5/6/2026
Distribution-Free Pretraining of Classification Losses via Evolutionary Dynamics
Evolutionary Dynamic Loss framework learns transferable classification losses using synthetic data and ranking-consistency objectives without real samples.
5/6/2026
Rethinking the Rank Threshold for LoRA Fine-Tuning
Re-examines LoRA rank thresholds for LLM fine-tuning, questioning prior NTK analysis and its practical implications for cross-entropy loss.
5/6/2026
Nora: Normalized Orthogonal Row Alignment for Scalable Matrix Optimizer
Introduces Nora, a matrix-based optimizer for LLMs combining Muon-like preconditioning with scale-invariance and computational efficiency.
5/6/2026
Memory-Efficient Continual Learning with CLIP Models
Proposes memory-efficient continual learning for CLIP models using distributed contrastive loss to reduce forgetting with small memory buffers.
5/6/2026
On Adaptivity in Zeroth-Order Optimization
Analyzes zeroth-order optimization for LLM fine-tuning, showing adaptive ZO methods offer no convergence advantage over ZO-SGD despite memory overhead.
5/6/2026
Pretrained Model Representations as Acquisition Signals for Active Learning of MLIPs
Uses pretrained MLIP representations as acquisition signals for active learning in machine learning interatomic potentials for chemistry.
5/6/2026
An Information-Geometric Framework for Stability Analysis of Large Language Models under Entropic Stress
Thermodynamic-inspired framework for analyzing LLM stability under uncertainty using entropy-based metrics beyond aggregate accuracy.
5/6/2026
The Kerimov-Alekberli Model: An Information-Geometric Framework for Real-Time System Stability
Kerimov-Alekberli model applying information geometry and non-equilibrium thermodynamics to AI safety and autonomous system alignment.
5/6/2026
Memorization In Stable Diffusion Is Unexpectedly Driven by CLIP Embeddings
Analysis of how CLIP embeddings unexpectedly drive memorization in Stable Diffusion text-to-image models, with categorization of token types.
5/6/2026
CreativityBench: Evaluating Agent Creative Reasoning via Affordance-Based Tool Repurposing
CreativityBench benchmark evaluating LLM creative problem-solving through affordance-based tool repurposing tasks.
5/6/2026
Reasoning-Guided Grounding: Elevating Video Anomaly Detection through Multimodal Large Language Models
VANGUARD system using multimodal LLMs for video anomaly detection with interpretable reasoning and spatial localization of anomalous events.
5/6/2026
When Should a Language Model Trust Itself? Same-Model Self-Verification as a Conditional Confidence Signal
Evaluation of same-model self-verification as confidence signal for LLMs compared against likelihood-based baselines on reasoning tasks.
5/6/2026
Where to Bind Matters: Hebbian Fast Weights in Vision Transformers for Few-Shot Character Recognition
Study of Hebbian Fast-Weight modules in vision transformers enabling rapid within-episode adaptation for few-shot character recognition.
5/6/2026
EvoJail: Evolutionary Diverse Jailbreak Prompt Generation for Large Language Models
EvoJail method for generating diverse automated jailbreak prompts for LLMs using evolutionary algorithms, designed for adaptability to updated safety-finetuned models.
5/6/2026
Analysis and Explainability of LLMs Via Evolutionary Methods
Framework applying evolutionary methods to analyze and explain LLMs by relating model weights to genotypes and outputs to phenotypes.
5/6/2026
EFGPP: Exploratory framework for genotype-phenotype prediction
EFGPP framework for genotype-to-phenotype prediction combining multiple data sources for complex trait prediction, tested on migraine prediction.
5/6/2026
Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection
SALO method for detecting jailbreaks by tracing dynamic refusal trajectories in LLM representations using causal analysis, robust against adversarial attacks.
5/6/2026
MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
MedStruct-S benchmark for semi-structured information extraction from OCR clinical reports with tasks for key discovery, QA, and key-value pair extraction.
5/6/2026
Geometric Deviation as an Unsupervised Pre-Generation Reliability Signal: Probing LLM Representations for Answerability
Method to detect whether LLM can answer queries before generation by measuring hidden state geometry deviation from answerable reference sets, tested on Llama, Qwen, and Mistral models.
5/6/2026
Sparse Memory Finetuning as a Low-Forgetting Alternative to LoRA and Full Finetuning
Sparse Memory Finetuning (SMF) method for adapting LLMs while reducing catastrophic forgetting by selectively updating memory rows during training. Compared against LoRA and full finetuning.
5/6/2026
RLDX-1 Technical Report
RLDX-1 robotic policy model extending Vision-Language-Action models with motion awareness, memory, and physical sensing for complex real-world tasks.
5/6/2026
Imbalanced Classification under Capacity Constraints
Imbalanced classification framework addressing underrepresented classes under operational capacity constraints for rare detection tasks.
5/6/2026
Contrastive Regularization for Accent-Robust ASR
Contrastive regularization method for accent-robust ASR using supervised contrastive learning as auxiliary CTC fine-tuning objective.
5/6/2026
Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems
Treats coordination as architectural layer for multi-agent LLM systems, addressing 41-87% production failure rates from coordination defects not capability gaps.
5/6/2026
Provable Accuracy Collapse in Embedding-Based Representations under Dimensionality Mismatch
Proves sharp dimension-accuracy tradeoffs in embedding representations showing accuracy collapse when embedding dimension mismatches ground truth.
5/6/2026
APEX: Large-scale Multi-task Aesthetic-Informed Popularity Prediction for AI-Generated Music
APEX predicts popularity for AI-generated music using multi-task learning with aesthetic quality metrics on new AI music landscape.
5/6/2026
GRPO-TTA: Test-Time Visual Tuning for Vision-Language Models via GRPO-Driven Reinforcement Learning
GRPO-TTA applies Group Relative Policy Optimization to test-time adaptation of vision-language models via prompt refinement and RL.
5/6/2026
Exposing LLM Safety Gaps Through Mathematical Encoding:New Attacks and Systematic Analysis
Demonstrates LLM safety bypasses using mathematical encoding (set theory, logic, quantum mechanics) achieving 46-56% attack success across eight models.
5/6/2026
FinSTaR: Towards Financial Reasoning with Time Series Reasoning Models
Proposes capability taxonomy for time series reasoning models applied to financial domain with 2x2 classification framework for entity and temporal analysis.
5/6/2026
MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents
Formalizes memory poisoning attacks on retrieval-augmented LLM agents using game theory, proposing MEMSAD detection method with unified evaluation framework.
5/6/2026
Revisiting Graph-Tokenizing Large Language Models: A Systematic Evaluation of Graph Token Understanding
Systematic evaluation of Graph-Tokenizing LLMs questioning effectiveness of treating graph data as prefix tokens for LLMs.
5/6/2026
SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation
SURE-RAG framework for selective retrieval-augmented generation that verifies evidence sufficiency before answering.
5/6/2026
Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies
Workspace-Bench 1.0 benchmark evaluates AI agents on file-dependent workspace tasks with real-world dependencies at scale.
5/6/2026
Where Paths Split: Localized, Calibrated Control of Moral Reasoning in Large Language Models
Convergent-Divergent Routing steers LLM inference toward desired ethical frameworks by gating transformer pathways at branch points.
5/6/2026
Agentic-imodels: Evolving agentic interpretability tools via autoresearch
Agentic-imodels introduces an autoresearch loop that evolves data-science interpretability tools designed for agent understanding rather than human interpretation.
5/6/2026
The Manokhin Probability Matrix: A Diagnostic Framework for Classifier Probability Quality
Manokhin Probability Matrix diagnostic framework separating classifier reliability and discriminatory power using 2x2 grid methodology.
5/6/2026
Multimodal Learning on Low-Quality Data with Conformal Predictive Self-Calibration
Conformal predictive self-calibration framework for multimodal learning on low-quality data with modality imbalance and noise.
5/6/2026
Steer Like the LLM: Activation Steering that Mimics Prompting
Framework that formulates prompt steering as activation steering for LLMs, investigating whether distillation can close performance gaps.
5/6/2026
Ecologically-Constrained Task Arithmetic for Multi-Taxa Bioacoustic Classifiers Without Shared Data
Task vector arithmetic for combining independently fine-tuned bioacoustic encoders without sharing data across taxa and institutions.
5/6/2026
Conditional Diffusion Sampling
Conditional diffusion sampling approach for sampling from unnormalized multimodal distributions with limited density evaluations.
5/6/2026
Safety and accuracy follow different scaling laws in clinical large language models
Framework measuring how safety and accuracy scale differently in clinical LLMs, showing size increases don't guarantee safer behavior.
5/6/2026
A Unified Framework for Tabular Generative Modeling: Loss Functions, Benchmarks, and Improved Multi-objective Bayesian Optimization Approaches
Unified framework for tabular generative modeling with loss functions, benchmarks, and multi-objective Bayesian optimization approaches.
5/6/2026
Deep deterministic policy gradient with symmetric data augmentation for lateral attitude tracking control of a fixed-wing aircraft
Applies symmetric data augmentation with deep deterministic policy gradient for aircraft control. Uses reinforcement learning with sample efficiency.
5/6/2026
Power-Softmax: Towards Secure LLM Inference over Encrypted Data
Method for privacy-preserving LLM inference over encrypted data by approximating Softmax and layer normalization as polynomials for homomorphic encryption.
5/6/2026
Understanding and Mitigating Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks
Studies bias inheritance when LLMs generate synthetic training data, showing how biases propagate and amplify in downstream tasks.
5/6/2026
Sample-Efficient Optimization over Generative Priors via Coarse Learnability
Research on zeroth-order optimization with generative priors using coarse learnability for sample-efficient model-based optimization.
5/6/2026
Super-fast Rates of Convergence for Neural Network Classifiers under the Hard Margin Condition
Convergence rate analysis for deep neural network classifiers under hard margin condition.
5/6/2026
Joint Relational Database Generation via Graph-Conditional Diffusion Models
Graph-conditional diffusion models for joint generation of multi-table relational databases.
5/6/2026
The Polar Express: Optimal Matrix Sign Methods and Their Application to the Muon Algorithm
Optimal matrix sign computation methods for neural network training via the Muon optimizer.
5/6/2026
Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding
Selective Jacobi decoding accelerates discrete autoregressive normalizing flows for generative modeling.