New Research Proposes Test-Time Safety Alignment Method for Large Language Models
A latest arXiv paper explores using input word embeddings as control variables to achieve safety alignment of large lang…
📂 Research
Latest articles in Research
EvoSelect: A New Data-Efficient Approach to Directed Task Evolution for LLMs
A latest arXiv paper introduces the EvoSelect framework, which achieves efficient iterative evolution of large language …
HIVE Framework: Capturing Hallucinations in Diffusion LLMs from Denoising Trajectories
Researchers propose the HIVE framework, which detects hallucinations in diffusion large language models by extracting co…
Neural Cellular Automata Achieve Structural Generalization Without Handwritten Rules
A new paper on arXiv proposes a novel approach based on Neural Cellular Automata (NCA) and discrete bottlenecks that ach…
Study Reveals: Large Language Models Are Systematically Producing 'Kitsch'
A latest arXiv paper argues that while content generated by large language models often outperforms human work in evalua…
CogRAG+: Diagnosing Memory and Reasoning Deficiencies in LLMs from a Cognitive Perspective
A latest arXiv paper proposes the CogRAG+ framework, which systematically diagnoses and repairs memory and reasoning def…
New Study Reveals: Partial Evidence Actually Worsens LLM Hallucinations
A new paper introduces the phenomenon of "Anchored Confabulation": providing large language models with partial intermed…
UniMatrix: Structured Recurrent States Enable Precise Associative Retrieval
A new paper introduces the UniMatrix architecture family, which leverages structured recurrent states as a compact assoc…
AI Agents Learn to Evaluate GUI Usability, Automating Human Testing
A new study on arXiv proposes training Computer Use Agents (CUAs) to automatically evaluate graphical user interface usa…
Can General-Purpose LLMs Handle Electricity Invoice Data Extraction?
A Spanish research team leveraged two general-purpose large language models — Gemini 1.5 Pro and Mistral-small — to extr…
BioGraphletQA: Driving Complex Question-Answering Data Generation with Knowledge Graph Subgraphs
Researchers propose BioGraphletQA, a QA data generation framework anchored in small knowledge graph subgraphs (graphlets…
MATH-PT: The First Mathematical Reasoning Benchmark for Portuguese Unveiled
A research team has released the MATH-PT dataset, a mathematical reasoning evaluation benchmark specifically designed fo…