Open huyinan opened 5 months ago
Some more to solution: modify BaseMessage Class so that ReasonType is considered as a new argument. Then consider modifying the functions that convert BaseMessage to Openaimessages.
Specifically, if ReasonType = 'REACT', we will add the following as the instruction for the assistant to produce output messages: """Solve a question answering task with interleaving Thought, Action steps. Thought can reason about the current situation, and Action can be searching the right function calling and applying the correct function calling based on {key words}. """
Use llf-bench as an example to implement the 'REACT' process. Create another .py file in the example folder where the example of llf-bench is implemented. Specifically, find how every function/class in the original llf-bench repo is realized in camel agent repo. Test this case with an embodied agent under another .py file in the '/test/'folder.
Required prerequisites
Motivation
Recent research results, in particular the paper of Shunyu Yao et al entitled' REACT: synergizing reasoning and acting in language models', have demonstrated the possibility of combining verbal reasoning with interactive decision making in autonomous systems. The tight synergy between 'acting' and 'reasoning' allows humans to learn new tasks quickly and perform robust decision making.
Solution
Add an extra type named 'ReasoningType' to specify the categories of output system messages. 'REASONING' -- the default reasoning type which only contains verbal messages produced by the language model; 'ACT' -- messages that involve function calling procedure; 'REACT' -- A combination of reasoning and acting.
Alternatives
No response
Additional context
No response