Steering Language Models in Multi-Token Generation: A Case Study on Tense and Aspect
This work addresses the problem of understanding and controlling syntactic knowledge in language models for researchers, but it is incremental as it builds on prior studies of grammatical contrasts.
The study investigated how large language models internally encode and can be controlled for verb tense and aspect, identifying distinct directions in residual space and demonstrating causal control across generation tasks, while finding that steering parameters like strength, location, and duration are crucial to minimize side effects such as topic shift and degeneration.
Large language models (LLMs) are able to generate grammatically well-formed text, but how do they encode their syntactic knowledge internally? While prior work has focused largely on binary grammatical contrasts, in this work, we study the representation and control of two multidimensional hierarchical grammar phenomena - verb tense and aspect - and for each, identify distinct, orthogonal directions in residual space using linear discriminant analysis. Next, we demonstrate causal control over both grammatical features through concept steering across three generation tasks. Then, we use these identified features in a case study to investigate factors influencing effective steering in multi-token generation. We find that steering strength, location, and duration are crucial parameters for reducing undesirable side effects such as topic shift and degeneration. Our findings suggest that models encode tense and aspect in structurally organized, human-like ways, but effective control of such features during generation is sensitive to multiple factors and requires manual tuning or automated optimization.