Meta 新论文:用跨视角一致性检验 LLM 答案真伪
Meta 这篇论文思路很妙:让模型用 3 个不同身份回答同一个问题,只留下大家都一致的部分,答案不一致就弃答。不用重新训练就能提升可信度,做问答或客服系统的人可以直接试试。
Meta 发表一篇新论文,核心论点是似然性不适合作为检验 LLM 答案真伪的标准,因为在下一词预测目标下,模型学到的是文本中的常见内容而非事实——错误说法在训练数据中重复越多,概率反而越高,罕见但真实的事件则被低估。
论文借鉴物理学和哲学的思路,提出把「在视角变换后仍然成立的事实」视为真实:让模型从多个视角作答,例如以 3 位医学专家的身份分别回答,保留所有答案中共同的部分,若无共同内容则选择弃答。
这篇论文没有跑实验,论证方式是算例展示:在一个二分类例子里,损失为 0.67 对应 100% 的准确率,而仅高出 0.04 的 0.71 却对应 0%。论文给出的实操建议是,在相信一个高置信度答案前,先检验它在变换 persona、措辞框架或语言后是否依然成立。
Such a simple yet powerful idea in new Meta paper.
Says that likelihood is the wrong test of truth for LLMs, and that answers which stay the same across contexts deserve more trust.
that LLMs are trained to predict the most likely next token, so they learn what is common in text, not what is true. A falsehood repeated often in training data becomes probable, while rare true events get discounted.
Drawing on physics and philosophy, the paper treats a fact as real if it survives a change of viewpoint. The model answers from several viewpoints, such as 3 medical specialists, keeps what every answer shares, and abstains if nothing is shared.
The paper runs no experiments, so its proof is worked examples. On a single binary example, a loss of 0.67 means 100% accuracy, while a nearly equal loss of 0.71 means 0%.
Before trusting a confident answer, check whether it survives changes in persona, framing, or language.
本条情报归属 AITOP「论文」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Meta、LLM、提示词工程、AI安全,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「rohanpaul_ai」(@rohanpaul_ai),经 AITOP 多维评分管线结构化提炼。