Framework converting multimodal LLM generative capabilities into zero-shot discriminative embedding models without extensive pre-training.
Model merging technique using task vector distillation to improve robustness of multi-task learning across diverse settings.
Theoretically motivated improvement to supervised fine-tuning for LLMs by rectifying reward structure to match RL generalization.
Offline multi-agent reinforcement learning using efficient flow-based policies for time-sensitive deployment.
Framework of strategies for improving LLM-based forecasting by integrating historical data and textual context with reduced computational cost.
Investigation of in-context learning in world models for embodied AI to adapt to novel environmental configurations.
Interpretable time series forecasting method using hierarchical prototypes to explain model decision-making.
Foundation inference model using in-context learning to predict marked temporal point process event sequences across different systems.
Process Reward Models that capture step-by-step reasoning dependencies in LLMs to improve reasoning alignment with final outcomes.
Benchmark dataset of 50 condensed matter theory problems for evaluating LLMs on advanced research-level physics problem-solving.
Permutation-invariant representation learning for privacy-preserving feature selection using generative intelligence.
Carré du champ flow matching: geometry-aware regularization technique for generative models improving quality-generalization tradeoff.
Hybrid tensor-EM method for learning mixtures of linear dynamical systems with improved performance on noisy time-series data.
Evaluation of zero-shot super-resolution capabilities in machine-learned operators for modeling continuous physical phenomena.
ToSFiT: Thompson sampling via LLM fine-tuning for Bayesian optimization in large discrete spaces without acquisition function maximization.
Differential privacy framework for decentralized learning using matrix factorization to enable collaborative training while preserving privacy.
FAPO method for LLM reasoning via reinforcement learning that filters flawed positive rollouts to improve policy optimization with verifiable rewards.
DiffuMamba: diffusion language model with Mamba backbone for efficient masked sequence modeling, achieving linear-time complexity vs Transformer quadratic overhead.
Analysis of stochastic gradient descent-based unlearning algorithms (D2D and R2D) with provable guarantees for removing training data impact.
Multi-agent reinforcement learning approach using attention for automated feature transformation in structured data processing.
Method for monitoring LLM API consistency over time by tracking log probability changes to detect undisclosed model updates.
Study on membership inference attacks for extracting training data from LLMs, demonstrating privacy risks through coordinated extraction and verification techniques.
Generalized Primal Averaging (GPA) optimizer for faster LLM training, extending Nesterov's method and unifying recent averaging-based approaches like DiLoCo.
Trust region masking technique for LLM reinforcement learning to address off-policy divergence in policy gradient training of large language models.
CSyMR benchmark for evaluating LLMs on compositional music information retrieval tasks requiring multi-step reasoning over symbolic music notation.
DUET method for LLM unlearning using distilled teacher models to remove undesirable knowledge efficiently while avoiding catastrophic forgetting.
Federated-inspired batch correction for single-cell RNA sequencing without centralizing high-dimensional datasets.
Position paper proposing agentic framework for time series forecasting with iterative refinement and adaptation.
KV-cache quantization to 2-bit precision enabling long video generation on resource-constrained hardware.
Machine unlearning method addressing superficial forgetting by targeting core feature representations rather than logits.
Online learning framework for training robust classifiers under adversarially chosen clean data and labels.
Pretrained variational bridge for efficient molecular dynamics trajectory generation across diverse molecular systems.
Automated detection pipeline for unverbalized biases in LLM chain-of-thought reasoning without predefined categories.
Performance characterization framework for small language models on edge devices using Roofline model analysis.
Benchmarking framework for time series foundation models addressing data quality, task alignment, and evaluation rigor.
Framework for measuring propensities (behavioral tendencies) in AI models beyond capability assessment using Item Response Theory.
Theoretical analysis of gradient descent convergence rates under large step sizes in separable logistic regression.
Discrete diffusion framework using sample-efficient conditional probability estimators for discrete state space generation.
Analysis showing test-time training with KV binding functions as learned linear attention rather than memorization.
Federated learning aggregation method (FedVG) using gradient guidance to address client drift and data heterogeneity.
Safety filtering framework for flow-based generative models with formal guarantees on constraint satisfaction.
Strategic risk aversion approach for training collaborative AI agents that generalize better with new partners.
Adversarial reinforcement learning dataset (AOT-SFT) to improve robustness of multimodal LLMs on visually complex scenes.
Maps failure regions in LLMs using MAP-Elites quality diversity search to characterize unsafe behaviors and vulnerabilities.
Energy-based theory for detecting concept drift in ECG signals, distinguishing physiologically plausible variation from true distribution shift.
Regularized online RLHF for Nash Equilibrium identification with generalized bilinear preferences modeling intransitive preferences.
ParamMem augments language agents with parametric reflective memory to improve reasoning through diverse self-reflection.
FinBloom presents a knowledge-grounding approach for LLMs to handle real-time financial queries with live data integration.
Apple proposes general active perception via reinforcement learning to handle uncertainty in partially observable robotic environments.
REA-RL uses online reinforcement learning with reflection to reduce overthinking and inference costs in large reasoning models.