arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2502.03490cs.AIcs.LG
收藏

通过信息容量缩放检验两跳推理

Examining Two Hop Reasoning Through Information Content Scaling

EleutherAI

David Johnston, Nora Belrose

AI总结:

本研究通过容量缩放探究Transformer两跳推理能力不稳定的原因,发现潜在两跳问答需模型对每个事实学习两次,还可通过数据集参数让小模型陷入独立记忆答案的状态,验证了容量缩放测量可补充可解释性方法。

摘要:

先前研究发现,Transformer在学习回答潜在两跳问题(形式如“谁是Bob母亲的老板?”的问题)方面能力不稳定。受Transformer用于简单事实记忆的知识容量相关前期研究启发,我们通过研究Transformer学习两跳问答(two-hop QA)数据集的能力如何随模型规模缩放,探究该现象的成因。我们发现,容量缩放与泛化表现均支持以下假设:潜在两跳QA需要Transformer对每个事实学习两次,而采用思维链(chain of thought)的两跳QA则无需如此。我们还表明,通过设置合适的数据集参数,可将极小模型“困在”一个独立记忆两跳问题答案的状态中——即便它们若能学会通过函数组合来答题会表现得更好。我们的研究结果表明,容量缩放测量可补充现有可解释性方法,不过将其用于该目的仍存在一些挑战。

英文摘要:

Prior work has found that transformers have an inconsistent ability to learn to answer latent two-hop questions -- questions of the form "Who is Bob's mother's boss?" We study why this is the case by examining how transformers' capacity to learn datasets of two-hop questions and answers (two-hop QA) scales with their size, motivated by prior work on transformer knowledge capacity for simple factual memorization. We find that capacity scaling and generalization both support the hypothesis that latent two-hop QA requires transformers to learn each fact twice, while two-hop QA with chain of thought does not. We also show that with appropriate dataset parameters, it is possible to "trap" very small models in a regime where they memorize answers to two-hop questions independently, even though they would perform better if they could learn to answer them with function composition. Our findings show that measurement of capacity scaling can complement existing interpretability methods, though there are challenges in using it for this purpose.

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。

↑