Grounding
Grounding refers to the process of linking symbols, language, or abstract representations to real-world perceptions, experiences, or concrete data. The concept is widely applied in cognitive science, linguistics, and artificial intelligence. In cognitive science, it denotes the connection between meaning and embodied experience. In AI, it specifically refers to anchoring model outputs to external knowledge sources or factual data in order to reduce hallucinations and improve the reliability and verifiability of generated responses.
Timeline (last 90 days)
Grounding refers to anchoring a language model's answers in verifiable external sources of fact, with common methods including retrieval-augmented generation (RAG), quoting passages from the source text, and tool queries for real-time data.
Sidekick's Movie Buddy uses a grounding approach, injecting subtitles as the source of fact, thereby reducing fabrication risk in areas such as film details where models tend to hallucinate.
Faithfulness, sometimes also called groundedness, measures whether every claim in a generated answer can be found in the retrieved context, focusing on whether the model fabricated content outside the context.
Grounding refers to constraining the output of large language models by providing external authoritative knowledge to prevent them from generating hallucinations.
Grounding prompts that instruct the model to answer only based on provided documents and cite the relevant passages can greatly reduce the probability of fabrication, but cannot eliminate it entirely.
All Facts (6)
Grounding（接地）是对抗大语言模型幻觉的核心工程手段，通过为模型关键输出绑定可验证的真实数据源来限制幻觉影响
85%UnverifiedGrounding refers to anchoring a language model's answers in verifiable external sources of fact, with common methods including retrieval-augmented generation (RAG), quoting passages from the source text, and tool queries for real-time data.
50%UnverifiedSidekick's Movie Buddy uses a grounding approach, injecting subtitles as the source of fact, thereby reducing fabrication risk in areas such as film details where models tend to hallucinate.
50%UnverifiedFaithfulness, sometimes also called groundedness, measures whether every claim in a generated answer can be found in the retrieved context, focusing on whether the model fabricated content outside the context.
50%UnverifiedGrounding refers to constraining the output of large language models by providing external authoritative knowledge to prevent them from generating hallucinations.
50%UnverifiedGrounding prompts that instruct the model to answer only based on provided documents and cite the relevant passages can greatly reduce the probability of fabrication, but cannot eliminate it entirely.
50%