Korean language model training changes how often and how much it answers
Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model
Artificial Intelligence
Summary
The researchers studied how training a Korean language AI model to write in a certain style can also accidentally change other behaviors. They found the model changes how often it chooses to answer questions and how much detail it gives, even when the training did not aim for these outcomes. For example, the model either answers ambiguous social questions more or less often, depending on the style used in training. These changes come more from whether the model decides to answer rather than the content of its answers.
What this means in practice
- •For ai product developers: Adjust how often a Korean language model answers questions by changing its response style training targets to improve user interaction.
- •For compliance teams: Monitor unprompted disclosures in AI-generated text by linking response style training to changes in answer frequency and detail.
Authors
Hyojung Han
Abstract
We post-train Qwen3.8-27B for Korean response style -- verbosity, list and markdown usage, discourse structure and register -- and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text -- observable without any reference labels.