Qwen2.5-7B builds its own synthetic steering vectors, injected into L0-L19, then X-rays all 28 layers

Nearby_Indication474 · reddit · 2026-09-27

AkbasCore KIZILELMA Live X-Ray is an open-source/Colab demo that performs activation steering live inside Qwen2.5-7B-Instruct during inference and visualizes hidden-state changes across all 28 layers, in a pipeline of internal extraction → synthetic vector construction → injection → engine off → 28-layer X-ray:

It captures the model's own hidden states and extracts the requested representational contrast.
It automatically builds 20 layer-local (L0–L19) × 3,584-dim synthetic steering vectors.
These are injected into L0–L19 with measured norm-relative doses; steering stops completely at the L19/L20 boundary.
It then measures Vanilla→SEASC hidden-state displacement across the full model, including L20–L27 with zero new injection, revealing how the effect propagates.

No external steering vectors, LoRA, fine-tuning, or weight modification. An interactive demo and a one-click Colab A100 experiment are provided (the full run takes seconds once ready), with releases on GitHub (titan-cognitive-core) and Zenodo (DOI 10.5281/zenodo.22979417). The author also suggests handing the code link to several AI systems to independently verify it.

Original post →

Qwen2.5-7B builds its own synthetic steering vectors, injected into L0-L19, then X-rays all 28 layers

More from Research