
15d ago
Stable On-Policy Distillation through Adaptive Target Reformulation
On-policy knowledge distillation method stabilizing training by reformulating targets to address distribution mismatch between student and teacher LLMs.
15d ago
BalDRO: A Distributionally Robust Optimization based Framework for Large Language Model Unlearning
BalDRO: distributionally robust optimization framework for LLM unlearning addressing sample-wise imbalance in forgetting difficulty.
15d ago
TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Models
TimeSAE: sparse decoding method providing causal explanations for black-box time series models with out-of-distribution generalization.
15d ago
HyperNet-Adaptation for Diffusion-Based Test Case Generation
HyperNet-Adaptation framework using diffusion models for generating realistic test cases for deep learning systems beyond simple adversarial perturbations.
15d ago
SFO: Learning PDE Operators via Spectral Filtering
Spectral Filtering Operator (SFO): neural operator for solving PDEs using Universal Spectral Basis to capture long-range nonlocal interactions.
15d ago
Rethinking Zero-Shot Time Series Classification: From Task-specific Classifiers to In-Context Inference
TIC-FM: in-context learning framework for zero-shot time series classification using foundation models without task-specific training.
15d ago
CoGenCast: A Coupled Autoregressive-Flow Generative Framework for Time Series Forecasting
CoGenCast: generative framework combining autoregressive LLMs and flow models for time series forecasting with semantic and stochastic modeling.
15d ago
Disentangling Intrinsic Importance from Emergent Structure in Multi-Expert Orchestration
INFORM: interpretability framework for analyzing multi-LLM orchestration policies, decoupling expert interaction structure from intrinsic importance in collaborative reasoning systems.
15d ago
Separation-Utility Pareto Frontier: An Information-Theoretic Characterization
Information-theoretic analysis of the trade-off between utility and fairness (separation) in predictive models, proving Pareto frontier concavity.
15d ago
SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data
Benchmark and toolkit for evaluating sparse autoencoders using large-scale synthetic data with realistic features.
15d ago
BRIDGE: Bridging Reasoning In Distillation Gap Elimination via Structure-Aware Masking
Knowledge distillation method for compressing chain-of-thought reasoning from large LLMs to smaller models using structure-aware masking.
15d ago
Turbo Connection: Reasoning as Information Flow from Higher to Lower Layers
Transformer architecture extension enabling multi-step reasoning by allowing information flow across layers, addressing fixed-depth computation limits.
15d ago
Dynamics of Learning under User Choice: Overspecialization and Peer-Model Probing
Analysis of machine learning platform dynamics when users select between competing learners, studying convergence and model specialization.
15d ago
Local Message-Passing for Discrete Graph Generation
GenGNN message-passing backbone for discrete graph generation using coupled node-edge refinement without graph transformers.
15d ago
ECoLAD: Selecting Anomaly Detectors for Automotive Deployment via Compute-Reduction Evaluation
Deployment-focused evaluation protocol for selecting time-series anomaly detectors suitable for automotive applications.
15d ago
Simple Recipe Works: Vision-Language-Action Models are Natural Continual Learners with Reinforcement Learning
Demonstrates vision-language-action models naturally handle continual learning with sequential fine-tuning and RL.
15d ago
The Cost of Reasoning: Chain-of-Thought Induces Overconfidence in Vision-Language Models
Shows chain-of-thought reasoning in vision-language models degrades uncertainty quantification reliability.
15d ago
Critical Damping as a Momentum Schedule: Multi-Seed Validation, a Hybrid Recipe, and an Exhaustive Negative Result on Surgical Layer Selection
Critical damping-based momentum schedule for SGD achieving faster convergence without hyperparameter tuning.
15d ago
ACES: Who Tests the Tests? Leave-One-Out AUC Consistency for Code Generation
Framework for selecting LLM-generated code using test voting consistency without determining individual test correctness.
15d ago
Retrieval with Multiple Query Vectors through Anomalous Pattern Detection
Retrieval method using multiple query vectors simultaneously via anomalous pattern detection for complex reasoning tasks.
15d ago
Attribution-Guided Continual Learning for Large Language Models
Attribution-guided approach for continual learning in LLMs that identifies and preserves parameters storing important knowledge.
15d ago
COSMOS: Model-Agnostic Personalized Federated Learning with Clustered Server Models and Pseudo-Label-Only Communication
Federated learning framework handling heterogeneous client architectures and data distributions via server-side personalization.
15d ago
CTFusion: A CTF-based Benchmark for LLM Agent Evaluation
Benchmark using Capture The Flag challenges to evaluate LLM agents on cybersecurity tasks with contamination safeguards.
15d ago
Selective Safety Steering via Value-Filtered Decoding
Decoding-time steering method for LLMs that selectively applies safety constraints without unnecessary interventions.
15d ago
Prune, Update and Trim: Robust Structured Pruning for Large Language Models
Post-training pruning method for reducing LLM inference costs by removing model parameters while maintaining performance.
15d ago
Student Capacity Moderates Knowledge Distillation Effectiveness: A Systematic Study Across ResNet Teacher-Student Pairs on CIFAR-10
Investigates how teacher-student capacity ratios affect knowledge distillation effectiveness in ResNet image classification on CIFAR-10.
15d ago
Revisiting Neural Processes via Fourier Transform and Volterra Series
Neural processes model using Fourier transforms and Volterra series for learning latent functions from irregular measurements with translation equivariance.
15d ago
From Performance to Representational Adequacy: A Representational Bootstrap Framework for Adaptive Biological Systems
Representational bootstrap framework for analyzing adaptive biological systems beyond aggregate performance metrics.
15d ago
IntraShuffler: A Privacy Preserving Framework for Heterogeneous DP Federated Learning
Privacy framework for federated learning with heterogeneous differential privacy budgets, addressing gradient pattern leakage in client updates.
15d ago
Online KL-Regularized Reinforcement Learning with Function Approximation under Misspecification
Analysis of KL-regularized reinforcement learning under function approximation with model misspecification, extending bounds to non-realizable settings.
15d ago
Bootstrap Theory of Representational Emergence: Explanatory Insufficiency as a Driver of Representation Learning and World Models
Theoretical framework (TBER) examining how new representational levels emerge in learning systems when existing representations become insufficient for explanation.
15d ago
The Injection Paradox: Brand-Level Suppression in Safety-Trained LLM Recommendations via RAG Context Injection
Research on prompt injection attacks against RAG-based LLM recommendation systems, revealing how safety training causes injected prompts to suppress target brands.
15d ago
Learning Dynamics Reveal a Hierarchy of Weight-Induced Layerwise Gram Metrics
Theoretical analysis of gradient descent dynamics in ReLU networks revealing layerwise Gram metrics hierarchy through activation field dynamics.
15d ago
Detecting Explanatory Insufficiency in Learned Representations: A Framework for Representational Vigilance
VER framework for detecting inadequate structure in learned representations beyond standard evaluation metrics, monitoring representational sufficiency.
15d ago
Gefen: Optimized Stochastic Optimizer
Gefen: memory-efficient optimizer reducing AdamW footprint by ~8x through moment state sharing and learned quantization for large-scale pretraining.
15d ago
FAST: A Framework for Aligned Sampling and Training in Parallel Reinforcement Learning for Autonomous Driving
FAST: parallel reinforcement learning framework addressing sampling efficiency bottlenecks in autonomous driving through straggler effect mitigation.
15d ago
Multi-Agent Routing as Set-Valued Prediction: A WildChat Benchmark and Cost-Aware Evaluation
Multi-agent routing benchmark from WildChat with 3,000 prompts and cost-aware evaluation protocol for set-valued prediction of tool/agent selection.
15d ago
Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers
Looped Transformers: recurrent architecture bridging latent and explicit reasoning in language models, achieving competitive performance with better efficiency.
15d ago
DemoPSD: Disagreement-Modulated Policy Self-Distillation
DemoPSD: self-distillation method for training LLMs to reason by modulating policy disagreement, reducing overfitting and improving generalization.
15d ago
Directional Curvature from Armijo Backtracking: A Low-Cost Sharpness Probe and a Calibration-Free Learning-Rate Safeguard for Adam
Method extracting directional curvature information from Armijo backtracking line search as low-cost sharpness probe for adaptive learning rate safeguards.
15d ago
RL Forgets! Towards Continual Policy Optimization
Study showing reinforcement learning for vision-language model adaptation exhibits catastrophic forgetting, challenging assumptions about RL continual learning.
15d ago
Minimum Block Width for Universal Approximation by Residual Neural Networks with Inner Width One
Theoretical analysis establishing minimum block width requirements for universal approximation in residual neural networks with inner width one.
15d ago
Layer-Parallel Inference Reduces Encrypted Nonlinear Depth in Transformers
Layer-parallel inference technique for encrypted Transformer computation using structured Newton methods to reduce FHE bottlenecks.
15d ago
x-Prediction Is All You Need:Training-Free Accelerated Generation via Endpoint Decodability
Training-free acceleration method for diffusion and flow matching models using x-prediction to reduce neural function evaluations during sampling.
15d ago
Efficient Long-Horizon Learning for Learned Optimization
Meta-learning approach to train neural network optimizers that outperform hand-designed optimizers like Adam, improving long-horizon training efficiency.
15d ago
Eigenbasis-Independent Learnable Spectral Positional Encodings for Directed Graphs via Hermitian Block Krylov Subspaces
Learnable spectral positional encodings for directed graphs using Hermitian block Krylov subspaces, addressing eigendecomposition efficiency.
15d ago
Robust Bayesian Decision Making under Adversarial Uncertainty
Bayesian decision-making framework for experimental design under adversarial uncertainty, focusing on reliable downstream decision-making.
15d ago
HiQA: A Hierarchical Contextual Augmentation RAG for Multi-Documents QA
HiQA proposes hierarchical contextual augmentation for retrieval-augmented generation in multi-document question-answering, improving accuracy and reducing hallucinations in LLM responses.