This work examines how post‑training alters a language model’s bias toward the poster’s occupation in fact‑checking tasks. Using base and post‑trained checkpoints from six model families, we evaluate fact‑checked claims with randomized poster profiles, focusing on “job‑based doubt” – the tendency to label statements false when the poster holds a lower‑status job. Findings show:
-
The effect appears only with the open T\"ulu-3 recipe (OLMo‑2 32B, T\"ulu‑3 70B); Meta, Alibaba, and Mistral recipes do not exhibit it. The shift is under one percentage point but statistically reliable, exceeds normal training drift, and replicates on independent claim sets.
-
In OLMo‑2 32B, the model does not increase attention to the job field; instead, it reinterprets it: a lower‑status job pushes the base model toward “true” and the post‑trained model toward “false”.
-
This flip is carried by a single direction in the residual stream and occurs during supervised fine‑tuning. Validation on 400 new claims with pre‑specified predictions confirms the mechanism. None of seven deployed commercial models show increased doubt toward lower‑status posters.
The study highlights that different post‑training recipes can introduce subtle yet measurable speaker bias, underscoring the need for recipe‑specific audits.
Review