PulseExploreJournal ClubResearchersJournals
Instagram
HomeJournal ClubExplore
Synapse
⌘+K
Synapse
October 13, 2025Open Access

Steering Language Models in Multi-Token Generation: A Case Study on Tense and Aspect

View Full Paper
Ask AI
Bookmark
Share

Authors

AKAlina KleringsJBJannik BrinkmannDRDaniel Ruffinelli

Discussion

Loading...

Member takes

Overview

Case study demonstrates causal control over tense and aspect in multi-token generation, suggesting sensitivity to various factors.

Key Points

  • Effective steering of tense and aspect in language models reduces topic shift and degeneration in generated text.
  • Employing linear discriminant analysis reveals distinct directions in encoding grammatical features within the model.
  • Manual tuning or automated optimization is necessary for effective control over complex grammatical features.
  • Parameters such as steering strength, location, and duration are crucial for managing multi-token generation outcomes.

Cite This Study

Klerings et al. (2025) studied this question.

synapsesocial.com/papers/68ed1896f29694dd1da78af8https://doi.org/10.48550/arxiv.2509.12065
View Full Paper
Ask AI
Bookmark
Share