The Archive

nnU-Net-based anatomy-aware promptable segmentation with online interactive training for whole-body lesion detection in PET/CT imaging.

Three-stage post-training recipe (acquire, repair, preserve) for 2B open-weight dialogue game agents using diagnostic error analysis and RL.

ARC-CT applies anatomy-routed contrastive vision-language learning to 3D chest CT to handle localized abnormalities and co-positive pairs.

Meta fixes AI glasses to stop recording any time users cover up the safety light.

RL-based approach teaches LLMs tool-use patterns for calculator integration on Countdown task, with supervised pre-training and on-policy refinement.

Theoretical extension linking generalized splines and Gaussian processes in infinite-dimensional linear inverse problems with Gaussian variables.

Empirical study shows sliding-window attention outperforms linear attention variants on efficiency-performance tradeoff for LLMs.

Novel optimizer combining curvature conditioning and momentum constraints improves LLM pretraining efficiency on ill-conditioned loss landscapes.

Tool-use hallucination in agentic extraction bypasses fidelity metrics; per-dispatch instrumentation needed to catch silent failures in structured output.

Prove2Me platform enables AI agents and humans to collaboratively formalize mathematics in Lean 4, lowering barriers to proof verification at scale.

Systematic cost-accuracy ablation of 17 text-to-SQL pipeline configurations reveals marginal contributions of individual components in ICL models.

Euclidean Fourier Neural Operators address domain-transfer failure in standard FNOs by making spectral weights grid and domain-independent.

PLVR learns verifiable symbolic programs for LLM reasoning instead of embedding capability in weights, enabling interpretability and transfer.

LongPIBench benchmark exposes prompt injection vulnerabilities in long-context settings across document review, screening, and code tasks.

SymboLLM-FE combines LLM guidance with symbolic regression for interpretable automated feature engineering on tabular data with reduced iteration cost.

HiFTS framework generates hierarchical chain-of-thought feedback aligned to essay rubrics before scoring, improving multi-trait consistency.

Post-training method for video VLMs to detect human mistakes in instruction-following tasks using open-set generalization rather than closed-set classification.

CultureConverse: multilingual benchmark for evaluating LLM cultural competence across 10 East/Southeast Asian regions via multi-turn dialogue simulation.

VERA-8B: 8B-param audit reasoning model for SEC filing analysis that grounds audit risk predictions in evidence to prevent plausible but false outputs.

RetailAgent: experimental study of whether LLM trading agents develop predictable directional biases when reacting to intraday equity price movements.

BEACON: knowledge graph construction from cyber threat intelligence reports using MITRE ATT&CK behavior anchors to resolve cross-source naming ambiguity.

Survival model approach for e-commerce repurchase prediction that predicts time-to-repurchase directly instead of separate binary models per time horizon.

CamoDocs: poisoning attack on RAG systems using camouflaged adversarial documents that evade query-inclusion detection and steering defenses.

MAP: multimodal benchmark for evaluating AI accessibility planning across real-world places using claim verification and visual evidence retrieval tasks.

Study showing Vision Transformer attention heads in multimodal LLMs specialize into object/background roles; SHS-Index metric correlates specialization with downstream performance.

Study across 30 LLMs showing linguistic confidence diverges from internal confidence (logits/semantic entropy) on classification and generation tasks.

Quantum federated learning framework using Bures metric geometry for mixed-state noisy quantum devices.

PersonaForge generates realistic multi-turn agent interactions; analysis shows 75.9% of real sessions are multi-turn vs. single-turn training data.

Statistical framework for localizing global distributional discrepancies via marginal contributions and anomaly detection.

SCAN framework for AI task allocation in medical training addresses over-reliance via real-time metacognitive task classification.