Fragility of Instruction-Tuned LLMs: One Token Collapse
Discover how instruction-tuned LLMs face major performance drops from minor constraints, revealing critical fragility in AI response quality.
LogicEval: Framework to Fix Logical Software Vulnerabilities
LogicEval evaluates automated repair methods for logical software vulnerabilities using the LogicDS dataset to improve real-world software security.
ROSE: Intent-Centered Metric for Accurate NL2SQL Evaluation
Discover ROSE, an intent-focused metric improving NL2SQL evaluation by prioritizing semantic accuracy over syntax for reliable SQL query assessments.
Parallax: Securing Autonomous AI Agents from Risks
Discover how Parallax protects AI agents by separating thinking from acting, blocking 99% of attacks and ensuring secure autonomous execution.
Hallucination in Video LLMs: Causes, Types & Solutions
Explore the causes, types, and mitigation strategies of hallucinations in Video Large Language Models for more reliable video-language systems.