跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.21921 · v1

Latent Fact Composition under Next-Token Training

首次提交 2026-11-17 05:33
版本更新 2026-11-17 05:33
雷达观测 2026-11-18 19:43
arXiv ↗PDF ↗

摘要

Controlled studies of latent two-hop reasoning train transformers to predict only the answer, whereas language models are trained to predict every token. We ask whether a model trained on every token learns to answer two-hop questions whose intermediate entity appears in no two-hop training example. We build a synthetic corpus with such questions. On this corpus every term of the loss can be computed exactly. A looped transformer, a standard transformer and DiscoLoop answer almost none of these questions. We therefore introduce the Baton family of looped transformers. They pass the intermediate entity to the second loop as a token embedding. When trained to predict only the answer, two of its three variants answer these questions almost perfectly. We then relate the success and failure of each model to three design conditions. Experiments that change one condition at a time agree with this hypothesis. When Baton is trained on every token, however, its accuracy on these questions rises and then declines. We further split the gradient of the next-token loss into three terms. Removing the two terms that depend on the particular training set raises the peak accuracy but does not remove the decline. The decline arises in the second hop while the model still recalls both facts.

作者备注:21 pages, 5 figures, 23 tables

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-17 05:33 · 最新观测版本

Latent Fact Composition under Next-Token Training

Controlled studies of latent two-hop reasoning train transformers to predict only the answer, whereas language models are trained to predict every token. We ask whether a model trained on every token learns to answer two-hop questions whose intermediate entity appears in no two-hop training example. We build a synthetic corpus with such questions. On this corpus every term of the loss can be computed exactly. A looped transformer, a standard transformer and DiscoLoop answer almost none of these questions. We therefore introduce the Baton family of looped transformers. They pass the intermediate entity to the second loop as a token embedding. When trained to predict only the answer, two of its three variants answer these questions almost perfectly. We then relate the success and failure of each model to three design conditions. Experiments that change one condition at a time agree with this hypothesis. When Baton is trained on every token, however, its accuracy on these questions rises and then declines. We further split the gradient of the next-token loss into three terms. Removing the two terms that depend on the particular training set raises the peak accuracy but does not remove the decline. The decline arises in the second hop while the model still recalls both facts.

查看此版本解读 →查看此版本原文 ↗