<ul data-eligibleForWebStory="true"><li>Large Language Models (LLMs) present new opportunities in natural language processing but also pose risks like ethical concerns and bias.</li><li>Capital One's Enterprise AI team focuses on safe and responsible AI integration into products.</li><li>They introduced a paper on refining LLM input guardrails to enhance safety and efficiency.</li><li>The study at Preventing and Detecting LLM Misinformation won the Outstanding Paper Award.</li><li>LLM post-training stages aim to improve output quality and comply with safety guidelines.</li><li>Guardrails are critical for user-facing applications to prevent biased or harmful outputs.</li><li>Developing guardrails is essential due to adversarial attacks targeting LLMs.</li><li>The input moderation guardrails act as a proxy defense to filter out unsafe interactions.</li><li>Using techniques like LLM-as-a-Judge helps identify safety violations in user inputs.</li><li>Chain-of-thought prompting and fine-tuning improve LLM's reasoning and classification performance.</li><li>Experimental results show significant enhancement in LLM performance with refinement and alignment techniques.</li></ul>

Refining input guardrails for safer LLM applications | Capital One

Discover more