研究人员正在探索新的方法来对抗大型语言模型 (LLM) 的幻觉并提高其事实性。语义熵分析答案的变化以检测捏造，而语言校准则训练模型以一种有助于读者预测的方式表达置信度。一致性事实性将正确性视为一个不确定性量化问题，将答案分解为子声明并过滤掉置信度低的声明。一致性语言建模将一致性预测应用于生成模型，旨在保证可接受的答案并标记可能产生幻觉的短语。 AI
影响 这些方法为提高 LLM 的可靠性提供了潜在的进展，旨在减少捏造并提高用户对 AI 生成内容的信任度。
排序理由 该集群描述了多篇学术论文，提出了检测和减轻 LLM 幻觉的新方法。
在 Mastodon — fosstodon.org 阅读 →
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
Semantic Entropy (Nature 2024) detects LLM confabulations by clustering sampled answers by meaning and computing entropy over the cluster distribution. "Paris" and "It's Paris" cluster together, so paraphrase noise doesn't inflate the signal. Cost: it only catches hallucinations …
Linguistic Calibration trains Llama 2 to emit confidence phrases that let a downstream reader make calibrated forecasts on related questions. The key move is defining calibration through reader utility instead of self-reported probability. Hedged text that doesn't help the reader…
Conformal Factuality casts LM correctness as uncertainty quantification. Decompose the answer into sub-claims, score each, drop the low-confidence ones until the retained set is ~1-α factual. The sub-claim decomposition is doing most of the work, and the conformal machinery rides…
Conformal Language Modeling (CLM) adapts conformal prediction to generative LMs: sample candidates, stop when a calibrated rule fires, return a set guaranteed to contain an acceptable answer. The more interesting half is the component-level filter — per-phrase coverage, not just …