LLM-automated system for constructing knowledge graphs and generating adaptive educational questions, addressing scalability limitations of manual curation.
Pareto-Conditioned Diffusion framework for offline multi-objective optimization via conditional sampling from diffusion models on static datasets.
Deep learning models for multi-horizon electricity price forecasting in volatile markets, comparing state-of-the-art time-series architectures.
Analysis of optimization instability in deep neural networks caused by parametric singularities, proposing gradient Frobenius norm solutions.
Study explores how LLMs support healthcare professionals in analyzing patient-generated health data from wearables and smartphones for cardiac risk reduction.
AEGIS method for concept erasure from diffusion models balances robustness against concept reactivation with retention of model utility under adversarial conditions.
Framework establishes first scaling laws for LLMs in recommendation systems using principled synthetic data to replace noisy user interaction data in pre-training.
Learnable Chernoff Baselines enable efficient inference-time reward-guided alignment for generative models without architecture modifications or computational overhead.
Bielik Guard provides efficient safety classifiers for Polish-language LLM content moderation, offering 0.1B and 0.5B parameter model variants based on RoBERTa.
GISA benchmark evaluates information-seeking AI agents performing multi-turn web interactions, addressing limitations of existing benchmarks with naturally-constructed real-world tasks.
LLaDA2.1 improves text diffusion model decoding speed by combining Token-to-Token and Mask-to-Token editing schemes for 100B-parameter block-diffusion models.
Deep learning model with explainability for predicting open source software project sustainability from temporal contribution patterns.
Foundation model for immune system combining multimodal patient-level data to capture multicellular interactions in immune diseases.
Traffic control strategy using jam-absorption driving to mitigate stop-and-go waves on freeways.
Method for synthesizing diverse post-training data for LLMs using Feature Activation Coverage metric based on model representations.
Investigation of learning dynamics in transformers showing training trajectories collapse onto low-dimensional manifolds in modular arithmetic tasks.
Security analysis of mobile LLM agents identifying vulnerabilities in screen-interface paradigm and proposing intent-centric architecture.
Computational analysis of temporal experience and unpredictability in autism using phenomenological and NLP methods.
Block floating-point data format for efficient LLM inference achieving 4.5 bits per value with hierarchical scaling.
Method for editing 3D Gaussian Splatting representations with improved cross-view consistency and editing flexibility.
Behavioral study examining user interaction with three LLM assistance modalities (Advisor, Coach, Delegate) in multi-party negotiation scenarios.
Lightweight 5B parameter multimodal model for image generation and editing competitive with much larger models.
Framework for synthesizing and optimizing CUDA kernels using LLMs combined with systematic exploration to achieve competitive hardware performance.
Method for identifying queries that cause LLM character specification violations using red-teaming approaches to detect deployment-level failures efficiently.
Research on adaptively merging multiple LoRA modules from open models to improve performance on downstream tasks.
Physics-informed neural network solving the Nirenberg differential geometry problem of prescribing Gaussian curvature on surfaces using mesh-free approach.
Knowledge-guided world model using reinforcement learning for simulating opioid policy interventions and counterfactual analysis of public health decisions.
Exploration method in reinforcement learning using ensemble error estimates to compute optimistic value bonuses for directed agent exploration.
D3-Net framework for longitudinal treatment effect estimation using deep learning with debiasing to address error propagation in iterative conditional expectation.
Hybrid machine learning ensemble framework combining temporal fusion transformer, attention-BiLSTM, and XGBoost for Bitcoin price forecasting.
Analyzes implicit bias in deep network training through Jacobian spectral properties, establishing theoretical signatures of depth-induced scaling and singular-vector alignment.
Theoretical analysis showing neural networks with trainable low-degree rational activation functions are more expressive and parameter-efficient than standard activations.
Active learning approach for high-dimensional level set estimation using trust regions and double acquisition functions to minimize function evaluations.
Addresses LLM personalization by generating synthetic user-specific interaction data at scale to optimize prompts for individual user preferences and constraints.
Applies deep reinforcement learning to automate analog and mixed-signal circuit design optimization across diverse non-differentiable design spaces.
Studies soft contamination in LLM training data through semantic duplicates, showing typical decontamination filters fail to detect near-equivalent benchmark test data.
Demonstrates stable training of LLMs from scratch using exclusively low-rank weight factorization, matching dense model performance while reducing computational costs.
Safe reinforcement learning framework using recovery-based shielding with Gaussian process models for non-linear continuous control with provable safety guarantees.
Training-free guidance method enabling continuous diffusion language models to satisfy formal syntactic constraints like JSON schema matching via regular expressions.
Regularized meta-learning framework addressing redundancy and overfitting in deep ensemble methods through redundancy-aware projection and statistical weighting.
RNA sequence design reframed as conditional sequence generation task using language models instead of traditional optimization approaches.
Theoretical analysis of logistic regression optimization with large stepsize gradient descent, proving tight convergence bounds for separable data in low dimensions.
Theoretical analysis of Mamba state space model training dynamics and generalization properties.
Analysis of robustness and reasoning consistency in vision language models fine-tuned with reinforcement learning.
Bench-MFG: standardized benchmark suite for mean field games and multi-agent reinforcement learning.
Multi-agent model-based RL with joint state-action learned embeddings for coordination in dynamic environments.
Constraint-rectified training method to reduce overhead and overthinking in chain-of-thought reasoning.
Analytical results for exponential family distributions in hierarchical Dirichlet process models.
Flow-Factory: unified modular framework for reinforcement learning with flow-matching and diffusion models.
Adaptive steering method to balance modality preferences in multimodal large language models.