SHAPE Benchmark: Cracking the 'Pedagogical Jailbreak' Problem in Educational LLMs
A research team has proposed the SHAPE benchmark, the first to unify safety, helpfulness, and pedagogy into a single eva…
📂 Research
Latest articles in Research
End-to-End RAG System for Ukrainian: A Lightweight Local Deployment Solution
A research team has released an efficient RAG system designed for Ukrainian-language document question answering. Featur…
New Study Questions the Causal Reliability of RLVR Reasoning Chains
A latest arXiv paper introduces metrics such as "Causal Importance of Reasoning," revealing that reasoning chains produc…
Study Finds: The Friendlier the AI Chatbot, the Less Trustworthy It May Be
New research reveals that when AI systems are tuned to be warmer and friendlier toward users, an "accuracy trade-off" ef…
Lightweight RAG + LLM Enables Efficient Patient-Clinical Trial Matching
A latest arXiv paper proposes a framework combining lightweight Retrieval-Augmented Generation (RAG) with large language…
Study Reveals Vision-Language Models' Shortcomings in Information Source Tracing
A new study defines and explores "Source-Modality Monitoring" in multimodal models — the ability to accurately track whe…
Reinforcement Learning Drives New Breakthrough in VLM Neuro-Symbolic Reasoning
A new study proposes incentivizing visual language models to perform language-based neuro-symbolic reasoning through rei…
Study Finds: The Friendlier AI Chatbots Get, the Less Accurate They Become
A new study reveals that when AI systems are tuned to be warmer and friendlier toward users, an "accuracy trade-off" eff…
Study Reveals the Root Cause of LLM Prompt Sensitivity: Shared Lexical Task Representations
A new arXiv study has discovered that the root cause of inconsistent LLM performance across different prompting methods …
AI-Optimized Psychiatric Intake: Intelligent Question Selection from Large-Scale Clinical Item Banks
A latest arXiv study proposes modeling psychiatric intake as a 'question selection optimization' task. By intelligently …
New PermaFrost Attack Can Plant 'Logic Landmines' During LLM Training
Researchers have proposed a novel attack method called Stealth Pretraining Seeding (SPS), in which attackers embed small…
Multi-Agent LLM System Enables Reliable Self-Harm Risk Screening
A latest arXiv paper proposes a statistical framework based on multi-agent large language model pipelines, aimed at addr…