PaperlayerView exact revision on arXiv

Source-linked AI summary

Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training

Zhizhuo Kou, Heming Li, Jian Yang, Junyu Luo, Jingshu Peng, Sirui Han, Yike Guo

arXiv:2610.28474v1cs.CV

TL;DR

Chat models often default to Standard English, and output audits alone cannot establish whether training updates treat dialect users alike. The paper tracks outputs and update preferences across open model lineages, then tests calibrated accommodation pairs. Post-training standardizes informal cues while updates selectively favor some varieties, and calibrated pairs raise held-out convergence without detectable dialect insertion for Standard-English users.

  • Problem

    Chat models tend to default to Standard English across user varieties, while outputs alone cannot show whether training updates treat users alike.

    Source: §1 · p. 1

  • Method

    The study tracks replies and content-matched update preferences across 10 open lineages, then tests preference pairs calibrated to user dialect density.

    Source: §1 · p. 2

  • Results

    Across open lineages, post-training standardizes informal cues while update diagnostics show selective preferences; on OLMo-2-1B, CAP raises held-out convergence by 0.173 over matched controls without detectable insertion for Standard-English users.

    Source: §6.2 · p. 7, §7 · p. 8

  • Takeaways & Limitations

    Output audits can miss selective training preferences, so the authors recommend reporting update diagnostics alongside audits when consecutive checkpoints are public.

    Source: §7 · p. 8

    The causal re-run is limited to OLMo-2-1B, with effects appearing only at the largest tested dose and varying across runs.

    Source: Limitations · p. 9

Abstract

from arXiv · show

People adapt how they talk to the person in front of them. However, chat models tend to respond in Standard English regardless of the dialect used by the user. A uniform answer can hide whether training treats these users alike, and output audits cannot tell as well. We trace how each post-training stage, from supervised fine-tuning (SFT) to preference optimization (DPO) and reinforcement learning, changes a model's tendency to match the user, and read what each update prefers from two content-matched replies, one matching the user and one not. Across 10 open model lineages, post-training standardizes replies to every informal cue, yet after adjusting for how likely each reply already was, the DPO update still favours matching British and Australian spelling over AAE and Singlish on 10 lineages, and models still produce AAE features when asked. We then introduce calibrated accommodation pairs (CAP), preference data that reward matching users' dialect density while preserving Standard English for Standard-English users. On a fully open model, CAP raises convergence on features it was not trained on without a detectable increase in dialect for Standard-English users or a cost in instruction following. In contrast, one-sided pairs reach a similar gain only by overshooting the user and imposing dialect. Our findings show that uniform replies can hide selective updates and audits should read the update as well as the output. Code: https://github.com/kouzhizhuo/cap-accommodation.

1 Introduction

Chat models often default to Standard English across varieties, leaving open which post-training stage creates that pattern and whether training updates treat users alike.

1 Introduction

  • The introduction asks whether supervised fine-tuning or preference optimization installs the Standard-English default and whether training treats varieties alike.Identical Standard-English outputs can conceal different learned preferences.Source: §1 · pp. 1–2
  • The authors track replies and training updates across 10 open lineages, using content-matched replies to compare matching and non-matching responses.They score only reply text not copied from the request.Source: §1 · p. 2
  • The paper reports uniform standardization alongside selective preference updates, and introduces diagnostics intended to separate output behavior from training preferences.The reported DPO preference favors British and Australian spelling over AAE and Singlish on all 10 lineages.Source: §1 · p. 2
  • The authors note that several preregistered predictions did not hold and leave decisions about replying in a user’s variety to its speakers.Source: §1 · p. 2

2 Background and Related Works

The background frames accommodation as socially motivated speech adjustment and contrasts it with priming and standard-language-ideology accounts of convergence.

2 Background and Related Works

  • Communication Accommodation Theory describes convergence, divergence, and maintenance as ways speakers adjust speech to interlocutors for social reasons.Convergence may flow upward toward a higher-status variety, but speakers able to converge may choose not to.Source: §2 · p. 2
  • Priming accounts predict convergence to every variety, while prior language-model studies measure style and this paper measures features distinguishing varieties (Pickering and Garrod, 2004; Blevins et al., 2026).Source: §2 · p. 2
  • Standard-language ideology predicts convergence toward the standard and prestige cues, but not stigmatized grammar.Prior work reports covert prejudice against AAE speakers, reward-model dispreference for AAE, and dialect disparities from preference tuning (Mire et al., 2025; Ryan et al., 2024).Source: §2 · p. 2
  • The paper treats the ratio between consecutive released stages as a descriptive update diagnostic, not an identified reward, to show which content-matched reply a stage up- or down-weights.DPO’s log-ratio between trained and reference policies is an implicit reward (Rafailov et al., 2023).Source: §2 · p. 2

3 Method

The method compares matched variety and Standard-English requests across released model stages, measuring both reply convergence and preference shifts between content-matched answers.

3 Method

  • The study renders 287 first-person requests across six task types in six varieties, Standard American English, and informal and formal Standard English controls.Local native speakers with Ph.D.-level education rendered the requests; content was held fixed.Source: §3 · p. 2
  • Ten open model lineages provide public checkpoints at every post-training stage, with two replies sampled per request and seed across three seeds.Source: §3 · p. 3
  • Convergence scoring excludes copied 4-gram spans and quotations, and variety-feature density is based on detected realizations and opportunities for those features.Source: §3 · p. 3
  • The convergence ratio normalizes reply differences by the request density gap: 1 means replies differ between twins as much as requests do, while 0 means not at all.For varieties with too few opportunities, the study uses a rate-based ratio; feature competence compares requested feature rates with authentic text.Source: §3 · p. 3
  • The matching bonus is positive when an update favors a matched reply more for a variety speaker than for a Standard-English speaker.The update diagnostic uses the length-normalized log-ratio between a stage and the preceding stage.Source: §3 · p. 3
  • The analyses use 95% bootstrap confidence intervals over requests and standard deviations over seeds, and the design, measures, and predictions were preregistered.Source: §3 · pp. 3–4
  • Detector outputs agree with blind expert-panel consensus labels on 28 of 30 requested replies (κ = 0.814).Recognizers are applied only to model replies because they over-fire on raw user requests.Source: §3 · p. 4

4 Replies Converge on Standard English

Across open model families, post-training reduces convergence to informal cues while models retain some ability to produce AAE features when explicitly asked.

4.1 Post-Training Standardizes Replies to Every Informal Cue

  • For OLMo-2-7B, British-spelling convergence falls from 0.171 at base to 0.120 after SFT and 0.058 after DPO, then is 0.060 after RLVR.Informal-register convergence falls from 0.413 at base to 0.009 after DPO; formal-register convergence remains 0.141 after DPO.Source: §4.1 · p. 4
  • After SFT, AAE grammar and Singlish lexis are near zero in OLMo-2-7B, whose base-model rates were 0.022 and 0.051, respectively.The reported AAE result also holds on human-written AAE tweets.Source: §4.1 · p. 4
  • The prestige ordering is present at base and remains after DPO, with convergence correlated to preregistered prestige tiers at ρ = 0.778 and ρ = 0.845, respectively.Post-training lowers convergence levels but does not install the ordering.Source: §4.1 · p. 4
  • From base to final release, informal-register convergence falls in all 12 further families, British-spelling convergence in 9, and AAE ends at most 0.001.The broader analysis finds post-training standardizes replies for users of every variety, not only stigmatized varieties.Source: §4.1 · p. 4
  • Across five deployed chatbots, convergence is at most 0.053 for informal register and 0.009 for AAE, while Nigerian Pidgin is reproduced at 0.116–0.502.These Nigerian Pidgin rates exceed those of post-trained open models, which reach at most 0.046.Source: §4.1 · p. 4

4.2 Models Still Produce AAE Features on Request

  • When asked to use AAE, OLMo-2-7B after DPO realizes features at 0.567 of their authentic rate and Tülu-3-8B at 0.588; OLMo-3-7B reaches 0.411.Requested imitation increases stereotyping, so the competence measure is not a measure of reply quality.Source: §4.2 · p. 4

5 The Update Is Selective

The update diagnostic reveals selective treatment beneath standardized replies: adjusted DPO updates favour matching British and Australian spelling over AAE and Singlish. Accommodation-consistent pairs raise convergence only at the largest tested dose, without calibrating replies to users.

5.1 The DPO Update Favours Matching Prestige Spelling over Stigmatized Varieties

  • The adjusted DPO update favours matching British and Australian spelling over AAE and Singlish on all 10 lineages.The pooled contrast is 0.020 nats per token (CI [0.014, 0.027]); the paired test postdates pre-registration.Source: §5.1 · pp. 4–5, Figure 3 · p. 5
  • The adjusted prestige-versus-stigmatized contrast was positive in 12 of 14 further updates, reversed in none, and inconclusive in two.The two inconclusive cases were SmolLM3’s preference stage and Qwen3.5-4B.Source: §5.1 · p. 5

5.2 Added Preference Pairs Raise Convergence Only at the Largest Dose Tested

  • At 60k added pairs (33%), held-out convergence rose in all four runs, averaging 0.043, and transferred to held-out varieties by 0.035 on average.At 1k and 5k pairs and at 20k (0.001, CI [−0.002, 0.004]), held-out convergence stayed at dose-0 levels; IFEval remained within the registered two-point bound.Source: §5.2 · pp. 5–6, Figure 5 · p. 6
  • Added pairs raised convergence without calibrating replies: Standard-English replies gained dialect features, while Singlish was over-produced and AAE under-produced.Across four 60k-pair runs, Standard-English replies gained 0.10–1.06 trained features per 100 words; Singlish replies reached 1.07–9.47 times the user’s trained-feature rate and AAE 0.17–0.47 times.Source: §5.2 · p. 6

5.3 Sensitivity Analysis

  • The findings persisted when the estimator, bootstrap, reply length, sampling seeds, or human-audit correction changed.The sensitivity analyses are reported in Table 19 and Section C.4.Source: §5.3 · p. 6

6 Calibrated Accommodation Pairs

Calibrated accommodation pairs (CAP) encode preferences for matching a dialect user’s feature density while avoiding over-accommodation and dialect imposition for Standard-English users. On OLMo-2-1B, CAP increased held-out convergence without detectable insertion or an instruction-following cost, though several registered calibration and compliance criteria failed.

6 Calibrated Accommodation Pairs

  • CAP re-renders the model’s reply at the dialect user’s trained-feature density and pairs it against under-, over-, and imposed-dialect alternatives.For Standard-English users, the preferences retain the model’s Standard-English-twin reply over a fully dialect-rendered alternative; the pair mix is U : O : A = 2 : 1 : 1.Source: §6 · p. 7, Figure 6 · p. 7

6.1 Experimental Settings

  • The evaluation pre-registered tests for held-out convergence, calibration, density tracking, instruction following, and Standard-English compliance; the pilot under-accommodated but proceeded without tuning.The study trained fully released OLMo-2-1B stages at a fixed budget and tested 300 fresh requests from a disjoint pool.Source: §6.1 · p. 7

6.2 Main Results

  • CAP raised held-out convergence by 0.173 over matched controls on OLMo-2-1B, without detectable insertion for Standard-English users or an IFEval cost.For held-out varieties, the gain was 0.150; insertion changed by 0.001 features per 100 words (CI [−0.008, 0.013]) and IFEval changed by 2.1 points (CI [−0.7, 4.9]).Source: §6.2 · p. 7, Table 4 · p. 8
  • CAP’s calibration and tracking criteria and Standard-English request compliance failed as registered, despite passing held-out convergence and IFEval criteria.Singlish replies no longer overshot user density (0.79 versus 2.59), but these registered criteria remained unmet.Source: §6.2 · pp. 7–8, Table 4 · p. 8

6.3 Ablation Study

  • Without O pairs, held-out convergence reached 0.583 but Standard-English replies gained 0.039 dialect features per 100 words; removing A pairs or the preference contrast removed the gain.One-sided pairs from CAP’s builder gained 0.201 but produced Singlish replies at 10.41 times user density and inserted 0.378 features per 100 words.Source: §6.3 · p. 8, Table 5 · p. 8

6.4 Sensitivity Analysis

  • CAP’s held-out gain persisted under alternative bootstrap and exclusion choices and on the earlier request set, while Qwen3-8B’s CAP gain was 0.017, below the one-sided recipe’s 0.154.The Qwen3-8B result fell short of registered criterion M1.Source: §6.4 · p. 8, §6.2 · p. 8

7 Discussion

Output audits can show the same Standard-English answer across users, while update diagnostics reveal that training moves the policy toward matching some users and away from others. The authors recommend reporting update diagnostics alongside output audits.

7 Discussion

  • The diagnostic reads public checkpoints without retraining, but raw and adjusted contrasts answer different questions and should both be reported.The prestige ordering is consistent with prestige-directed convergence but does not establish prestige as its cause; the standard default arises with little dialect-specific data and is not inevitable.Source: §7 · p. 8

8 Conclusion

Across released stages, post-training standardizes replies even as updates favour some varieties over others. CAP improves held-out convergence without detectable insertion for Standard-English users, whereas one-sided pairs reach a similar gain by overshooting and imposing dialect.

8 Conclusion

  • Adding accommodation-consistent pairs to a third of DPO data raised held-out convergence over matched controls in every run, but did not calibrate replies.CAP addresses this failure on OLMo-2-1B; explicit requests for Standard English still receive some dialect, and one-sided pairs reach similar gains only by overshooting and imposing it.Source: §8 · p. 8

Limitations

The study covers English varieties and public open-model lineages, but its requests, feature inventory, and validation constrain how its findings generalize. Its update audit is correlational, and the causal re-run is limited to OLMo-2-1B and varied across runs.

Limitations

  • The study covers English and six varieties, with every stage public for the open lineages studied.Closed chatbots enter only through their behaviour.Source: Limitations · p. 9
  • Rendered requests are not calibrated to authentic text’s feature densities, so convergence on real users’ writing may differ.Only competence and a check on human-written AAE twins are anchored in authentic text.Source: Limitations · p. 9
  • Detectors cover a finite feature inventory, and human validation uses consensus labels without independent double annotation or reported inter-annotator agreement.The main audit sample contains 84 items.Source: Limitations · p. 9
  • The update diagnostic shows which content-matched replies stages favour, not why; the preference-data audit is correlational.The released models were not trained by the authors.Source: Limitations · p. 9
  • CAP’s causal re-run uses OLMo-2-1B, its effect appears only at the largest tested dose and varies across runs, and several registered criteria fail.CAP passes M1 and M4 but fails M2, M3 and M5 on OLMo-2-1B, and does not pass M1 on Qwen3-8B.Source: Limitations · p. 9

Ethical Considerations

The paper describes safeguards for corpus use and limits on data release, reports how human labels were obtained, and flags possible caricature from over-accommodation. It also discloses AI assistance in code writing and polishing.

Ethical Considerations

  • Corpora were used under their licences, raw corpus text was not redistributed, personal data were masked before display, and synthetic dialect text was labelled synthetic.WildChat requests were used only where the release flags no personal data.Source: Ethical Considerations · p. 9
  • The release includes code and derived results, but not raw text from corpora whose licences do not state redistribution or the retrained checkpoints.Replies from deployed chatbots will be released subject to provider terms.Source: Ethical Considerations · p. 9
  • No crowdworkers or external participants were involved; human labels came from a panel of five or more PhD-level annotators who reached consensus through discussion.The items were not annotated independently, so inter-annotator agreement is not reported.Source: Ethical Considerations · p. 9
  • In one of four runs, a large dose of accommodation-consistent pairs overproduced trained features, which risks caricature.The paper notes that accommodation is not a quantity to maximize and that Black American participants want choice.Source: Ethical Considerations · p. 9
  • AI assistants were used for code writing and polish, while researchers designed and checked the research.Source: Ethical Considerations · p. 9

A Data and Resources

This section documents the model and data resources, experimental and auditing procedures, CAP design, and validation of dialect detectors against expert judgments. It reports that CAP did not meet its pre-declared held-out-convergence criterion on Qwen3-8B, identifies limits from detector error, and finds that alternative measures and explanations do not account for the observed pattern.

B.1 Experimental Details

  • The stimuli comprise 287 first-person requests across six task types, tagged formal or casual and rendered with fixed content across conditions.Variety renderings use feature-specific rules and lexicons, but were not calibrated to authentic feature densities or rated for naturalness.Source: §B.1 · p. 15

B.2 Detectors, Human Audit and LLM Judges

  • The detector audit found that some apparent default-reply hits were false positives, while rare post-training AAE hits were mostly multiple negation.The panel confirmed 6 of 14 audited default-reply hits; multiple negation was not confirmed in any of its four audited hits, and detector hits also arose on templated prompts, code, and names.Source: §B.2 · pp. 17, 20
  • The human audit validated the formality ranker but not the part-of-speech score, and competence estimates use human-corrected detector rates.The ranker matched the panel on 12 of 15 pairs; requested-reply detectors had 6 of 7 hits confirmed, but precision uncertainty required correction, and neither audited Singlish feature entered the competence score.Source: §B.2 · p. 19
  • The two LLM judges matched consensus on 43 of 44 feature items for gpt-oss-20b and 34 of 44 for Qwen3-8B.They called only 3 and 0 of 40 register pairs equally formal, respectively, versus 25 for the panel; their labels were not used in validation rules or reported effects.Source: §B.2 · p. 20

B.3 Calibrated Accommodation Pairs

  • CAP pairs compare density-matched dialect replies with standard or over-dense alternatives, plus Standard-English replies against dialect-imposed alternatives.The U:O:A mix is 2:1:1; before training, only 0.39 of mirrored replies fell within 25% of user density, and over-dense replies were less fluent by 0.18 nats per token.Source: §B.3 · p. 20
  • On Qwen3-8B, CAP raised held-out convergence by 0.017 (CI [0.009, 0.026]), below the pre-declared 0.02 minimum.In a blind check, the panel marked 0 of 20 CAP replies and 0 of 20 one-sided replies as exaggerated, so the check did not distinguish them.Source: §B.3 · p. 20

A.2 Artefacts and Licences

  • The release includes rendered requests, validity gates, feature detectors, and audited materials, with artefact licences and uses listed in Table 8.The code and data are described as research-use artefacts; no corpus text is redistributed, and only short anonymized examples are quoted.Source: §A.2 · p. 15, Table 8 · p. 18, §B.4 · p. 21

B.4 Further Method Details

  • The study tracks public post-training lineages and supplements them with broader model families and deployed chatbots; the 70B Tülu-3 and Tülu-2 lineages were not run for compute reasons.Variety features must pass false-positive checks, and the convergence ratio measures differences from Standard-English twins rather than absolute matching to user density.Source: §A.1 · p. 15, §B.4 · p. 21
  • Expert labels and detectors agreed on 28 of 30 requested replies (κ = 0.814), and the panel confirmed 6 of 7 detector hits.The GYAFC-trained formality ranker selected the panel’s more formal reply in 12 of 15 pairs; the part-of-speech score failed at 9 of 15, and no claim rests on it.Source: §B.4 · pp. 21–22
  • Detectors over-flagged raw user requests, so the analysis applies them only to model replies; default convergence remains subject to detector error.Experts confirmed 3 of 18 flagged WildChat requests and 6 of 14 default-reply hits; false hits, mostly multiple negation, occurred similarly for Standard-English twins (146 versus 121 over 30,996 replies each), so they largely cancel in the ratio.Source: §B.4 · p. 22
  • Prior function-word measures barely changed across varieties and stages, while instructions, personas, and copying did not explain the pattern.Source: §B.4 · p. 22

B.5 Further CAP Details

The CAP ablations show that matching and anchoring contrasts contribute to held-out convergence, while density matching curbs overshoot. Across additional results, outputs often remain standardized even as updates reveal selective preferences, and the authors recommend auditing both.

B.5 Further CAP Details

  • The pilot passed 3 of 4 checks but under-accommodated, reproducing 0.50 of user density for AAE and 0.26 for Singlish without inserting dialect for Standard-English users.The ablations compare components at the same budget; the test uses 300 fresh requests disjoint from training and validation.Source: §B.5 · p. 22
  • CAP’s one-sided comparison gains 0.201 held-out convergence but inserts 0.378 features per 100 words for Standard-English users, versus 0.001 for CAP.One-sided Singlish replies reach 10.41 times the user’s density, while CAP’s reach 0.79 times it.Source: §B.5 · p. 23
  • Without A pairs, held-out convergence barely moves (0.008); raising the mirror target to 1.5 times user density multiplies it by 2.0 (0.340) without insertion.The matching and anchoring contrasts carry the gain, over-accommodation contrast buys restraint, and density matching curbs Singlish overshoot.Source: §B.5 · pp. 22–23
  • CAP’s held-out gain is 0.137 on 287 earlier requests and remains 0.173 when the 2% false-positive exclusion is removed.The bootstrap interval over requests alone is [0.151, 0.198], compared with [0.138, 0.207] over runs and requests.Source: §B.5 · p. 23

C.1.1 Case Study

  • After post-training, replies to human-written AAE tweets show at most 0.017 convergence, while the standard-English case-study replies also omit the AAE cues.The post-training outcome on human-written AAE meets the near-zero condition on all five lineages, but both standardization conditions hold on only one of four lineages with a base model.Source: §C.1 · p. 23, §C.1.1 · p. 24

C.2 Preference-Data Audit and DPO Re-runs

  • Dialect-bearing prompts make up 0.81%, 0.81%, and 0.64% of audited preference mixes, but the audit is correlational and recognizer-selected WildChat requests were excluded.On AAE-bearing prompts, chosen and rejected replies differ in AAE density by −0.006 for OLMo-2-1B and −0.007 for OLMo-2-7B; both estimates have intervals spanning zero.Source: §C.2 · pp. 26–27

C.3 Extended Discussion

  • Standard replies arise with little dialect-specific data, but models retain some ability to produce AAE features when asked.The paper notes that accommodation-consistent pairs raise convergence without calibration to the user.Source: §C.3 · p. 28
  • The diagnostic finds update preferences that output audits miss: post-training lowers convergence, while adjusted DPO updates favour prestige spelling over stigmatized varieties on all 10 lineages.The diagnostic compares content-matched replies and describes what an update favours; it does not identify why.Source: §C.3 · pp. 27–28

C.4 Further Results Details

Further results report the scope and robustness of the paper’s findings: preference updates lower convergence, while adjusted comparisons retain a prestige-spelling advantage. The section also records the measurement framework and pre-registered predictions against which the results are interpreted.

C.4 Further Results Details

  • The preference stage lowers informal-register convergence on all 10 lineages and British-spelling convergence on 9 of 10; adjusted contrast is positive in 12 of 14 further updates.The pooled adjusted contrast is 0.020 nats per token.Source: §C.4 · p. 28
  • OLMo-2-7B’s informal convergence falls from 0.413 at base to 0.039 after SFT and 0.009 after DPO under the chat template.After DPO, AAE detector-hit density is 0.001 for AAE requests and 0.000 for their Standard-English twins.Source: §C.4 · p. 28
  • The preference stage lowers informal convergence on all 10 lineages and British-spelling convergence on 9 of 10, while the prestige–stigmatized split falls from the first released stage to DPO on all 10.Base-to-SFT declines are counted only when they hold under all three base framings.Source: §C.4 · p. 28
  • After adjustment, the OLMo-2-32B prestige contrast is 0.012 (CI [0.002, 0.022]) despite a raw contrast of −0.031, because its update up-weights unlikely replies generally.OLMo-3.1-32B does not reverse; its raw contrast is 0.030 and its likelihood slope is −0.014.Source: §C.4 · p. 29

D.1 Theory and Pre-Registered Hypotheses

  • The paper’s measurement framework defines convergence as reproduced user-variety feature density on novel spans relative to a matched Standard-American-English request.Its split index is mean convergence for prestige cues minus mean convergence for stigmatized varieties.Source: §D.1 · p. 30, Table 20 · p. 31
  • The pre-registered hypotheses test where the split arises, selectivity by prestige, retained competence on request, update preferences, and whether added pairs raise convergence with matched helpfulness.Secondary predictions concern convergence’s relation to matched-guise prejudice and validity against prior accommodation measures.Source: §D.1 · pp. 30–31

D.2 Pre-registered Predictions and Outcomes

  • The paper reports all pre-registered outcomes, including predictions that did not hold, alongside its statistical procedures.Falsifiers and alternative predictions were specified before the main results were read.Source: §D.1 · p. 31, §D.2 · p. 31

D.2.1 Register Divergence Is Not Confirmed

The validated formality analysis does not confirm register divergence after post-training, and the section documents analytical procedures and deviations from preregistration. It situates the work among accommodation, alignment, and dialect-bias research while distinguishing its feature-based and update-focused measurements.

D.2.1 Register Divergence Is Not Confirmed

  • After DPO, 0 of 24 cells are more formal and 19 are less formal; for OLMo-2-7B AAE, the difference is −0.015 (CI [−0.023, −0.008]), suggesting replies to vernacular users are slightly less formal.Source: §D.2.1 · p. 32
  • The validated ranker finds 2 of 24 cells more formal and 4 less formal after SFT, unlike the unvalidated part-of-speech score’s 9 suggested cells.In base models the formality difference is positive in 20 of 20 cells.Source: §D.2.1 · p. 32

D.3 Statistical Procedures

  • The analysis uses request-cluster bootstrap intervals and paired within-request contrasts, with the adjusted update comparison added after preregistration.The adjustment controls for matched-reply SFT likelihood and token length.Source: §D.3 · p. 33

D.3.1 Pre-registration

  • The final request set contains 287 rather than 300 requests, and bootstrap resampling without task-type stratification was a recorded deviation.Resampling within task type leaves all 10 adjusted contrasts above zero, with intervals 0.93–1.02 times as wide.Source: §D.3.1 · p. 33

E.1 Related Work

  • Earlier accommodation measures capture style, whereas this study measures variety-distinguishing features and separates reply behaviour from what updates favour.The reported function-word measures change by less than .04 between varieties and stages on the replies.Source: §E.1 · p. 34
  • The discussion frames post-training as a directed shift toward a standard, rather than only a reduction in output diversity.The cited account links alignment tuning to mostly stylistic token shifts.Source: §E.1 · p. 34
  • Unlike studies scoring fixed texts with a reward model or judge, this work compares the policy update on content-matched replies differing in whether they match the user’s variety.The update ratio is used descriptively, not as an identified reward.Source: §E.1 · p. 34
Loading 2610.28474v1…