Fahmida Alam
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Efficient Safety Guardrails through Latent Dual-Hypothesis Reasoning with On-Demand Trace Reconstruction
According to the description, a preprint presents LR-Guard, a safety guardrail that reasons in compact latent form and can rebuild a readable explanation on demand. The authors report it beat its slower explicit-reasoning teacher in most in-domain and out-of-domain tests on safety classification and explanation extraction, with a 19.4x inference speedup at batch size one.