TL;DR
This paper reveals that while post-training (SFT and DPO) makes LLM outputs appear uniformly "Standard English," the underlying training updates are highly selective, favoring prestige dialects like British English over stigmatized ones like African American English (AAE). To address this, the authors introduce Calibrated Accommodation Pairs (CAP), a preference tuning method that enables models to match a user's dialect density without over-correcting or imposing dialects on standard users.
What Problem Does This Paper Address?
Humans naturally adapt their speech to match their interlocutors—a phenomenon known as Communication Accommodation Theory (CAT). However, most state-of-the-art chat models default to a rigid "Standard English" regardless of the user's variety.
The authors argue that this uniformity is deceptive. An "audit" of model outputs might show that everyone gets the same standard response, but it hides what the model learned to prefer during training. If a model is trained to favor matching British spelling but disfavor matching AAE grammar, it reinforces a "Standard Language Ideology" that treats certain varieties as "correct" and others as "errors." The challenge is to enable models to accommodate diverse users authentically without falling into the traps of over-accommodation (caricature) or imposition (forcing a dialect on someone who didn't use it).
Methods
The researchers developed a multi-stage framework to "read" the training process:
- Twin Requests: They created a dataset of 287 requests, each rendered in six English varieties (e.g., AAE, Singlish, British English) and two registers (Formal, Informal).
- Convergence Ratio ($CR_v$): A metric to measure how much of a user's dialect density is reflected in the model's response, specifically looking at "novel spans" to ensure the model isn't just copying the user's words.
- Update Diagnostic: By calculating the length-normalized log-ratio between consecutive training stages (e.g., SFT vs. DPO), they could determine which reply types the update actually "favors," even if the final output remains standard.
- Calibrated Accommodation Pairs (CAP): A novel preference data recipe. Instead of just "dialect vs. standard," CAP uses three types of pairs:
- U (Under-accommodation): Prefers a dialect-matched reply over a standard one for a dialect user.
- O (Over-accommodation): Prefers a reply matching the user's specific density over a "hyper-dialect" version.
- A (Against imposition): Prefers a standard reply for a standard user.
Datasets
- Twin Request Dataset: 287 first-person requests across six task types, manually rendered into varieties including AAE, Singlish (SGE), Nigerian Pidgin (NGP), Appalachian English (APP), British English (BRE), and Australian English (AUS).
- Multi-VALUE: Used to apply morphosyntactic features for dialect rendering.
- WildChat: 162,842 English turns used to audit the rarity of dialect features in real-world chatbot traffic.
- EnDive: A cross-dialect benchmark used for evaluating held-out variety convergence.
Key Insights and Ideas
- The "Standard" is a Post-Training Default: Base models actually show some tendency to mirror user dialects. It is the post-training pipeline (SFT and DPO) that aggressively standardizes replies across all informal cues.
- Updates are Selective: Even when outputs look identical, the DPO update favors matching prestige dialects (British/Australian spelling) over stigmatized ones (AAE/Singlish) across all 10 model lineages tested.
- Uniformity Hides Bias: Output audits are insufficient. Researchers must look at the update (the change in likelihood) to see the true preferences installed during alignment.
- Prestige Tiers Exist: The models' internal preferences align with social prestige hierarchies found in human attitude studies, where British English sits at the top and AAE/Singlish are at the bottom.
Key Results
- Standardization is Universal: Across 10 lineages, convergence to informal cues drops significantly after DPO. For example, in OLMo-2-7B, AAE convergence is near zero by default.
- CAP Effectiveness: On the OLMo-2-1B model, CAP raised held-out convergence by 0.173 over matched controls. Crucially, it did this with zero detectable increase in dialect insertion for Standard English users.
- Overshooting in One-Sided Pairs: Simple "one-sided" preference pairs (always preferring dialect for dialect users) failed to calibrate, often leading the model to "overshoot" and produce caricatured, hyper-dialectal responses.
- Competence Remains: Despite standardizing by default, models still can produce AAE features when explicitly asked, realizing them at over 50% of authentic rates.
Limitations and Future Work
The authors acknowledge several limitations:
- Rule-Based Stimuli: The requests are rule-based renderings, which may not perfectly capture the nuance and density of authentic dialectal text.
- Finite Feature Inventory: The detectors only cover a specific set of linguistic features.
- Scale: The causal re-runs using CAP were primarily performed on smaller models (1B and 8B).
- Social Context: The paper measures asymmetry and ability, but notes that the decision of when a model should use a variety is a complex social question that should involve the speakers of those varieties.
Future work could involve expanding CAP to more diverse languages and varieties and exploring how these preferences interact with safety and toxicity filters, which often disproportionately flag stigmatized dialects.
Summary updated