Language models learn better when training targets match bottlenecks
What Should Data Teach? Moving Bottlenecks Across Circuit, Store, and Use
Machine Learning
Summary
Training language models to learn facts depends on what kind of problem is currently hardest for the model to solve. The authors identify three main bottlenecks: figuring out the right computation, making the right information available, and choosing between different ways to use that information. They show that giving training examples that focus on whichever bottleneck is limiting the model’s learning helps it improve more efficiently. Their experiments demonstrate that changing how models get supervised during training, instead of just ranking harder examples, leads to better fact learning over time.
What this means in practice
- •For language model trainers: Adjust training signals to focus on current bottlenecks for more efficient fact learning in language models.
- •For ai system designers: Improve long-context answering by modifying supervision signals to better guide memory use and route decisions.
Authors
Yixiao Chen, Ke Cheng, Jiangtao Guan, Shuo Huang, Yue Liu, Jun Zhang, Yuhong Liu, Jie Jiang
Abstract
What should data teach a language model at a particular point in training? A circuit view reveals three distinct bottlenecks: forming a computation, making its required content available, and selecting among available routes. A shared diagnosis-to-data principle connects them: localize the missing operation, preserve its causal relation, vary shortcut-bearing context, and re-audit the residual. Formation-sensitive selection and prerequisite ordering accelerate a binding-matching-transport path; a brief early prefix from the same training multiset retains a validation advantage through 100B tokens. Availability counterfactuals then distinguish writing content from invoking available memory, while paired supervision and context-opportunity ranking improve matched route decisions and long-context answer likelihood. A continuous 350M-model experiment connects the three interventions on the same facts: early circuit training improves subsequent learning, and the complete sequence outperforms stage-replacement controls on facts withheld from Use teaching. Independent query surfaces and opposed-source decisions expose conditional arbitration as the remaining frontier. Together, these results show why a change in the limiting operation calls for a change in supervision, not merely a new ranking of difficult examples.