idea

  • Improving ā€œLLM as a judgeā€ schemes by ensuring minimal semantic correlation between evaluation dimensions
  • ā€LLMā€™s donā€™t want to do workā€ - input a stipulation where an LLM is rating something that is supposed to be neutral, but stipulate that when it decides to give a bad rating it has to explain more
    • first would have to research bias of LLMs in rating tasks in general.
  • Align a model to ask questions more often, ā€œask and you shall receive"
  • "Entering the flow state with LLMsā€, use the sparse autoencoders idea where you can add a vector at the last layer and find out how much that improves LLM output
  • For robots arms that use cameras that are directly on the grippers, how can we introduce behaviour to ā€œtake a step back and look at the bigger pictureā€, before going back in?