A language model can receive an instruction about language, tone or intent before generating text—and still fail to reflect it in the response. In controlled experiments, Embedding Consistency Regulation (ECR) achieved language routing accuracy of 99.93% ± 0.12%, compared with 34–36% for random, shuffled and dummy controls, research by Chung-Wei Victor Yuan reports.
The study examines persistent semantic routing: whether an assigned route stays reflected in a model’s output throughout generation. It addresses language models operating under tight memory and compute constraints, where routes can be assigned before decoding but do not necessarily remain visible in the resulting text.
How Embedding Consistency Regulation tests route persistence
Embedding Consistency Regulation uses a frozen embedding model to derive a short, discrete prefix for each input. It applies the same prefix-construction process during both training and inference, allowing the research to test whether an assigned route persists through generation.
The reported accuracy gap comes from controlled prefix-assignment experiments. While ECR achieved near-perfect language routing accuracy, the random, shuffled and dummy controls stayed near chance levels at 34–36%.
The research also finds that likelihood measures do not reliably capture routing consistency.
Yuan introduces Response Routing Recoverability (RRR), an encoder-independent diagnostic that checks whether assigned routes remain observable after generation. It focuses on the response itself: does the generated text still reflect the route assigned before decoding?
The study reports that routing recoverability persists across multilingual settings, unsupervised partitions, reduced-precision conditions and transfer tasks.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
