Intel / article

Register Bias in Complexity-Based Large Language Model Routing

Source: arxiv.org
Author: Simran Koul
Date: Published Sep 17, 2026

Context

A study of complexity-based , the practice of sending easy-looking queries to a small model and hard ones to a large one, found that the routing step is not neutral to how a query is written (arxiv.org). Text in a non-standard English register, such as African American English or the English of second-language writers, gets systematically assigned a lower-capacity model tier than a meaning-equivalent standard-English version of the same question. The author traces this to a specific routing signal, input length: non-standard registers tend to omit function words, so they look shorter and are read by the router as simpler, even though other complexity signals used by the same router don't show this bias. Testing across a device-edge-cloud model ladder on 37,704 authentic learner sentence pairs, the study found every tier, including a frontier cloud model, answered non-standard-register queries less accurately, while the routing decision itself added comparatively little additional quality cost on top of that pre-existing gap.

Key takeaways · AI-distilled

The bias traces to one specific routing signal: input length. Non-standard registers omit function words and so read as shorter and simpler to the router, while other complexity signals tested don't carry this effect.
Every tier of the device-edge-cloud model ladder tested, including a frontier cloud model, answers non-standard-register queries significantly less accurately — the harm comes from bias present in every model, not just from routing to a weaker tier.
The marginal quality cost of the routing decision itself was not statistically significant on this — routing compounds a bias that already exists in every model rather than being the primary source of harm.
The disparity was demonstrated on 37,704 authentic learner sentence pairs plus a separate controlled parallel corpus, giving both real-world and controlled evidence for the same effect.

Terms in this piece · Glossary

model routing — Sending each request to a model chosen by the difficulty of the task, rather than using one model for everything.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source arxiv.org