DPPrefSyn algorithm generates differentially private synthetic preference data for privacy-preserving LLM alignment training without exposing sensitive human feedback.
SCOUT framework reframes prompt-injection defense as adaptive detector allocation, routing requests to appropriate detectors or LLM judges based on pre-hoc reasoning.
Theoretical analysis of Bayesian inference in deep MLPs studying how neural networks make predictions with large model and dataset sizes.
Generative learning framework for optimizing quantum data embeddings with learnable gate sequences and fidelity-based surrogate objectives.
Auto-research framework using tool-calling LLMs to propose and test machine-learned interatomic potential training strategies with HPC integration.
Unifying operator-side perspective on anchored fixed-point methods with applications to obtaining last-iterate convergence guarantees.
Benchmark evaluating LLM agents on financial spreadsheet tasks including synthesis, manipulation, and comprehension in professional finance domain.
Theoretical analysis of stochastic linear bandits under batching and extreme quantization constraints with single-bit feedback per batch.
Formalizes cognitive fatigue in autoregressive transformers with online diagnostics for detecting degradation in long-horizon generation.
Few-shot learning approach leveraging benchmark evaluations as side information with weak monotonicity assumptions for rapid task adaptation.
Collective communication framework enabling efficient LLM training on mixed-vendor heterogeneous hardware clusters with diverse network characteristics.
Evaluation framework replacing monolithic benchmarks with persona-based synthetic judges to assess pluralistic alignment in generative AI systems.
Multi-agent differential transformer system for autonomous resource management in heterogeneous satellite clusters conducting Earth observation missions.
Framework studying interaction between ML model quality, auction mechanisms, and automated bidding behavior in online advertising platforms.
Training-free sparse attention mechanism for long video diffusion models reducing quadratic computational cost while maintaining quality.
Domain incremental learning approach combining masked autoencoder with LoRA adapters for non-stationary video streams, exploiting catastrophic forgetting.
Research on self-training language models with synthetic data, identifying when models can improve from self-generated text through latent capability resurfacing.
Method enabling vision-language models to localize objects in images from few examples without training, applicable to image editing and visual search.
Theoretical analysis of approximation rates for deep ReLU networks learning smooth functions with focus on overcoming curse of dimensionality in high-dimensional spaces.
Training-free inference acceleration for interactive video world models enabling real-time simulation and embodied AI.
Probabilistic sequence layer design using Bayesian filtering for efficient recurrent neural networks with uncertainty tracking.
Agentic framework using experience-aware reasoning for object detection across diverse scenes and image degradations.
Study showing sparse autoencoders outperform baselines for steering LLM outputs on AxBench benchmark.
Zero-shot cross-lingual confidence estimation method for multilingual LLMs without retraining across languages.
Method to generate multi-hop training data from unannotated text using graph constraints for improving LLM compositional reasoning.
Proposes Entropic Projection Alignment framework addressing distribution shift: estimating performance, identifying responsible features, and improving target domain accuracy.
Presents COLLEAGUE.SKILL for automated AI skill generation via expert knowledge distillation from heterogeneous traces for person-grounded agents.
Addresses representation fragmentation in few-shot layout-to-image generation by disentangling semantic identity from visual details.
Introduces GLIDE, open-source Python library for prediction-powered inference combining human annotation and LLM-as-judge for reliable agentic systems evaluation.
Presents Contextual Scalarisation Thompson Sampling for multi-objective recommendations in public media adapting to changing priorities.
Proposes activation steering framework for interpretable attribute control in music generation via mechanistic interpretability of Transformer architectures.
Proposes S³LDBO, snapshot single-loop algorithm for decentralized bilevel optimization in networked multi-agent systems.
Introduces CoSee framework analyzing failure modes of shared-memory collaboration in resource-constrained visual agents through noise accumulation lens.
Proposes latent teammate modeling in world models for multi-agent RL to handle teammate-induced uncertainty in cooperative settings.
Controlled study on how skill document granularity affects LLM agent task success using SkillsBench benchmark across multiple models and conditions.
Evaluates LLM agents in bargaining scenarios under different information regimes, measuring performance against game-theoretic solutions and honesty metrics.
Proposes consolidating rewarded perturbations for LLM post-training by sampling Gaussian perturbations and ensembling top-K specialists as alternative to gradient descent methods.
Studies how vision-language embedding models like CLIP represent concept binding in multi-object scenes, showing limitations in cross-modal retrieval.
LongTraceRL: Reinforcement learning approach using search agent trajectories and rubric rewards to improve LLM long-context reasoning with intermediate supervision.
KLIP: Out-of-distribution detection using KL-divergence with diffusion priors for localized distribution shifts in inverse problems.
Technical review of meta-learning methods enabling systems to adapt quickly to new tasks with limited data, covering state-of-the-art approaches and applications.
Survey of graph machine learning integration with LLMs, covering GNN architectures, applications in knowledge graphs, molecules, and emerging LLM-GNN combinations.
Adaptive NAD: Online unsupervised network anomaly detection system for IoT using self-adaptive learning to handle evolving traffic patterns.
Bitween: Automated learning system for discovering randomized self-reductions that previously required expert manual derivation.
Causal representation learning framework for climate analysis combining latent variable discovery with observable-to-observable causal relationships.
Auto-Discovery-Bench: Diagnostic benchmark evaluating AI agents' ability to maintain structured beliefs and discover hidden structures through iterative hypothesis-feedback cycles.
Proposes Bayesian sampling approach for membership inference attacks using Bayesian neural networks to reduce computational overhead of existing methods.
NeUQI: Initialization method for post-training quantization of LLMs to reduce memory/latency on consumer hardware, enabling efficient LLM deployment.
Analyzes DropEdge data augmentation technique for GNNs with theoretical framework explaining limited performance gains in supervised learning.
Proposes methods to attribute model behavior changes across pretraining, fine-tuning, and alignment stages to identify responsibility for AI system successes/failures.