I-BBS extends BBS theory for coordinate-free inference of latent manifolds from distance matrices using random matrix theory, without accessing ambient space.
Observable Matrix Dynamics framework uses random matrix theory to diagnose neural network internal representations and training dynamics via distance matrices.
IG-Lens provides exact additive probability attribution across transformer layers using integrated gradients, improving upon logit lens methods for interpreting decoder-only models.
Diagnostic framework (EPC) for detecting preference collapse in self-adapting LLM agents with multimodal evaluators.
Analysis of how SMOTE and resampling degrade probability calibration in imbalanced classification tree ensembles.
Maritime anomaly detection using equation-grounded synthetic anomalies for AIS vessel tracking data.
ScaleAware-JEPA: self-supervised framework for multiscale physical field representation learning for scientific discovery.
Low-rank mixture models for flow matching to simplify normalizing flow transformations and improve generative modeling.
Vehicle routing optimization for nursing care taxi dispatch combining integer linear programming with machine learning.
PS-PPO: critic-free RLHF method for LLMs using prefix-sampling to optimize long reasoning traces efficiently.
GLIP: joint pretraining framework combining graph neural networks with LLMs for graph-level tasks.
LLM agent for NMR molecular structure elucidation providing interpretable results beyond black-box methods.
MemLeak: method for diagnosing information leaks in multimodal agent memory systems via implicit visual cues.
Theoretical analysis of inlier-memorization effect in early training dynamics for unsupervised outlier detection.
Efficient value-sharing method for Q-learning that accelerates learning by sharing information across actions.
Dual-flow reinforcement learning approach for multimodal action exploration in continuous-control tasks.
First certified unlearning theory for continual learning addressing privacy protection with sequential model updates.
Theoretical framework for understanding data poisoning attacks and defenses in continual learning systems.
Chatbot enhancement using persistent homology for mental health support with privacy-preserving training methods.
RoAd-RL: open-source benchmarking framework for robust adversarial reinforcement learning with standardized evaluation protocols.
Sparse autoencoders applied to vision-language joint embeddings to decompose entangled features into interpretable monosemantic components.
Kernel density estimation method for selecting optimal bandwidth in model calibration for uncertainty quantification in deep learning.
Framework combining deep neural networks with linear model interpretability via local fidelity regularization for more reliable explanations.
Post-hoc OOD detection method using loss-landscape curvature for uncertainty estimation in pre-trained networks without retraining.
DuoMem dual-space distillation framework enabling capable memory-augmented LLM agents on resource-constrained devices.
Study mapping boundaries of reasoning capability elicitation in LLMs across diverse cognitive tasks.
Atompack: append-oriented storage format for large immutable atomistic ML training datasets.
Learned stochastic stopping mechanism for improving length generalization in looped Transformers.
Test-time adaptation for GNNs under distribution shifts via gradient rotation without labeled data.
Two-phase distillation approach for building multi-task agentic LLMs from separate RL experts.
Diagnostic methodology for industry-scale audio-visual LLM evaluation in video moderation systems.
Data-driven probabilistic framework using Gibbs measures on hierarchical structures for energy-based learning.
Neural Subspace Reallocation reframes continual learning as retrieval-based memory management over LoRA subspaces.
Study showing JEPA-style predictive objectives discard exogenous control-relevant features in learned representations.
GAIA framework for global online data selection in LLM instruction tuning via Gaussian processes.
Query-aware spreading activation method for multi-hop retrieval over knowledge graphs in Graph RAG systems.
ML research on robust strategic classification when agents manipulate inputs under cost uncertainty.
Federated learning framework using CRFs to optimize client aggregation weights with heterogeneous data.
NMO benchmark for generative molecular design beyond drug discovery, addressing domain transferability in ML.
Analysis of neural network training dynamics by studying how Hessian eigenvectors evolve and affect learning trajectories.
Scalable batch Bayesian optimization framework using Boltzmann distribution sampling for parallel simulation workflows.
Unsupervised community detection algorithm using discrete Forman-Ricci curvature for heterophilic graphs via sheaf diffusion.
Knowledge-informed fine-tuning of tabular foundation models using knowledge graphs for improved performance in niche domains with scarce data.
Theoretical analysis of token acceptance conditions in speculative decoding with greedy decoding and tree-based candidates for practical LLM acceleration.
Methods using LoRA variants for continual learning in motion-language agents handling both motion-to-text and text-to-motion without catastrophic forgetting.
Hybrid active-online learning framework for optical network failure detection adapting to concept drift with margin-based selective labeling.
Encoder-decoder architecture for in-context learning on tabular data producing target-agnostic row embeddings reusable across diverse downstream tasks.
Self-distillation method for LLM reasoning that routes training by problem difficulty and maintains success buffer for stable improvement without external supervision.
Analysis showing parameter-level defenses against model merging are vulnerable due to small task vector magnitudes enabling reconstruction attacks.
Method for aligning generative flow models via online RL that addresses trajectory likelihood tractability and training-inference inconsistencies.