A preprint by independent researcher Jason Van Pham proposes translating continuous representations between frozen models to enable steering, recall, and reasoning without weight editing. The work introduces a method where concepts are compiled into residual directions in one model's space and mapped via an affine transformation into another model's hidden layers.
- In the first experiment, adding a negative gain to a concept direction in Qwen2.5-0.5B shifted generations from living-animal readings to object-language (e.g., "portable stove") for a fictional creature prompt.
- The second regime maps ordered fragments from Qwen3-Embedding-8B into Llama-3.1-8B input slots, allowing the target model to reconstruct nonce propositions and support matched inference.
- Rank 128 compression in the embedding stage preserved pairwise similarity (r=0.937) but reduced reconstruction quality, demonstrating that retrieval geometry survives compression before token identity does.
- The author explicitly limits claims, noting the method is not fine-tuning, KV-cache memory, or a proof of general semantic inversion.
The research establishes a sign-sensitive residual-state transition on specific prompt families and demonstrates that compact inversion and ordered recall are distinct operating points for cross-model information transfer.