Theoretical analysis of generalized Euler logarithm properties with applications to natural gradient, backpropagation and optimization.
Buffer-free continual learning framework exploiting network redundancy for class and task-incremental learning without replay memory.
Framework for bi-criteria combinatorial optimization with noisy evaluations and bandit feedback.
Hyformer transformer model jointly optimizing molecule generation and property prediction with synergistic benefits.
Research on identifiability challenges in learning sparse linear differential equation models from data.
Survey of LLM applications to time series analysis tasks including forecasting, anomaly detection, interpretation and captioning.
Foundation model for inductive link prediction on knowledge hypergraphs with novel entities and relation types using hypergraph embeddings.
Method for automatically discovering learning-friendly generation orders in autoregressive models by identifying orders with faster early-stage loss drops.
MaPPO: preference optimization for LLM alignment incorporating prior reward knowledge into Direct Preference Optimization framework.
Equilibrium Propagation with intermediate error signals for biologically-inspired local learning in convolutional recurrent neural networks. Reduces BPTT computational cost.
Normalized Maximum Likelihood code-length formulation for Riemannian data spaces including hyperbolic geometry for hierarchical graph structures.
Scalable Option Learning: hierarchical RL algorithm achieving 35x speedup in high-throughput environments for long-horizon decision-making.
Comparative study of ensemble voting and stacking methods for obesity risk prediction using machine learning. Healthcare application with limited novelty.
SpikingBrain: spiking neural network-based large models addressing quadratic training and linear inference complexity bottlenecks in Transformers for long-context.
Hammer and Anvil: theoretical framework categorizing backdoor attacks in federated learning and identifying fundamental defenses against adaptive adversaries.
Fine-tuned LLaMA 3.2 vision-language model applied to neutrino event classification in high-energy physics detector data.
Inverse reinforcement learning method using classification and regression instead of specialized architectures. Addresses reward identifiability in maximum-entropy models.
BoHA: blockwise Hadamard adaptation for parameter-efficient fine-tuning of LLMs. Addresses sequential adaptation and task retention vs. single-task accuracy.
Fidelity benchmark for multimodal time series forecasting addressing data contamination and leakage issues in existing datasets.
Two-stage adaptive personalization framework for foundation models in federated learning with heterogeneous tasks and modalities. Original research on PEFT.
ThinKV: adaptive KV cache compression framework for reasoning models using thought-type detection and hybrid quantization.
Flock: knowledge graph foundation model using random walk learning for zero-shot link prediction on novel entities and relations.
Frequency-Aware attribution method for neural networks using selective high/low-frequency perturbations for improved explainability.
Closed-form optimization method for neural network last layers using known linear solution during training.
Amortized multi-objective optimization across task families using generative solution modeling for expensive multi-objective problems.
W4A4 LLM quantization method using closed-form rotations on Stiefel manifolds to address convergence issues in ultra-low-bit quantization.
FaVeX: algorithm for computing verified explanations of neural networks faster by dynamically combining verification techniques.
ATHENA: agentic framework for autonomous scientific computing and ML research. Hierarchical evolutionary algorithms loop for end-to-end computational lifecycle management.
Predictor-Corrector framework using neural controlled differential equations to correct forecast errors in learned time-series models.
Exact Flow Linear Attention: novel attention mechanism using continuous-time dynamics and closed-form solutions for improved efficiency in transformers.
DT-PBO: interpretable tree-based surrogate model for Bayesian optimization with preference learning. Addresses interpretability in high-stakes domains.
DiffeoMorph: differentiable framework for learning 3D shape morphing using agent-based distributed control. Application to developmental biology and multi-agent systems.
Bloom filter-based data encoding method for ML that compresses features into fixed-length bit arrays with optional keyed hashing for obfuscation.
Deep Bayesian RL using generalised linear models with learnable basis functions for meta-RL task adaptation.
SB-TRPO algorithm for safe reinforcement learning satisfying hard zero-cost safety constraints while optimizing rewards.
FANoS-v2 PyTorch optimizer augmenting RMS-preconditioned momentum with feedback controller and thermostat damping.
DeepFedNAS framework for efficient hardware-aware neural architecture search across heterogeneous IoT device federations.
ART: reinforcement learning approach to optimize diffusion model time discretization schedules for efficient sampling.
Neural scaling laws study showing diverse task-level scaling behaviors diverge from aggregate validation loss predictions.
Establishes formal correspondence between sparse attention mechanisms in transformers and compact kernel regression.
Minerva applies reinforcement learning with verifiable rewards to improve cyber threat intelligence LLM structured output generation.
ESSAM: evolution strategies with sharpness-aware maximization for memory-efficient LLM fine-tuning via reinforcement learning.
TopoPrune data pruning method using latent space topology to improve stability during cross-architecture transfer and with feature noise.
SLOPE framework for model-based RL using optimistic potential landscape shaping to improve planning in sparse reward environments.
Continual learning framework (CaRE) using bi-level routing mixture-of-experts for class-incremental learning on 300+ tasks with pre-trained models.
Recurrent neural network study on how predictive models learn world models with path integration and object-location binding in token scenes.
Chess mixture-of-experts model with small GPT experts routing to different player styles, addressing mode-averaging in dense transformers.
arXiv paper introducing SOCKET, a soft collision kernel estimator for sparse attention in long-context LLM inference using locality-sensitive hashing.
arXiv paper analyzing why sign-based optimization algorithms like Lion and Muon outperform AdamW for training LLMs under heavy-tailed noise.
MCP server for biopharma research that integrates LLM with multi-source data aggregation from FDA, SEC, PubMed for catalyst analysis.