A new research paper proposes putting the safety checker inside the language model itself, instead of bolting one on.
Researchers introduce SingProbe, a lightweight guardrail that reuses the hidden states an LLM already produces during inference, rather than running a separate model to check outputs. It predicts query intent, response safety, and hallucination risk token by token, layered directly into the decoding process. The system adds roughly 2 million parameters and less than 0.5% extra inference overhead, according to the paper. The team also built a benchmark, SingStreamBench, to test whether such guardrails stay quiet on safe text while catching unsafe content as it emerges, and extended the approach to medical text generation with a variant called SingProbe-Med.
Most current guardrails are separate models bolted onto the outside of an LLM, which means added latency, added cost, and a lag between when unsafe content is generated and when it's caught. If a checker this small and this fast can match larger standalone guardrails, that changes the calculus for anyone deciding whether to run safety checks at all on cost-sensitive deployments.
The results come from the paper's own benchmark on the paper's own model, so competitive performance against bigger detectors is a claim worth independent replication before anyone builds a product on it.