LLM DNA

The model’s text responses are embedded using an external sentence-embedding model. To obtain a compact representation while preserving the model’s functional space, the method uses random linear projection (RLP) as the core mechanism. In the extraction step, a fixed set of prompts is fed into the model ; the resulting response embeddings are concatenated into a high-dimensional vector , and the final DNA is computed as . Here, is a Gaussian random matrix designed according to the Johnson–Lindenstrauss lemma, and is the set of embedding vectors capturing the semantic content of each response. The mapping is claimed to satisfy properties analogous to inheritance (small model changes do not cause abrupt DNA changes) and genetic determinism (similar DNA implies similar functionality), and it can be applied even to closed, API-only models where weights or internals are inaccessible.