arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.04465cs.CLcs.AIcs.LG
收藏

概念标记:通过概念定义学习行为嵌入

Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions

Instituto de Computación, Facultad de Ingeniería, Universidad de la República(共和国大学工程学院计算研究所)

Ignacio Sastre, Aiala Rosá

AI总结:

Concept Tokens通过概念定义学习行为嵌入,能有效控制冻结语言模型的行为,减少幻觉并提升合规性。

摘要:

我们提出了Concept Tokens,一种轻量级方法,通过在预训练的语言模型(LLM)中添加一个新的特殊标记,并仅从多个自然语言定义中学习该标记的嵌入。在此方法中,目标概念的出现被替换为新的标记。语言模型保持冻结,嵌入通过标准的语言建模目标进行优化。我们评估了Concept Tokens在三种设置中的表现。首先,我们在HotpotQA上的封闭书问答中研究了幻觉现象,发现存在方向性影响:否定幻觉标记主要通过增加弃答来减少幻觉答案,而肯定则增加幻觉并降低精度。其次,我们诱导了重述,这是一种用于第二语言教学的教育反馈策略,并观察到相同的方向性影响。此外,与提供完整的定义语料库上下文相比,概念标记在保持对其他指令的合规性(例如,要求后续问题)方面表现更好。最后,我们通过埃菲尔铁塔和虚构的“Austral Tower”进行了定性研究,以说明所学习的嵌入捕获了哪些信息以及其局限性在哪里。总体而言,Concept Tokens提供了一种从定义中学习的紧凑控制信号,可以引导冻结的LLM行为。

英文摘要:

We propose Concept Tokens, a lightweight method that adds a new special token to a pretrained LLM and learns only its embedding from multiple natural language definitions of a target concept, where occurrences of the concept are replaced by the new token. The LLM is kept frozen and the embedding is optimized with the standard language-modeling objective. We evaluate Concept Tokens in three settings. First, we study hallucinations in closed-book question answering on HotpotQA and find a directional effect: negating the hallucination token reduces hallucinated answers mainly by increasing abstentions, whereas asserting it increases hallucinations and lowers precision. Second, we induce recasting, a pedagogical feedback strategy for second language teaching, and observe the same directional effect. Moreover, compared to providing the full definitional corpus in-context, concept tokens better preserve compliance with other instructions (e.g., asking follow-up questions). Finally, we include a qualitative study with the Eiffel Tower and a fictional "Austral Tower" to illustrate what information the learned embeddings capture and where their limitations emerge. Overall, Concept Tokens provide a compact control signal learned from definitions that can steer behavior in frozen LLMs.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑