Biomedical questions: when retrained on its own answers, a model loses accuracy with each successive generation

On PubMedQA, Qwen2.5 models with 0.5 and 3 billion parameters, trained on their own outputs, perform less well at recognising diseases and molecules, and rely less on the context provided. The larger model performs even worse.

What happens when a medical model is retrained, generation after generation, on the responses of its predecessor? A study on PubMedQA compares two approaches using Qwen2.5 with 0.5 and 3 billion parameters: a synthetic loop, where generation G(k+1) learns from the responses of G(k), and a control group that is consistently retrained on the original human data. Four generations (G0 to G3) are tracked using two random seeds, on a fixed set of 1,000 expert-annotated examples.

In all configurations, the synthetic loop performs worse than the control group in terms of disease and chemical recognition, the proportion of context-supported responses, and lexical and semantic proximity to the reference responses. Responses become longer, without an increase in repetitions — the decoding process prevented the repetition of the same trigram anyway.

Counterintuitively, the gap is more pronounced for the 3-billion-parameter model. The authors remain cautious: these results do not measure a rate of clinical hallucination and do not prove a general collapse of the models.