Spectral Compact Training (SCT) for LLMs on consumer hardware. Uses permanent truncated SVD factors to avoid materializing dense weight matrices during training.
Multiscreen attention mechanism for language models. Introduces absolute query-key relevance to reject irrelevant keys, addressing softmax attention limitations.
Adaptive stopping mechanism for multi-turn LLM reasoning. Determines optimal stopping points for agents using retrieval-augmented generation and ReAct-style interactions.
Vision-based robotic process automation (RPA) using sequential Monte Carlo localization. Enables stable GUI automation from single demonstrations with improved robustness.
Framework for analyzing agent communication protocols across three layers: communication, syntactic, and semantic. Systematically studies 18 representative protocols for LLM systems.
Method scaling determinantal point processes for RAG systems to improve diversity of retrieved context while maintaining relevance.
Framework for monitoring safety of tool-using LLM agents through latent reasoning that decouples safety judgment into trainable stages.
Novel neural network architecture for solving PDEs addressing limitations of physics-informed neural networks.
Binary encoding scheme for ternary neural network weights enabling efficient storage and computation for compressed LLMs.
AI framework combining spatio-temporal and graph learning for electricity theft detection in smart grids.
Analysis of computational efficiency for Kolmogorov-Arnold Networks on hardware-constrained deployment scenarios.
Multi-stage pipeline combining experimental design and machine learning surrogates to explore agent-based models efficiently.
Cross-scale evaluation of LLMs on biomolecular modeling tasks revealing performance gaps compared to mechanistic understanding.
Bayesian fine-tuning method for LLMs using low-rank adapters to improve uncertainty quantification in safety-critical applications.
Using LLMs and vision models trained on human preferences to improve network visualization aesthetics beyond traditional heuristic metrics.
Control-theoretic analysis of state-space model robustness under adversarial perturbations, examining Spacetime SSM forecasters and Kalman filter representations.
MetaSAEs: sparse autoencoder training with decomposability penalty producing more atomic, single-concept latents for safety-relevant LLM applications.
OLMo Hybrid: theoretical and empirical analysis of hybrid models combining linear RNNs and attention as alternatives to pure transformers with scaling benefits.
Neural operator methods for multi-task optimal control problems, mapping task descriptions to control policies using permutation-invariant architectures.
Benchmark of Earth embedding models (AlphaEarth, Prithvi, Clay) for neighborhood-scale urban monitoring from satellite imagery.
Analysis of trajectory prediction models revealing that surrounding agents often degrade accuracy due to learned confounders, using Shapley attribution.
Study of data intervention techniques for improving fairness across demographic subgroups in ICU prediction models using real healthcare data.
Data-driven approach using trained autoencoders as fast projectors to enforce complex nonconvex operational constraints in learning and control systems.
Empirical survey comparing regularization frameworks (Ridge, Lasso, ElasticNet, Post-Lasso) across 134,400 simulations with historical development context.
BlazeFL: lightweight federated learning simulation framework enabling fast, deterministic training of hundreds or thousands virtual clients on single node.
Neural approach for black-box global optimization from noisy samples using iterative refinement to avoid local minima in multi-modal functions.
Reinforcement learning approach for handling delayed feedback by replacing state augmentation with homomorphic methods to reduce sample complexity.
Mechanistic interpretability method for discovering repeated attention patterns in large language models at scale without resource-intensive controlled settings.
CountsDiff: diffusion model framework for generating and imputing count-based discrete ordinal data using survival probability schedules.
Framework for automated mathematical conjecture resolution combining LLMs with formal verification to improve reliability of research-level mathematical problem solving.
Research on representational collapse in multi-agent LLM committees using majority voting, measuring agent diversity via cosine similarity and effective rank on mathematical reasoning tasks.
k-Maximum inner product attention for graph transformers addressing quadratic complexity while maintaining expressive power of GraphGPS.
DDCL-Attention: Prototype-based readout layer for transformer encoders using soft probabilistic token matching for compact summaries.
Bayesian information-theoretic approach to training data attribution for tracing model predictions to influential training examples.
Method for input-dependent layer selection in steering vectors to improve LLM alignment at inference time, adapting intervention layer per input.
SODA: Semi on-policy knowledge distillation method for LLMs balancing off-policy simplicity with on-policy effectiveness without adversarial training instability.
Theoretical research on multi-task representation learning for reinforcement learning with shared representations across related RL tasks with different rewards.
Framework combining structure pretraining with diffusion models for generating molecular dynamics trajectories with limited MD data.
ACES: method for selecting LLM-generated code using LLM-generated tests via leave-one-out AUC consistency without determining test correctness.
Low-bit mixed-precision attention kernel using MXFP for efficient transformer inference with reduced memory bandwidth.
BWTA: binarized transformer quantization scheme with ternary activations and algorithm-hardware co-design for efficient inference.
Analysis of LLM reasoning models under noisy labels in reinforcement learning with verifiable rewards, identifying label noise vulnerabilities.
ArrowFlow: novel ML architecture operating in permutation space using ranking filters and permutation-matrix updates without gradients.
Generalization analysis of stochastic bilevel optimization with applications to hyperparameter optimization, meta-learning, and RL.
Spectral Path Regression using directional Chebyshev harmonics for interpretable learning on tabular data without exponential scaling.
Analysis of geometric alignment cost in scientific foundation models for biology/physics, showing discrete tokenization degrades continuous geometry preservation.
Framework for uncertainty-aware foundation models on clinical data, addressing incomplete and irregular measurements in healthcare.
ClawArena benchmark for evaluating AI agents in dynamic environments with evolving information, contradictions, and implicit user feedback.
Graph-assisted retrieval framework for reasoning about defects in laser powder bed fusion manufacturing using structured scientific knowledge.
Framework using Temporal Behavior Trees to repair suboptimal trajectories before using them for robot control policy learning.