arXiv: Open-source biomedical knowledge graphs (Pathways, Clinical Trials, Drug-Gene) with AI agent access via Samyama database.
arXiv: Addresses LLM limitations in private-library code generation; shows API documentation retrieval alone is insufficient.
arXiv: HindSight framework evaluates LLM-generated research ideas by matching against future publications and citation impact.
arXiv: Analyzes how wider beam search can hurt LLM output quality due to overestimation bias in noisy scoring.
arXiv: RSGen framework for remote sensing image generation using diffusion models with layout and edge guidance constraints.
arXiv: PokeAgent benchmark for multi-agent AI decision-making with partial observability, game theory, and long-horizon planning in Pokemon RPG.
arXiv: Physics-informed neural networks for simulating EUV electromagnetic wave diffraction in lithography. Domain-specific neural networks.
arXiv: Analyzes tokenization design choices for foundation models trained on structured electronic health records.
Reinforcement learning framework extending RLHF with multi-dimensional contextual rubric rewards and alternating optimization.
Inference-time steering mechanism for frozen LLMs using adaptive prompt routing to enable evolving safety alignment without retraining.
Prototype-based OOD detection method with dynamic prototype count adaptation based on category complexity.
Federated learning framework combining knowledge graphs and temporal transformers for early sepsis prediction across multi-center ICUs.
Study of Gini Index role in detecting and debiasing class accuracy disparities in prompt-based classification tasks.
Defense mechanism against steganographic collusion in multi-agent RL using dynamic representational circuit breaking at optimization substrate.
Attribution-guided framework using rank-one model editing to rectify unreliable neural network behavior on non-robust features.
Analysis of transformer training dynamics via spectral edge detection showing parameter updates concentrate in few coherent directions.
Graph neural network for flash flood susceptibility mapping using watershed connectivity with conformal uncertainty quantification.
Domain adaptation method for remaining useful life prediction with incomplete degradation trajectories using evidential learning.
Hypergraph neural network approach using Ricci flow to address over-smoothing and improve message passing.
Multi-expert framework with uncertainty guidance for imbalanced sequence learning and minority class detection.
Method bridging learned embeddings and interpretable handcrafted features for temporal event sequences in financial systems.
Metacognitive test-time reinforcement learning framework for unified multimodal models enabling knowledge accumulation across similar prompts.
Physics-grounded multimodal LLM agent combining language models with PDE solvers for scientific reasoning without domain-specific fine-tuning.
Zero-shot forecasting method for time series with exogenous variables using prior-fitted networks.
Masked data training paradigm for discrete diffusion language models using information density-driven noise scheduling.
Breast cancer risk prediction model using longitudinal mammography with incomplete or irregular screening histories.
Empirical study showing prediction-equivalent ML models produce substantially different feature attributions across 24 datasets, challenging assumptions in explainable AI.
Evaluates LLM failure modes in scientific decision-making when stability doesn't guarantee agreement with statistical ground truth.
Privacy-preserving machine learning technique using informational compression for anonymization without performance degradation.
Applies optimal transport theory to evaluate ML model vulnerabilities through Wasserstein-constrained data perturbations.
Proposes counteractive reinforcement learning approach addressing computational complexity in high-dimensional MDPs.
Evaluates electrodermal activity signals for aerobic exercise detection in wearable sensors.
Python library for unit circle based computing using complex phasors and unitary wave interference gates.
Federated learning approach for Alzheimer's disease classification on MRI data with site-aware partitioning and privacy guarantees.
Hybrid approach combining game theory and reinforcement learning for adversarial scenarios using analytical solutions for early termination.
Practical guide for using AI systems and agents in mathematics and machine learning research with discussion of responsible guardrails.
Analyzes 10,469 experiments from LLM agents performing architecture search vs hyperparameter tuning using ANOVA decomposition.
Introduces diagonal flow matching for inverse design problems with better stability than conditional flow matching.
Benchmarks causal discovery algorithms on synthetic healthcare data for structural recovery and fairness decomposition.
Data-driven framework learning interaction kernels in stochastic multi-agent systems via sparse regression on trajectory data.
LLM-guided neural architecture search for time-series classification in privacy-constrained domains using data-local constraints.
Hardware-in-the-loop architecture search methodology for designing efficient on-device LLMs with real-time latency constraints for mobile deployment.
Proposes guided asymmetric self-play method for post-training coding LLMs with better problem selection to improve model capabilities.
Derives hyperparameter scaling laws for modern optimizers enabling transfer across model sizes, batch sizes, and training horizons.
Develops continuous-time framework for spiking neural networks using charge conservation to achieve deterministic computation in temporal stochasticity.
Analyzes whether LoRA checkpoint weights encode task performance information readable without running the base model, enabling efficient adapter analysis.
Theoretical work on smooth calibration as robust calibration measure and step toward omniprediction guarantees.
Reinforcement learning approach for temporal feature generation in cross-user activity recognition from wearable sensor data.
Bayesian method with Tucker decomposition for adaptive regularization in high-dimensional inverse problems.
Masked diffusion model optimization using binary encoding and index shuffling for improved scaling of diffusion language models.