Steering Sweet Spots: Finding the Optimal Uncertainty Window for Activation Interventions

TL;DR: Inspired by Rahn et al.'s work on activation steering, we'll investigate whether there's a "Goldilocks zone" of uncertainty where activation steering works best—too little uncertainty means the model is already committed, too much means the representation space is too chaotic for reliable steering.

Research Question: Is there an optimal range of token-level uncertainty for effective activation steering, and can we identify and predict these "steering sweet spots" in advance?

Hypothesis: Activation steering will be most effective within an intermediate uncertainty window (neither too low nor too high), and we can predict these windows by analyzing the geometry of the model's hidden state space using the path prediction methods from Zur et al.

Experiment Plan: Use EAST (Entropic Activation Steering) from Rahn et al. to implement activation interventions across different uncertainty levels. Develop a metric to quantify the "richness" of the path space using the hidden activation predictions from Zur et al.'s methodology. Create a 2D landscape mapping steering effectiveness against (uncertainty, path richness) pairs. Test whether we can predict steering success by identifying regions of this landscape before applying interventions. Explore whether the optimal uncertainty window varies across different reasoning domains or problem types. Expected outcome: We'll identify an inverted-U relationship between uncertainty and steering effectiveness, with the peak occurring at intermediate uncertainty levels where the path space is rich but not chaotic.

References: ["Rahn, N., D'Oro, P., & Bellemare, M.G. (2024). Controlling Large Language Model Agents with Entropic Activation Steering. arXiv.org.", 'Zur, A., Geiger, A., Lubana, E., & Bigelow, E.J. (2025). Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamics.']