Large-scale multimodal dataset and benchmark for plant disease detection using vision-language models in agricultural domain.
Benchmark dataset (DTBench) for document-to-table extraction evaluating LLM ability to produce structured tables from unstructured documents.
Empirical comparison of social network topology between Reddit and Moltbook, an AI-agent-driven social platform, examining agent-mediated systems.
Adaptation of LLaVA framework for Polish language vision-language models using automated pipeline without manual annotations.
Multi-agent simulation studying how community discussion and social memory improve LLM-generated comedy writing in controlled sandbox environment.
Theoretical analysis of sample complexity for constrained MDPs in reinforcement learning with safety constraints for real-world applications.
Hybrid approach combining time series embeddings with statistical features using Granite TinyTimeMixer for HVAC equipment anomaly prediction.
Extension of neural ODEs to variable-dimensional spaces (M-polyfolds) enabling flexible geometric deep learning beyond fixed-dimensional manifolds.
Self-supervised representation learning from sparse, irregularly-sampled electronic health records using dual-masked autoencoding for clinical tasks.
Analysis showing vision-language models outperform text-only LLMs on pure text retrieval tasks due to visual training correcting binding shortcuts.
Neural operators trained with physics knowledge to improve data efficiency and generalization for modeling PDE-governed physical systems.
Transformer compression technique using calibration-optimized matrix orthogonalization instead of SVD for improved accuracy at moderate compression rates.
Framework for learning reward functions from heterogeneous feedback types (demonstrations, comparisons, ratings, stops) using Bayesian inference and amortized variational inference.
Study of multilingual data curation across 13 languages for foundation models, addressing performance interference and uneven data availability in multilingual training.
Research on automatically detecting biases in reward models used for LLM post-training, proposing iterative LLM-based approach to identify spurious attributes like length, format, and hallucinations.
tensorFM model for efficient prediction on tabular categorical data by capturing high-order feature interactions with low-rank approximations.
Method to improve adversarial training robustness in LLMs by addressing distribution gaps that cause vulnerability to simple prompt variations.
Theoretical analysis of Graph Transformers with GNN-based positional encodings through manifold limit models for graph sequences.
First systematic study of neural scaling laws for masked-reconstruction transformers on single-cell RNA sequencing data using CELLxGENE Census.
On-policy distillation method sampling from reasoning prefixes to reduce training cost while maintaining benefits of token-level teacher supervision.
Quantum-inspired classification head using complex-valued unitary representations in Hilbert space for improved uncertainty quantification in neural networks.
Information geometry analysis of softmax distributions examining how semantic structure encodes into representation spaces for model behavior.
Hybrid federated and split learning framework for privacy-preserving clinical prediction keeping feature extraction on clients and heads on coordinator.
Analysis showing random masking of parameter updates in adaptive optimizers like RMSProp outperforms state-of-the-art, inducing curvature-dependent regularization.
Study of prescriptive scaling laws for foundation models estimating capability boundaries and downstream accuracy given pre-training compute budgets.
Curiosity-driven game-theoretic framework for multi-label classification handling long-tail label distributions in large-scale data mining applications.
Interpretability method identifying critical reasoning turns in language models using directional trajectory change to reveal causal relationships in reasoning.
Framework modeling behavioral staleness in asynchronous federated learning to address performance degradation from asynchronous training delays.
Method to discover implicit LLM alignment objectives causally tied to model behavior, addressing reward misalignment risks beyond pre-defined rubrics.
First systematic study of black-box adversarial attacks targeting memory injection in memory-augmented LLMs through similarity-based retrieval mechanisms.
Reinforcement learning framework incorporating cerebellar and dendritic computational strategies to improve sample efficiency and generalization under partial observability.
Fractional-order variant of federated averaging addressing slow convergence, communication costs, and non-IID data distribution challenges.
Extension of mean-shift clustering introducing randomness in both trajectory updates and kernel bandwidth for improved handling of data-scarce regimes.
Framework coupling diffusion models for signal enhancement and classification jointly, integrating semantic information from classifier outputs during denoising.
Multi-agent reinforcement learning approach for sequential social dilemmas using fairness incentives without global information access.
Theoretical analysis showing logit distance bounds representational similarity in discriminative models including language models.
Benchmark protocol for IoT anomaly detection with event-level evaluation under realistic perturbations like sensor dropout and drift.
Evaluation framework examining how chain-of-thought reasoning generalizes in multimodal LLMs on visual planning tasks, testing out-of-distribution robustness.
Meta-learned optimizer predicting coordinate-wise step sizes to address hyperparameter sensitivity in gradient-based optimization for non-convex settings.
Addresses optimal query allocation for LLM-as-judge evaluation under fixed computational budget to minimize estimation error in prompt-response scoring.
Develops methods to compute exact Lipschitz constants for piecewise linear neural networks for robustness guarantees and regularization.
Analyzes geometric coherence issues in federated learning with graph neural networks when aggregating heterogeneous client updates.
Presents conditional entropy-penalized autoencoders for counterfactual inference on time series data in finance, healthcare, and marketing applications.
Introduces K-means quantization approach for quantization-aware training to reduce LLM memory footprint while maintaining performance in low-bit regimes.
Proposes feedback alignment method to improve training efficiency in predictive coding networks by addressing vanishing gradient problems in early layers.
Uses neural networks to estimate parameters in large-scale agent-based labor market models, addressing computational constraints in ABM parameter exploration.
Per-instance unlearning using adaptive noise calibration with differential privacy guarantees and individual sensitivity bounds.
GLM-5 foundation model for agentic engineering with reasoning and coding capabilities, using DSA for efficiency and asynchronous RL for alignment.
Study of how fine-tuning aligned LLMs on benign tasks degrades safety guardrails due to unstable orthogonality in parameter space.
Deep RL applied to reachability problems in control tasks, addressing mismatch between maximizing safe state sets and optimizing expected returns.