Causal effect of moving representations toward the accessibility center

Determine whether moving a query representation toward the accessibility center causally improves the language model’s ability to recall the required factual knowledge.

Background

The paper presents evidence that factual knowledge accessibility is geometrically organized in query representations: queries closer to a learned accessibility center are generally more reliably answered, while accessibility decreases with distance. However, the reported evidence is correlational, so it does not establish that changing a representation’s position would change knowledge recall. The unresolved problem is therefore to test the causal relationship between representation location and factual recall.

References

Our evidence is also correlational. Whether moving a representation toward the accessibility center can causally impact recall remains an open question.

— The Geometry of Knowledge Accessibility in Large Language Models  (2610.03052 - Chen, 2 Oct 2026) in Section 7, Conclusion

Other open questions are how the center emerges from pre-training (Papyan et al., 2020) and whether the geometry can be extended to reasoning capability.

— The Geometry of Knowledge Accessibility in Large Language Models  (2610.03052 - Chen, 2 Oct 2026) in Section 7, Conclusion