Key intuition and motivation:

the generation of api calls can be explicitly represented using states. By eliminating tokens outside the predefined states vocabulary, we can greatly reduce hallucinations. Core contribution: beyond a extensible method in augmented LLM, it provides a general method of incorporating domain-knowledge to eliminate hallucinations.

Future direction:

In LLM applications (in augmented LLM or biomedical), a core problem is how to incorporate prior knowledge with LLM to avoid mistakes and hallucinations. FSM is good for tool usage, but not transferable for other domains such as biomedical. Following this research line, the next step could be developing a general method that use prior knowledge to find a high-quality reliable subset of tokens which eliminate hallucinations as much as possible.

For example, use “phrase” (in tool usage, api name is a phrase, in API decode, api name and parameter type are somehow “phrase”) to eliminate hallucinations.

Paper link: ToolDec

Brief Summary

This paper provide a pipeline for choosing and using biomedical functions/apis. The pipeline combines

• a BERT-based API retriever. It will take a task instruction and return a short list of API candidates.
• a GPT-4-based API predictor. It will (with in-context learning) choose the most relevant API.
• a GPT-4-based generation pipeline of api calling that takes several steps to give the final script and show the result.

Limitation

The backend of API calling (and prediction) is completely gpt4-based, which leads to unavoidable hallucination and limits the performance this work can possibly attain.

One possible direction is to finetune the backend of the API calling.

Paper link:BioMANIA