# A Training-Free Socket for DNA in Frozen Language Models

DevFeed: [A Training-Free Socket for DNA in Frozen Language Models](<https://devfeed.tech/articles/a-training-free-socket-for-dna-in-frozen-language-models-118318.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.05401>)

Author: Arash Fatehi

Published: 2026-11-05T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [LLM Techniques](<https://devfeed.tech/topics/llm-techniques.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [linear-regression](<https://devfeed.tech/topics/linear-regression.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [dna](<https://devfeed.tech/tags/dna.md>), [free](<https://devfeed.tech/tags/free.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [models](<https://devfeed.tech/tags/models.md>), [q-bio-gn](<https://devfeed.tech/tags/q-bio-gn.md>), [socket](<https://devfeed.tech/tags/socket.md>), [training](<https://devfeed.tech/tags/training.md>)

## AI overview

The socket transfers gene-level information from frozen Evo 2 DNA features into a frozen language model using one closed-form ridge regression, with no gradients through the host. On GeneHop, real DNA outperformed shuffled DNA across four hosts by 16.7-26.1 points; all results survived Holm correction, and null controls stayed within two points of sham. A text-only linear map let one plug serve other hosts at a test cost of at most 0.9 points. Nearest-neighbour retrieval captured 72-84% of the effect, while the plug scored 3.0-5.2 points higher on each host and retained most of its effect without training homologs. These results come from a single held-out test run under a frozen analysis plan.

## Source excerpt

arXiv:2611.05401v1 Announce Type: cross Abstract: Connecting a DNA foundation model to a large language model (LLM) usually requires training a connector, and often the LLM too. We show that a closed-form map is enough for gene-level questions. Our method, the socket, fits one ridge regression (the plug) from frozen Evo 2 features to the host LLM's own representation of the gene's symbol, name, and description at an early layer. The plug's output replaces the hidden state at placeholder positions, and no gradient passes through the host. We test it on GeneHop, a five-way multiple-choice benchmark split by homology. In a single run on the held-out test set, under a frozen analysis plan, real DNA beats shuffled DNA on all four hosts: by +22.2 points [19.8, 25.0] on Qwen3.5-4B, +26.1 [24.3, 27.9] on Qwen3.6-27B, +16.7 [14.2, 19.7] on Mistral-7B, and +18.4 [16.2, 21.0] on OLMo-2-7B. All four survive Holm correction, every null control stays within two points of the sham level, and the development results agree within 2.3 points. A layer rule that uses no task items was selected on three hosts; it predicted the working layer of the fourth before any DNA-plug scoring on it. A text-only linear map lets one plug feed any other host, at a cost of at most 0.9 points on test. The socket conveys gene identity: nearest-neighbour retrieval recovers 72-84% of the effect, but the plug beats it on every host by 3.0-5.2 points and keeps most of its effect without training homologs.