Metraipolis › archive › dc_106a5deb494b

Read-Best Is Not Steer-Best: A Probing--Steering Layer Dissociation in Omni-Modal Large Language Models — arXiv:2609.22135

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.22135
License (as archived)
CC0-1.0
Archived
2026-10-09T06:48:54.947Z
SHA-256 of the archived text
19223084771865eb40a8330ee0bbcc2458eccca101667b710ae6d915e7ed3b16
Stable URL of this page
https://metraipolis.com/documento/dc_106a5deb494b

How to cite

Read-Best Is Not Steer-Best: A Probing--Steering Layer Dissociation in Omni-Modal Large Language Models — arXiv:2609.22135 — arXiv (en), https://arxiv.org/abs/2609.22135 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_106a5deb494b (Metraipolis archive dc_106a5deb494b, sha256 19223084771865eb40a8330ee0bbcc2458eccca101667b710ae6d915e7ed3b16)
BibTeX
@misc{dc106a5deb494b,
  title = {Read-Best Is Not Steer-Best: A Probing--Steering Layer Dissociation in Omni-Modal Large Language Models — arXiv:2609.22135},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.22135},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_106a5deb494b; sha256 19223084771865eb40a8330ee0bbcc2458eccca101667b710ae6d915e7ed3b16},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.22135. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Omni-modal large language models integrate text, audio, and image signals into a shared residual stream, where concepts such as emotion can be linearly decoded and causally modified by activation steering. A common but rarely tested assumption is that the layer with the highest probing accuracy is also the best layer for steering, so injection layers are often selected by probe performance. We provide the first causal test of this assumption across three independently developed omni-modal models and find that it fails. Reading and intervention rely on different layers, a phenomenon we call the probing-steering layer dissociation. Using emotion as a controlled testbed, we measure layer-wise readability and steerability across text, audio, and image inputs. Probe-best layers vary widely across architectures, while steering-effective layers consistently fall within a narrow mid-to-late range of normalized depth. Paired random-direction controls show an approximately 26-fold causal gap, ruling out random perturbation and direction quality as explanations. Logit-lens analysis reveals a staged forward process: causal handle, probing saturation, and vocabulary commitment, and motivates a two-factor account in which steering effectiveness depends on both representational readability and downstream plasticity. These results show that probing accuracy is a poor heuristic for selecting intervention layers and suggest a cross-architecture mid-to-late selection criterion. We also identify a cross-modal emotion subspace organized by valence and arousal, with joy acting as a stable anchor across models. Code and data: this https URL .

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-22 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.22135 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-09T06:48:44.839Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-22?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).