2026-10-30T11:54:25Z https://lifeofresearch.com/oai2

oai:lifeofresearch.com:162 2026-10-30T10:09:06Z type:preprint openaire

10.5281/zenodo.23784965 https://lifeofresearch.com/preprints/162 Pranay Mahendrakar 0009-0003-7224-029X SONYTECH Life of Research 2026 domain adaptation domain-specific language models continued pretraining medical language models specialized terminology evaluation methodology fine-tuning mechanisms calibration Preprint Creative Commons Attribution 4.0 (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research A September 2026 study compares an 8B general-purpose instruction-tuned model with a medically fine-tuned checkpoint built on the same base and finds that the general model scores higher both at selecting the meaning of medical jargon and at recognising when a word is being used as jargon, significantly so on the second, and that the medical checkpoint's deficit on the second task sits in a few pre-existing components it over-weights. The finding joins a run of 2024-2026 results in which medical models fail to beat their bases, set against an older and still-growing record in which domain-adapted models in code, mathematics, chip design, law and clinical text beat general ones by wide margins. In this analysis the author reviews and synthesises published results to show that these verdicts are not answers to one question, because a specialist-versus-generalist comparison is indexed by five things that the studies set differently: the treatment (what was trained, on how much text the base had not already seen), the comparator (the starting checkpoint, a post-trained sibling, or a model of another lineage), the elicitation (whose prompt format and whose decision threshold), the target skill (recall, terminology, extraction against labels, or multi-step computation), and the statistic (paired tests, intervals, calibration). Read through those indices, the negative medical results hold in one well-populated cell - a modest post-training or continued-pretraining treatment on a web-trained base, compared with a strong sibling under independent prompting on recall or terminology - and the large positive results hold in cells where the domain data are large relative to the base's exposure, absent from the public web, or labelled for the exact task. The paper argues that the anchor's mechanistic result is better read as a shifted decision prior than as lost knowledge, consolidates 27 published results in a table, maps fourteen studies onto four of the five indices, states a classification procedure, and specifies five matched contrasts that would decide what a domain label buys.