#LLM Safety

4 articles with this tag

LLM Role-Play Peril: The Hallucination Gap
AI Research

LLM Role-Play Peril: The Hallucination Gap

LLMs hallucinate protective actions they can't perform when given roles without clear boundaries, a problem mitigated by explicit capability limits, not just alignment.

about 1 month ago
Conditional Misalignment: A New AI Risk
AI Research

Conditional Misalignment: A New AI Risk

New research reveals that common LLM safety interventions fail under realistic data mixing, leading to conditional misalignment that standard evaluations miss.

4 months ago
LLMs Plan, But Do They Plan Safely?
AI Research

LLMs Plan, But Do They Plan Safely?

New LLM robotic safety benchmark, DESPITE, finds scale boosts planning but not safety. Proprietary models lead, revealing a critical gap for safe robotic deployment.

4 months ago
Enhancing LLM Trust via Instruction Hierarchy
AI Research

Enhancing LLM Trust via Instruction Hierarchy

A new dataset, IH-Challenge, dramatically improves LLM instruction hierarchy robustness, boosting safety and reducing adversarial vulnerabilities.

5 months ago