[
    {
        "id": "osp-16775",
        "type": "article-journal",
        "title": "Learning Cross-Model Activation Alignments with Explicit Many-to-Many Layer Maps",
        "author": [
            {
                "family": "Sudakov",
                "given": "Alina"
            },
            {
                "family": "Bar-Shalom",
                "given": "Guy"
            },
            {
                "family": "Frasca",
                "given": "Fabrizio"
            },
            {
                "family": "Maron",
                "given": "Haggai"
            }
        ],
        "URL": "https://omanscience.com/en/articles/learning-cross-model-activation-alignments-with-explicit-many-to-many-layer-maps",
        "language": "en",
        "issued": {
            "date-parts": [
                [
                    2026
                ]
            ]
        },
        "abstract": "LLMs are released at a rapid pace, raising a natural question: how do two independently trained models relate, both in which layers correspond and in how features transform between them? We study this by learning an activation alignment, a map from a source model's layerwise activations to a target's. Our method, MATCHA, factors this map into a layer map, whose output is an explicit target-by-source matrix that can be extracted and inspected, and a layer-shared feature map between hidden spaces. Most of prior work fixes the layer correspondence in advance, pairing layers at roughly the same relative depth; in contrast, we learn both factors jointly from prompts. Across 42 pairs of seven models spanning three different families, MATCHA reconstructs the target's activations more faithfully and improves retrieval-based metrics substantially, w.r.t. previous approaches. The recovered maps are broadly monotone in depth but, in contrast with most previous approaches, are consistently many-to-many: each target layer draws on a band of source layers. Our alignments also enable transfer of activation-space interventions, allowing steering vectors and probes developed for one model to transfer to another."
    }
]