Procedural Memory Distillation: method for language models to retain and reuse procedural information across episodes for self-improvement through online reflection.
Semi-CoT: framework for semi-supervised chain-of-thought learning that reuses generated reasoning traces as learning signals to improve LLM reasoning capabilities.
Monograph introducing mean field reinforcement learning through Markov decision processes and large-population stochastic control with mathematical framework.
Framework for learning robust room embeddings from reverberant speech with uncertainty quantification using dispersion-calibrated scoring without downstream supervision.
OPINE-World: programmatic world modeling using LLMs and counterexample-guided synthesis to generate data-efficient, reusable environment models for agent adaptation.
Proposes MMAO-Cls using metabolic multi-agent optimization as outer-loop optimizer for joint feature selection and classifier hyperparameter tuning.
AgenticDataBench: benchmark for evaluating LLM-based data agents on automating data science workflows including data wrangling, analysis, and visualization tasks.
Studies adversarial robustness and explainability stability of cybersecurity classifiers using SHAP-based explanations across multiple datasets and attack methods.
Introduces Goggles, a learned module using gradient editing to improve language models' ability to recognize fictional content, addressing the negation neglect problem.
COMFYCLAW: agentic system with self-evolving skill harnesses for image generation workflows, enabling agents to recall patterns and user preferences from prior runs.
Theoretical framework applying wave-particle duality concepts to low-illumination image enhancement via Data Relativistic Uncertainty paradigm.
Full Bayesian reinforcement learning approach via Likelihood-Free Iterative Bayesian Importance Sampling for data-scarce settings.
Decentralized optimization framework for nonsmooth nonconvex problems with communication compression and error compensation.
PARTREP method enabling decoder-only LLMs to learn selective prompt repetition patterns, improving reasoning by redistributing contextual grounding across positions.
Review of Koopman operator theory for linearizing nonlinear dynamical systems, covering data-driven techniques like EDMD and machine learning methods.
Lynx: progressive speculative KV cache quantization technique for accelerating long-context LLM inference in retrieval-augmented generation and agentic systems.
Study of post-hoc calibration methods for semantic segmentation to improve confidence estimate reliability in safety-critical applications.
PhysMani framework coupling physics-principled 3D Gaussian world model with action policy for dynamic object manipulation in embodied AI.
Statistical analysis of k-means clustering with missing data, establishing asymptotic risk bounds and convergence guarantees.
Model-agnostic methodology for measuring lag relevance in time series forecasting using Ghost variables and Shapley values.
Framework auditing multilingual text-to-speech systems against language-specific phonological patterns using classifier-based evaluation.
Object Aligner: configurable JSON schema similarity scoring for measuring LLM output alignment with structured schemas, enabling agentic planning and tool calling evaluation.
Evaluation of Vision-Language Model reliability for medical image quality assessment under image corruption and demographic bias.
Scalable distributed algorithm for computing silhouette coefficients to assess k-clustering quality on large datasets.
Adaptive reinforcement learning approach for zero-shot cross-platform control of autonomous surface vehicles with unknown dynamics.
Maven RL framework with editable evidence memory for long-context reasoning, rewarding intermediate evidence state changes rather than just final answers.
Open-weights constitutional classifier for multilingual AI safety filtering, achieving SOTA on prompt-safety benchmarks at 1/10th the size of competing models.
Hierarchical state-space model for video prediction using surprise-based chunk boundary detection instead of fixed-length or similarity approaches.
Emotional Self-Correction method improves vision-language model reliability by activating latent self-correction without post-training.
WBMM efficiently implements large kernel depthwise convolutions via windowed batch matrix multiplication.
Structured Gaussian process method for high-dimensional omics classification with small samples and class imbalance.
Deep learning method for medical image segmentation of penile tissue in MRI for reproductive health phenotyping.
Physics-based synthetic pipeline for film degradation and archival restoration benchmark with temporal coherence.
Proves tight lower bounds for multi-secretary online selection problem using Bellman certificates.
Evaluates frontier LLMs on expert-authored clinical reasoning scenarios, showing open-ended medical performance remains unsolved with 32% hard subset score.
Proposes Fourier-based preconditioning for mutual information-inspired feature learning, proving H-Score invariance properties.
Studies uncertainty quantification via conformal prediction for counterfactual decision-making in high-stakes applications.
Additive deep learning framework for drug discovery that separately models chemical descriptors and molecular graph structure for solubility prediction.
Addresses failure of on-policy self-distillation on long chain-of-thought reasoning, proposing method to maintain model thinking capability.
Proves aggregation with exponential weights is minimax-rate optimal in expectation for model selection, settling open problem from 2013.
Enables in-context learning in spiking neural networks via dendritic computation, making biologically plausible SNNs pass Garg-2022 ICL benchmark.
Redesigns symbolic parser backend using CCG directed types for improved structural generalization on SLOG benchmark with 30K parameters.
HNSW search framework adding theoretical correctness guarantees to hierarchical navigable small world graphs via graph spanner verification.
Proposes quantum sequence modeling using variational circuits with self-modulating gates and bounded memory for stable long-sequence processing.
Studies whether LLM personas from psychometric questionnaires are intrinsic or frame-dependent using geometric analysis on manifolds.
NASA deploys agentic search system using LLMs to help geoscience researchers discover relevant datasets and tools from thousands of available resources.
WattGPU predicts power consumption and latency for LLM inference across unseen GPUs without exhaustive profiling, addressing data center energy optimization.
arXiv paper on fast multi-dimensional refusal subspace extraction in LLMs for safety and interpretability.
arXiv paper on object-centric LeJEPA for more data-efficient self-supervised image representation learning.
Q-GAIN Python package for machine learning and physics-informed analysis of cold-atom experiment images with classification and detection.