Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by ben_w on Monday October 09, 2023 @03:57PM
(#37824720)
Sometimes I solve maths problems by visualising: perhaps I want to differentiate something, so I imagine the graph of the function, and then the slope of the graph, and then try to compare that to various simple functions that I could give as a guess of the answer, which I only then feed into a symbolic process of "assume this is true and work backwards".
[ Parent
| on HN ]
Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by og_kalu on Monday October 09, 2023 @04:12PM
(#37824899)
an LLM trained on just othello moves will reconstruct a board state from the sequence of moves to aid in prediction of the next move. You have no idea what an LLM is or is not doing to predict. Prediction is just the objective. Don't confuse that for the process.
Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by ben_w on Monday October 09, 2023 @04:37PM
(#37825199)
> You have no idea what an LLM is or is not doing to predict.
Indeed, but I'm not trying to with that comment, which is just about how my own seems to work on self-reflection. That said, I do have reason to doubt the accuracy of human introspection of our own thought processes, and therefore my own judgment in this may also be flawed.