FactCheckingLLMs

Context: LanguageModels

There are lots of examples of ChatGPT and now MicrosoftBing getting things disturbingly wrong.

This is a serious concern.

However, it should be fairly straightforward to add supplementary "fact-checking" and "consistency-checking" systems

For example, when looking at LangChain the other day I and a friend came across this module for doing maths by (rather hilariously) telling the LLM that it shouldn't even try to do the calculation itself, but should defer the actual calculation to another engine : https://langchain.readthedocs.io/en/latest/modules/chains/examples/llm_math.html

Any time an LLM sees a number, this could easily trigger invoking extra parallel systems to infer further calculations which can be done by more traditional methods and used to cross-check. Dates could also trigger searches in online databases.

What an LLM "knows" is highly dependent on exactly what question it's asked. But you can imagine "query expansion" type setups where we infer useful secondary questions from the original one, which can be combined to get a "more rounded" response from the LLM

This is the equivalent of "fact-checkers". But it's also standard engineering practice to have secondary backup / fallback / monitoring systems. In fact it's kind of weird and unusual that we would expect any critical system (AI or not) to get everything right within a single / unsupervised / unchecked method or technique.