Very interesting topic. I also wonder why other signs of AI writing, such as negative parallelism ("It's not just X, it's Y"), are preferred by the models.
Also, I wrote a small extension that automatically replaces ChatGPT responses with em dashes with alternative punctuation marks: https://github.com/nckclrk/rm-em-dashes
As a CLI, this tool is most efficient when it can see text outputs from the commands that it runs. But you can help it with visual tasks by putting a screenshot file in your project directory and telling claude to read it, or by copying an image to your clipboard and pasting it with CTRL+V
I agree, it kind of reminds me of this paper that shows LLMs, just like humans, will preferentially remember information, which can lead to biased outputs.
The best way to identify the values LLMs hold is not to give them a survey with questions of the form "I believe in the value of justice. Agree or disagree?"
Instead you need to present scenarios which put these beliefs into effect. This is similar to how humans may claim to value sustainability but that doesn't mean when making a purchasing decision they will always opt for an eco-friendly option.
I'm interested if this might be a more robust method to resolve conflicting information across retrieved documents. Instead of having the LLM reason over which source to trust, the solution is to incorporate the knowledge into parametric memory and see how this combination of perspectives produces a final response.
I included a section at the bottom that provides a sample bibtex citation. I didn't expect this much attention so I didn't even bother with a License but I'll include a MIT license later today and release 0.2.1
Also, I wrote a small extension that automatically replaces ChatGPT responses with em dashes with alternative punctuation marks: https://github.com/nckclrk/rm-em-dashes