Persistent semantic routing
A route is fixed before a compact model speaks. Is it still there in what it says?
YVIC Research studies compact, on-device language models — especially what they have already committed to before they decode.
Current work asks whether semantic routes such as language, tone, and intent survive generation, and whether prompt-injection structure is already visible before any token is produced.
Enter the research index ↓RRR / The assigned route, read back from the response with encoders independent of the one that assigned it.
Position
Language models change when they leave the lab.
YVIC Research is an independent effort in Taipei, focused on how language models behave once they run under real deployment limits — on phones and edge devices, with no cloud model behind them.
The current line of work treats a routing decision — language, tone, intent, task — as something fixed before a model decodes, and asks whether that route stays legible in the response. ECR reads a route from a frozen embedding space and carries it as a short discrete prefix through training and inference; RRR measures, with independent encoders, whether the route can still be recovered after generation. Routing can persist even where likelihood does not move — so it is worth measuring on its own.
A second thread approaches the same pre-decoding question from the safety side. PreDetect asks whether a prompt-injection or jailbreak attempt is already legible in a model's representation before it produces a token — read with a training-light, non-parametric margin over benign and injected reference anchors, with no change to the model. Where ECR asks whether an assigned route survives generation, PreDetect asks whether an adversarial intent is already there before it.
Research index
Current lines of inquiry
Researcher Engineer
Chung-Wei
(Victor) Yuan
袁崇惟
Selected work
Manuscripts & research systems
System study / fully offline
Luna lives
on the device.
Luna runs a compact language model and a local retrieval index entirely on the device — no network, no cloud model. It is the testbed for studying how routing and conditioning hold up once a model is quantized and deployed for real.
Maintained as a research system for local, private language-model deployment.
