The Archive

HiveTraceGuard-Pro, a 0.6B LoRA guardrail model tuned on Qwen3, detects Russian and English prompt injection and jailbreaks with binary safety scoring.

AgentFactory automates agentic system design and optimization by jointly optimizing workflow architecture and model selection for real-world constraints.

Committed reveal sampling (CRS) makes token selections persistent in discrete diffusion, improving generative quality without retraining.

ViTAMINS integrates synthetic hard negatives into vision transformer pretraining, improving ImageNet and transfer learning with +11.3% gains.

Causal Evidentiary Governance (CEG) framework enables regulated ML systems to commit to versioned causal DAGs partitioning allowable vs. disallowed pathways for fairness compliance.

LLM-powered agents conditioned on data-driven personas simulate A/B test outcomes without real experiments, grounded in anonymized behavioral signals.

Progressive Risk Vesting (PRV) manages capability activation across recursive LLM-agent trees with trajectory-level risk budgets and anytime harm bounds.

Perturbative analysis derives Cohen et al.'s central flow as formal limit of gradient descent at edge of stability; theoretical contribution with limited empirical scope.

Passive learning framework for Metric Interval Temporal Logic synthesis reduces timed specifications to untimed learning problems.

Survey of price extraction methods from websites comparing wrapper-based, rule induction, and browser-based approaches for e-commerce data.

PCoMoE enables fine-grained path composition during MoE inference to reduce intra-expert redundancy and improve LLM throughput.

Language-Contrastive Guidance applies training-free inference to preserve native accents in code-switched text-to-speech synthesis.

Network science framework evaluates deep graph generative models against topological properties of real-world networks for synthetic data validation.

Stratechery analyzes Nvidia's earnings results and business strategy around power consumption, open models, and Hugging Face's role in AI infrastructure.

FigTree multi-agent system generates editable scientific figures as recursive SVG programs for iterative refinement of methodology diagrams.

SinkPruner removes high-norm outlier visual tokens from MLLMs to reduce inference cost while maintaining multimodal understanding quality.

Analysis reveals LLMs misalign on Islamic finance questions when cultural cues override correct framework selection, exposing normative pluralism failures.

Inspicio links historical/low-resource language tokens to English WordNet synsets without source-language sense inventories for word sense disambiguation.

Perceptual study measures alignment between audio similarity metrics and human judgments for music replication detection across musical attributes.

Semi-supervised virtual staining using unpaired histopathology images via morphology preservation and realism constraints.

Disclosure-gated user simulator reduces over-cooperation bias in LLM-based companion agent evaluation via layered information gating.

zbMATH Open Knowledge Graph integrates 250+ years of mathematical scholarship with expert-curated semantic content and disambiguation.

CoBRA learns tool-use boundaries in LLMs via counterfactual margins, optimizing when to invoke tools by estimating instance-level marginal benefit.

Covariance-corrected Mahalanobis distance for few-shot out-of-domain intent detection in conversational agents.

LLM-labeled NER training for Persian customer chat anonymization, comparing DeepSeek-V3, Qwen3, and GPT-OSS labeling quality.

Action-class diagnostic framework decomposes multi-turn tool-calling failures into miscalibration vs. execution modes across open and closed models.

Terminology-driven framework generates large-scale diagram instruction data for improving VLM performance on scientific diagrams.

Embedded conditional independence tests detect whether LLM-generated text carries information beyond source, applied to German parliament speeches.

Dataset for modeling iterative problem-solving tracks solver performance, error persistence, and strategy shifts across repeated attempts.

Fal's H3 Max Live enables real-time video generation faster than playback speed, potentially enabling live streaming applications.