Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model

• URL: http://arxiv.org/abs/2609.11291v2
• Date: Mon, 14 Sep 2026 22:24:38 GMT
• Title: Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model
• Abstract summary: We measure two behaviours the objective never targets: abstention on ambiguous social questions and unprompted disclosure in securities guidance.<n>Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone.
• Score: 2.8678936590409028
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Do Large Language Models Know What They Don't Know II? A Fully Behavioral, Non-Cognitive Measure of Epistemic Honesty [8.148357447770893]
  Large Language Models (LLMs) are frequently confident, eloquent, and well versed.<n>Do they know what they don't know?<n>We develop a novel metric to evaluate whether an LLM appropriately acknowledges the boundaries of its knowledge.
  arXiv  Detail & Related papers  (2026-09-07T18:40:08Z)
• FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect [69.8983512616053]
  Large language models often change their responses to subtle rephrasings that align with an implied stance by users.<n>This can leave users with advice tainted by how they happened to phrase a question rather than by the underlying facts.<n>We introduce FramingQA, a benchmark that measures the model sensitivity to question framing across law, medicine, finance, and robotic simulations.
  arXiv  Detail & Related papers  (2026-09-07T12:56:44Z)
• How Correct Is Your Answer? A Semantic Correctness Framework for Open QA Evaluation [38.482877747109946]
  We introduce a semantic correctness taxonomy that assigns open-ended answers to eight ordered classes, separating verbose-but-correct answers from those contaminated by hallucinated content.<n>Second, we release CAP-Correctness, an 8.8k-example benchmark spanning widely used QA datasets, and CAP-Statements, an 11k-example dataset for converting question-answer pairs into declarative statements.<n>Third, we introduce CAP (Context-Aware Precision), a reference-based metric that scores question-conditioned statements using bidirectional NLI.
  arXiv  Detail & Related papers  (2026-09-01T15:06:39Z)
• Consistency Without Alignment: Item-Sensitive Language Models Indistinguishable From Random [0.0]
  Item-sensitivity, defined as whether a model's choice depends on the specific input rather than on its own output prior, is widely reported as evidence of task competence.<n>We show this evidence is necessary but not sufficient using a forced-choice signalling task abstracted from the board game Deception: Murder in Hong Kong.<n>Across seven language models, two model families, a post-training ablation, and three independent scoring rules, every one of 21 model-by-rule cells is reliably item-sensitive.<n>We call this consistency without alignment and argue it generalises to any evaluation that relies on item-sensitivity, permutation consistency
  arXiv  Detail & Related papers  (2026-09-01T02:12:10Z)
• Feature Priming in Online Linear Regression: Sparse-Regret Lower Bounds and a Tight Univariate Rate [47.36630149538994]
  In high-dimensional online prediction, the best predictor may depend on only a few features, so regret should scale with sparsity rather than the ambient dimension.<n>Warmuth and Amid asked at COLT 2023 whether any of three such rules admits a competitive online regret guarantee.<n>Our analysis identifies a common obstruction: cheap nuisance causes the refit to underweight the truly predictive coordinate.
  arXiv  Detail & Related papers  (2026-08-18T09:32:03Z)
• Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families [51.56484100374058]
  Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models.<n>As a disposition probe we use 21,600 decisions under uncertainty, replayed through a frozen pipeline so the decision-layer model is the only variable.<n>Three effects replicate across both families (weeks-clustered CIs excluding zero)<n>A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping)
  arXiv  Detail & Related papers  (2026-07-19T22:27:00Z)
• How Much Human Label Variation Does Formal Semantic Structure Explain?: Group-Level Effects and Item-Level Ceilings in NLI [1.9177583131035885]
  Human label variation in natural language inference is increasingly treated as signal rather than noise.<n>We measure it on the 3,113 SNLI and MNLI items of ChaosNLI using a rule-based operator and monotonicity tagger validated against MED.<n>In this sample, formal semantic structure shifts how much annotators disagree by a small amount and does not detectably change what they disagree about.
  arXiv  Detail & Related papers  (2026-07-17T11:34:30Z)
• Detecting Hallucinations in Retrieval-Augmented Generation through Grounding-Aware Sensitivity by Perturbation (GASP) [1.827510863075184]
  Grounding-Aware Sensitivity by Perturbation (GASP) is a span-level detector that scores each answer sentence by how strongly its likelihood depends on the retrieved evidence.<n>GASP holds the answer fixed and re-scores it under the full context, under no context, and with each chunk removed, then measures the log-likelihood drops and Jensen-Shannon divergences.<n>We evaluate GASP on three benchmarks (RAGTruth, TofuEval, RAGBench) with three instruction-tuned scorers from two model families.
  arXiv  Detail & Related papers  (2026-07-05T10:35:30Z)
• On the Structural Non-Preservation of Epistemic Behaviour under Policy Transformation [51.56484100374058]
  We formalise such information-conditioned interaction patterns as behavioural dependency.<n>This induces a probe-relative notion of $$-behavioural equivalence and a within-policy behavioural distance.<n>Results identify structural conditions under which probe-conditioned behavioural separation is not preserved under common policy transformations.
  arXiv  Detail & Related papers  (2026-02-24T22:55:21Z)
• Towards Self-Supervised Covariance Estimation in Deep Heteroscedastic Regression [102.24287051757469]
  We study self-supervised covariance estimation in deep heteroscedastic regression.<n>We derive an upper bound on the 2-Wasserstein distance between normal distributions.<n>Experiments over a wide range of synthetic and real datasets demonstrate that the proposed 2-Wasserstein bound coupled with pseudo label annotations results in a computationally cheaper yet accurate deep heteroscedastic regression.
  arXiv  Detail & Related papers  (2025-02-14T22:37:11Z)
• Constructing interval variables via faceted Rasch measurement and multitask deep learning: a hate speech application [63.10266319378212]
  We propose a method for measuring complex variables on a continuous, interval spectrum by combining supervised deep learning with the Constructing Measures approach to faceted Rasch item response theory (IRT) We demonstrate this new method on a dataset of 50,000 social media comments sourced from YouTube, Twitter, and Reddit and labeled by 11,000 U.S.-based Amazon Mechanical Turk workers.
  arXiv  Detail & Related papers  (2020-09-22T02:15:05Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.