zlacker

It can also "correct" proper reasoning. :)

~"When told where it's wrong, LLM can correct itself to improve accuracy."

Similar to cheating in chess- a master only needs to be told the value of a few positions to have an advantage.

>>seekno+(OP)
This is said in the abstract as well:

> recent attempts to self-correct logical or reasoning errors often cause correct answers to become incorrect, resulting in worse performances overall (Huang et al., 2023)

>>seekno+(OP)
I have noticed this several times. When I give feedback that a mistake was made (with no details on what the mistake is), often smaller and medium size LLMs then give a correct response.

>>mark_l+Bp
Which I take full advantage of when the output is like 90% correct but the "fix" requires a bit of refactoring, I just tell it what I want and presto. Faster than doing it by hand.

>>tines+k4
Yeah but in this complex way that really glosses over what's going on here.

Plus, sometimes the corrections aren't accurate. So of course if you tell it where it's wrong, and it gets a second chance, the error rate will be less...