All Articles
17680 articles total
A Comparative Study on Affective Cues in Text Embeddings Across Psychological Emotion Theories
Comparative study on how well modern text encoders capture psychological emotion theories at word and sentence levels.
From Tool Connection to Execution Control: Benchmarking Security Invariants in MCP-Style Agent Runtimes
Proposal of the Handle-Capability Protocol (HCP) to add an execution-control layer to MCP-style agent runtimes for improved security invariants.
Diff-Based Code Corruption using LLMs for Large-Scale Bugfix Benchmarking
Introduction of MegaBugFix, a large-scale benchmark of 12,629 buggy Python programs created via LLM-generated diffs for bugfixing evaluation.
AB-RAG: Adaptive Budgeted Retrieval-Augmented Generation for Reliable Question Answering
AB-RAG, a training-free framework that adaptively decides the retrieval budget for RAG systems based on confidence estimates.
Meta is adding ridiculous ‘rate limits’ and a soft paywall to its smart glasses
Meta is introducing monthly rate limits on AI features for its smart glasses, requiring a paid subscription for extended use.
Modification-Considering Value Learning for Reward Hacking Mitigation in RL
Researchers propose Modification-Considering Value Learning (MCVL) to mitigate reward hacking in reinforcement learning by filtering transitions that decrease a bootstrapped-return estimator.
RGLD: Randomized Global-Local Density Estimation for Tabular Anomaly Detection
RGLD is a new randomized global-local density estimator designed for efficient unsupervised tabular anomaly detection, outperforming deep detectors in speed and accuracy.
Evidence-Based Text-Conditioned 3D CT Synthesis for Ovarian Cancer
OvESyn is a framework for text-conditioned 3D CT synthesis for ovarian cancer, enabling the generation of synthetic medical imaging cohorts without original radiology reports.
Compositional Dynamics in Learning and Mechanics
This paper presents a compositional setting where gradient-based learning and Hamiltonian mechanics are treated as functorial semantics using an operad of smooth adaptive arrangements.
Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B
The authors introduce AgriTune-R, a reproducible framework for adapting general-purpose LLMs (specifically Qwen3-8B) to specialized agricultural applications.
Arbitrary Reduction of Validation Error for AI Decision Tests using Homomorphic AI and Repetition Codes
The paper describes HbHAI (Hash-based Homomorphic AI), which allows processing encrypted data with native AI algorithms while reducing computation time and energy footprint.
Reward-Free Code Alignment from Pretrained or Fine-Tuned LLM: Unpacking the Trade-offs for Code Generation
An empirical study on reward-free code alignment (DPO and BoNBoN) reveals that aligning pretrained models yields larger improvements than aligning already finetuned models.
BERTomelo: Your Portuguese Encoder Best Friend
BERTomelo is a new monolingual Portuguese encoder based on ModernBERT, trained on a massive corpus to outperform previous Portuguese-specific models.
Semantic-Aware, Physics-Informed, Geometry-Grounded Weather Video Synthesis
A new framework for weather video synthesis uses semantics, physics-informed dynamics, and scene geometry to generate realistic adverse weather effects for video editors.
Google copybara: moving code between repositories
Google's Copybara is a tool for moving code between repositories while maintaining history and handling transformations.
Anthropic to restoring access to Claude Fable 5 and Mythos 5 from tomorrow
Anthropic is restoring access to Claude Fable 5 and Mythos 5 models.
Claude Fable 5 export control lifted
Export controls on Claude Fable 5 have been lifted.
Exploring the Value of Diverse LLM Explanations in Introductory Programming
Research indicates that providing students with diverse LLM-generated explanations for programming exercises improves understanding and accuracy.
A Task-Driven and Quality-Assured Agent Framework for SAR Data Generation
SAGA is a new agent framework designed for reliable and quality-assured synthetic aperture radar (SAR) data generation and augmentation.
Latent Bridges for Multi-Table Question Answering
GRAB is a new pipeline that uses a graph encoder and latent bridges to improve multi-table question answering for LLMs without freezing the LLM itself.