Steering off Course: Reliability Challenges in Steering Language Models

TL;DR

The paper conducts a large-scale, cross-model evaluation of three steering methods for language models—DoLa (logit-lens–based), function vectors, and task vectors—across 36 decoder-only transformers from 14 families (1.5B–70B). It reveals substantial variability and widespread brittleness: many models show little or negative improvements under steering, and the underlying assumptions of these methods often fail to generalize. Through targeted analyses, the authors identify issues such as incorrect reliance on layer-contrasting signals, variable sensitivity to hyperparameters, and non-uniform effects of model pretraining and architecture. The work argues for rigorous, multi-model, multi-task evaluation paradigms to assess the reliability of surgical steering techniques before practical deployment. Overall, steering methods exhibit limited generalizability and reliability as scalable tools for modifying LM behavior.

Abstract

Steering methods for language models (LMs) have gained traction as lightweight alternatives to fine-tuning, enabling targeted modifications to model activations. However, prior studies primarily report results on a few models, leaving critical gaps in understanding the robustness of these methods. In this work, we systematically examine three prominent steering methods -- DoLa, function vectors, and task vectors. In contrast to the original studies, which evaluated a handful of models, we test up to 36 models belonging to 14 families with sizes ranging from 1.5B to 70B parameters. Our experiments reveal substantial variability in the effectiveness of the steering approaches, with a large number of models showing no improvement and at times degradation in steering performance. Our analysis demonstrate fundamental flaws in the assumptions underlying these methods, challenging their reliability as scalable steering solutions.

Figures (21)

• Figure 1: We study the generalization of various LM steering methods to previously unstudied models and find high variance in steering performance. (Top plot) Building on logit lens §\ref{['sec:VD']}, DoLa contrasts token probabilities across layers to discover factual answers. We show that models do not have similar patterns, limiting the effectiveness of this method. (Middle and bottom plots) Function vectors and task vectors are two methods for steering based on activation patching §\ref{['sec:activation']}. We show that activation patching results in highly variable performance across many model families and sizes.
• Figure 2: Projected token probabilities from hidden states at each layer of 4 selected LMs on the TruthfulQA dataset (the remaining results can be found in Appendix \ref{['asec:logit_lens_analysis']}). The correct and incorrect token probabilities begin spiking at the same layer, which suggests that a contrast with early layers would be relatively uninformative.
• Figure 3: Performance recovery with different activation patching methods across tasks and models. There is large variance in performance across tasks, models, and tools. Tasks: a) Antonym, b) Present-Past, c) Country-Capital, d) [lang] to eng, and e) eng to [lang].
• Figure 4: A subspace in the hyperparameter search across ($\ell, \lambda, \mathcal{A}_n$) for Mistral-v0.3 7B on the Country-Capital task. The FV does not become effective until 128 (10%) heads, and it continues to grow in performance until 512 (50%) heads. This provides some evidence against localization for this model and task, as many heads are required before performance emerges.
• Figure 5: (Left) Some tasks work well with few heads, while translating from English requires more heads to be effective. (Right) Post-trained models outperform their base counterparts on average, especially when more heads are used to construct the function vector. However, there is significant variance when averaging across models and tasks.