The Archive

Weakly supervised semantic segmentation for seagrass habitat mapping in side-scan sonar imagery using image-level labels.

RAT: Bayesian framework for evaluating RAG systems by jointly modeling retrieval, abstention, and answer correctness across pipeline components.

Sociological study of AI sensemaking via text analysis of millions of articles and interviews with 57 AI professionals in 2021–2023.

SkillForge: Framework for continuous skill evolution in RL agents via verification and refinement of reusable knowledge across episodes.

DNN²: Doubly non-negative relaxation for ReLU neural network verification using Burer-Monteiro factorization to close safety guarantee gaps.

Hardware-realized Turing Machine with autonomous multi-step execution and reprogrammable optical input for extended computations.

Meta^n: Recursive self-improvement framework for LLM agents that increases meta-depth beyond two by applying fixed meta-operation to its own outputs.

Minimax-optimal alternating regret bounds for online linear and convex optimization with O(log d) constant regret independent of horizon T.

Parameter-efficient adaptation of EEG foundation models via self-supervised learning using only 9% of parameters for clinical transfer tasks.

Arc-length velocity polygonal synthesis extends audio generation to arbitrary shapes, morphing, and 3D polyhedra via unified DSP pipeline.

KENDO framework integrates ensemble Gaussian processes with disagreement-aware acquisition for Bayesian optimization and active learning without MCMC.

Deep learning super-resolution methods downscale SEVIRI cloud masks and enable cross-sensor cloud mask translation for satellite data.

Lifted inference under approximate commutativity identifies and handles deviation in learned parameters for scalable probabilistic inference.

Constrained hyperparameter optimization for streaming/online learning with bounded search spaces and self-adjusting parameter tuning.

Multilingual spoken hallucination benchmark with 12k samples across English, Russian, Kazakh in text and audio with controlled hallucination types.

Parameter-level symmetry attribution via functional sensitivity analysis determines if learned network symmetries lift to parameter space.

Persian NLP resource review of 34 datasets reframes Persian as annotation-scarce rather than low-resource, with quantitative web visibility data.

Parallel predictive coding training pairs generative and encoding networks for time series forecasting and online anomaly detection without sequential propagation.

Combines on-policy distillation with RL-verified rewards to overcome sparse feedback and teacher-bound performance limits in LLM post-training.

LLMs exhibit poorly calibrated confidence in hidden-information tasks, with high-confidence errors in agentic decision-making contexts.

oFM, a multimodal foundation model trained on 1.67M cancer patients, integrates clinical trajectories, genomics, and pathology for precision oncology.

Wolfram Language paclet synchronizing algorithmic composition timelines across Csound, MusicXML, OSC, and click-track outputs.

A pioneering AI scientist once predicted computers would replace human radiologists. They haven't.

Data leakage in power outage prediction models inflates generalization metrics; identifies spatial, temporal, and event-based evaluation gaps.

Maia 200 AI accelerator delivers 10.145 Pflop/s FP4 via software-defined dataflow architecture, shifting from thread-centric to data-movement-centric design.

Modern LLM inference pipelines apply undisclosed probability redistribution post-training, complicating attribution of observed behavior to weights and alignment.

Markov model analysis of Debussy's Syrinx using sliding-window kernels to extract emergent musical form from symbolic sequences.

Parason exposes subtask and trial parallelism in LLM reasoning, reducing latency of test-time scaling from days to minutes.

Reddit analysis shows Anthropic Claude releases correlate with strongest positive sentiment vs. OpenAI/others; user perceptions shift dynamically with updates.