In https://arxiv.org/pdf/2501.09274 a standard LLM (Llama-3.1-8B-Instruct) without fine tuning is used to successfully predict protein sequences fitter for certain purposes (not specified specifically!) by guiding mutations in an evolutionary optimisation framework.
Telling title of the paper is "LARGE LANGUAGE MODEL IS SECRETLY A PROTEIN SEQUENCE OPTIMIZER"
Are there universal patterns in statistical distributions of sequences that have a generating mechanism and therefore grammar like patterns?
I.e. something learnt about grammars of languages that can be transferred?
And if so, are biological mutations maybe not random either, because cells may beat random guesses in a similar fashion?
Reminds me of methods for causal deconvolution based on algorithmic probability.