Turning local geometric insights into practical decomposition and steering tools

Develop practical, scalable methods that leverage the geometric structure of activation spaces in transformer language models to perform activation decomposition and causal steering, converting empirical observations of locally organized, low-dimensional subspaces into operational algorithms for feature discovery and model control.

Background

A growing body of work indicates that LLM activation spaces exhibit meaningful, locally low-dimensional geometric structure rather than being well captured by a single global set of directions. However, existing activation decomposition and steering methods often rely on global directions and therefore fail to fully exploit this local geometry.

The paper frames the problem of transforming these geometric insights into usable, scalable tools for both decomposition and steering as an open challenge. The authors propose Mixtures of Factor Analyzers as one approach to address this challenge, but the general task of developing practical methodologies remains explicitly highlighted as open.

References

Yet, while recent work shows meaning- ful geometric structure in activation space, how to turn these insights into practical tools for decomposition and steering remains an open challenge.

— From Directions to Regions: Decomposing Activations in Language Models via Local Geometry  (2602.02464 - Shafran et al., 2 Feb 2026) in Section 1 (Introduction)

It is therefore unclear how to select stable feature combinations, convert them into smooth tangent directions on the activation sphere, and perform geodesic updates efficiently without introducing noise or instability.

— GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models  (2609.10658 - Ngo et al., 9 Sep 2026) in Section Limitations, bullet “Integration with other direction-discovery methods”

The behavioral gap is thus better read as a failure to act on a represented ambiguity than a failure to register it, though establishing this causally (e.g. via activation steering) remains future work.

— FinInteract: Benchmarking Clarification and Intent Integration in Ambiguous Financial Question Answering  (2609.24002 - Wang et al., 21 Sep 2026) in Appendix, Section 6.4, “Mechanistic Case Study: Where Ambiguity Lives in the Network,” paragraph “Finding 17”

However, completely disentangling this bias from all correlated factors in the representation space is a broader open problem in representation-level intervention.

— Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation  (2609.04676 - Shin et al., 4 Sep 2026) in Limitations, second limitation

Future work could ask whether the aspects can be discovered rather than imposed from rhetorical theory, and whether the geometric structure generalizes across model families and languages beyond English.

— Can Activation Steering Capture Multidimensional Authorship Style?  (2609.04792 - Tran et al., 4 Sep 2026) in Section 6, Conclusion

Our work leaves a number of questions unanswered. First, where does the control we observe in Experiment~2 come from? Causal interventions could reveal how the scores, which the model sees in the token space, influence the representations of the sentence in the activation space. Moreover, can we fully characterize the functions of the activations that have the property of being controllable at all, and explain why?

— Cross-Model In-Context Neurofeedback Reveals (Limited) Activation Control in LLMs  (2610.20179 - Baldelli et al., 14 Oct 2026) in Section 5, paragraph “Future work” (Discussion and conclusion)