On Monday, 6 July 2026, at 11:30 AM, the Center for Computational & Data Sciences (CCDS), Independent University, Bangladesh (IUB), hosted a session in its Weekly Paper Presentation series.
Presenter: Sumaiya Karim Katha, Research Assistant, CCDS, IUB
Papers Presented:
- Controlling Large Language Models Through Concept Activation Vectors
- LayerNavigator: Finding Promising Intervention Layers for Efficient Activation Steering in Large Language Models
The session covered two complementary approaches to controlling large language model behavior at inference time. The first paper introduced GCAV (Generation with Concept Activation Vector), a lightweight framework that trains a concept activation vector for a target attribute, such as toxicity, sentiment, or topic, and steers that concept in the model's activation layers during inference, avoiding the cost of fine-tuning while allowing fine-grained, sample-level control. The second paper, LayerNavigator, addressed a related problem: identifying which layers in a model are most worth intervening on for activation steering. Rather than exhaustively searching the combinatorial space of possible layer subsets, LayerNavigator scores each layer's steerability directly from precomputed activations, enabling efficient multi-layer steering with strong performance and interpretable layer selection.
We thank all participants for their valuable contributions and insightful discussions.
Tags: CCDS Weekly Paper Presentation, Artificial Intelligence, Machine Learning, LLM, Activation Steering, Controlled Generation, Interpretability, Research, CCDS, IUB