Following up on the issue in the previous post, this problem isn't just once or twice; it has been the most painful part of several days of use.
After careful analysis, I suspect the following possibilities:
- The criterion during training is "looking comprehensive" rather than considering the reading efficiency of responses. Being comprehensive for the sake of being comprehensive results in unclear organization.
- To prove what it did, the model orders its response as what it looked up, how it analyzed it, how it solved it, and how it verified it, and only gives the conclusion at the end, causing everything before that to be useless filler.
- To ensure the response is accurate, it tends to add various premises and assumptions, which makes the information volume of a single response very large and the reading cost extremely high.
The last point in particular is also the core reason for the frequent appearance of "X points you need to know" and "X points explained."
The current workaround can only be to use system prompts to constrain it, but this is clearly a problem at the model level. If the model itself has this problem, then the content it generates will naturally have this problem too.