The Archive

Deep neural network regression under covariate shift from dependent data using quantile and Huber loss.

Quantum autoencoders for anomaly detection in high-energy physics compiled to FPGA hardware accelerators.

PortLLM fine-tuning portability via LoRA explained by high-dimensional near-orthogonality across continual pretraining.

232k dataset-model-app chains show license obligations stripped in AI supply chains; 'license laundering' quantified.

Multi-robot task planner anticipates future tasks to reduce cumulative cost in shared persistent environments.

PAC bounds on LLM harmful output probability via latent-space-guided tree exploration and Clopper-Pearson intervals.

Interpretable Mamdani fuzzy regression extension for Ex-Fuzzy library with target-aware partition initialization.

18 medical image encoders on 650k radiographs show self-supervision drives representational convergence more than clinical labels.

Study evaluates 21 instruction-tuned LLMs on recognizing Schwartz's ten basic values in 1,000 Russian texts; finds Acc@1 of 0.683.

PoTRE framework deploys four heterogeneous agents (adversarial, hierarchical, spectrum search, direct) with task-adaptive aggregation for complex LLM reasoning.

Maskability Index (MI) metric predicts alignment between knowledge relations and prompting strategies (masked vs. prefix) in T5/BERT pretraining.

Theoretical algorithm achieves O(T^7/10) regret for CDF-related online learning objectives, improving prior O(T^3/4) bound in 2D.

Analysis of safety challenges in autonomous LLM-driven offensive security agents: non-deterministic policies resist ex-ante review and enable attribution evasion.

Unified music generation framework combining hierarchical autoregressive planning and flow-matching for full-song synthesis from lyrics and descriptions.

Using Filtered-Corpus Training, study shows LLMs acquire unlike coordination (non-same-category conjunctions) without direct exposure, suggesting emergent compositionality.

Evaluates LLM-based machine translation on culturally loaded expressions using Dream of the Red Chamber Chinese-Japanese dataset; identifies systematic gaps.

Bayesian online learning framework aggregates expert update rules (learning rates, priors, variational families) to match best expert on data streams.

PhaseAware compact temporal model with phase/body-group descriptors for rehabilitation scoring; achieves 88.9% RMSE reduction on UI-PRMD deep-squat protocol.

Levi-Civita coordinates outperform Cartesian formulations for learning Hamiltonian dynamics in perturbed Kepler systems with high eccentricity.

PIER framework augments retrieval-augmented generation with physics constraints to improve transfer learning for environmental time-series modeling.

Personal travel tip about activating mechanical instruments at a San Francisco museum.

Hybrid CoLES + Mamba approach for user-centric transactional sequence modeling combines contrastive learning with selective state-space models.

DQAOA-GPT integrates distributed quantum approximate optimization with GPT for scaling variational quantum algorithms to large combinatorial problems.

AMD says it's going to invest up to $5 billion in Anthropic, while helping to expand the AI company's computing power, according to an announcement on Wednesday. As part of the new partnership, Anthropic will deploy up to 2 gigawatts of AMD's Instinct MI450 AI GPUs using the chipmaker's new Helios rack-scale system, as reported earlier by The Wall Street Journal. The companies plan to deploy the first gigawatt in the first half of 2027, building upon recent data center deals Anthropic has reached with SpaceX and TeraWulf. Anthropic has also signed AI infrastructure deals with Google, Broadcom...

HalluTruthQA: 2,400 expert-curated examples for fine-grained hallucination detection, localization, and explanation in Arabic LLM question answering.

Orchestrated open-weight small LLMs achieve malware analysis performance competitive with frontier closed-weight models at lower computational cost.

ELSAA: efficient attention mechanism combining low-rank and sparse approximations without decomposing projection matrices, extends Transformer context length.

Surprisal theory critique: LLM-based surprisal metrics obscure representational choices; argues black-box models don't eliminate computational-level design requirements.

Statistical framework for LoRA rank allocation: formulates importance-score design with explicit statistical interpretation for parameter-efficient fine-tuning.

Quadrilateral loss: differentiable penalty measuring additivity in neural networks via second-order mixed differences, connects to Shapley-GAM interaction quantification.