Wisent (https://www.wisent.ai/) just released fully open source latent space guardrails that monitor and stop unwelcome outputs of your LLM on the latent space level. Check it out here and happy to adopt it to your use case! https://github.com/wisent-ai/wisent-guard
On hallucinations it has not been trained on in TruthfulQA, this results in a 43% detection of hallucinations just from the activation patterns.
You can use them to control the brain of your LLM and block it from outputting bad code, harmful outputs or taking decisions because of gender or racial bias. This is a new approach, different from circuit breakers or SAE-based mechanistic interpretability.
We will be releasing a new version of the reasoning architecture based on latent space interventions soon to not only reduce hallucinations but use this for capabilities gain as well!
