The Archive

Study identifies localization gaps in VLMs used as condition encoders for diffusion image editing in multi-entity scenes.

Multi-agent system automates H. pylori case finding from gastric biopsy reports with evidence linking and contextual assertion/negation handling.

TILDE method enables concept unlearning in text-to-image diffusion models while preserving benign generation quality and diversity.

Experimental design framework quantifies variability and factors in LLM coding agents' autonomous model discovery via stochastic evaluation.

sqlite-utils 4.0 ships with database schema migration support.

RuBench 1.0 repository-level coding benchmark with 25 natively-authored Russian task specifications across Python, PHP, TypeScript, JavaScript projects.

ExplAIner declarative query language enables uniform specification and analysis of explainability notions for Boolean classification model interpretability.

SCENT multimodal framework bridges vision and olfaction via language guidance from VLMs for scene-smell alignment beyond visible cues.

Perspective article defines world models as environment simulators, discusses key technical aspects, and charts staged research roadmap across AI subfields.

RL method for robotic throwing in cluttered environments using potential field representations to balance basket attraction and obstacle avoidance.

Theoretical analysis proving neural networks exponentially outperform NTK on compositional tasks, quantifying Fourier vs. architectural complexity trade-offs.

ActionCache accelerates Vision-Language-Action models for real-time robotic deployment by caching and refining intermediate denoising steps without retraining.

Agentic systems turn model reasoning into action through multi-step workflows that combine inference, tool use, code execution, retrieval, orchestration, and... Agentic systems turn model reasoning into action through multi-step workflows that combine inference, tool use, code…

Domain-adapted Sentence Transformers automate cloud security compliance mapping using 13k synthetic training pairs from five European standards.

TopoBrick uses agentic topology sampling on building knowledge graphs for zero-shot IoT forecasting without training or fixed covariates.

Physics-Informed Neural Networks learn finite-dimensional embeddings of PDE solution families via multihead architecture with orthogonalization penalties.

Framework for responsible personalisation in human-robot interaction examining ethical risks from embodied agents across lifecycle and interaction contexts.

Code agent framework for automated software verification outperforms fixed proof strategies, proving larger fraction of Coq theorems than prior LLM approaches.

Dictionary learning interpretability module decomposes end-to-end autonomous driving models into causal semantic concepts for safety analysis.

First large-scale evaluation of LLM uncertainty estimation across 22 languages and model sizes, comparing nine methods on multilingual MCQA.

DT-Guard uses reasoning-active training to build efficient LLM safety guardrails that detect concealed intent without excessive latency.

Dithered Gaussian mechanism improves differential privacy efficiency by discretizing outputs instead of noise, avoiding floating-point vulnerabilities.

Token-based dual-view fusion framework for breast cancer classification from mammography using prompt-based vision model adaptation.

UI2App benchmark evaluates LLM-generated web applications on interaction fidelity, not just visual rendering, using image-driven inputs.

Quantum-inspired hybrid system for automated music harmony generation combines candidate exploration with rule-based optimization.

Quantitative Gaussian-process convergence theory for infinite-width neural networks provides finite-width error bounds architecture-agnostic.

Cross-lingual transfer from Sinhala improves Dhivehi speech recognition using continual pre-training for low-resource language.

Kernel-based operator learning derives budget allocation conditions balancing training pairs, input observations, and output resolution.

SkillReranker improves agent skill selection by decomposing tasks and reranking candidates, addressing semantic ambiguity and skill applicability.