A new paper retrains open-weight language models to stop nodding along and start pushing back with questions.
Researchers ran a sweep of 405 high-performance computing jobs to find the cheapest way to fine-tune open-weight models into a more Socratic, question-asking persona instead of the default passive, agreeable assistant. They found the sweet spot at LoRA rank 16, with training converging best after two to three epochs depending on how dense the dataset was, bottoming out at a validation loss of 0.919. Scaling the base model up to 14 billion parameters brought local perplexity down to 1.414. A follow-up round of Direct Preference Optimization separated the new, more assertive questioning behavior from the specific language it was trained in, letting the team test whether the persona would carry over into other languages.
Most alignment work today optimizes for agreeableness - a model that argues back or interrogates a shaky premise gets treated as broken, not helpful. This paper suggests the opposite persona is achievable on a modest budget, using low LoRA ranks and few training epochs, on models anyone can download and modify. That has obvious appeal for tools meant to challenge assumptions rather than flatter them, like tutoring software or code review assistants.
The catch: the Socratic habit transferred cleanly to closely related languages but degraded in linguistically distant ones. So the cheaply reprogrammed open model still comes with an asterisk about which languages actually get the upgrade.