✳flâneur — a map of the web's best reading

Training on Mistakes Improves Problem Solving Performance

gytis.co · 1,117 words · saved by 1 readers

Error correction helps problem solving agents, even those which are limited in capability, to solve a much broader range of problems. This is particularly important as the Internet weighs towards finished artifacts (like books, posts, and articles) that have scant few examples of making an erroneous claim and then backtracking and fixing it. Can model problem solving performance improve if it was trained on examples of mistakes and their corrections? And how does this relate to the inherent difficulty of the problem in question? To investigate this problem, I trained a toy maze-solving model that was allowed to backtrack on its prior moves. I found that the maze solving performance improved when errors and their corrections were injected into the paths of otherwise perfect solutions. I show that the probability of moving in the correct direction improves at decision points when mistakes are allowed. This post also shows that, even though the training data teaches the model to make mist

Training on mistakes improves problem solving performance Training on mistakes improves problem solving performance 2024-04-25 Home Projects Blog Error correction helps problem solving agents, even those which are limited in capability, to solve a much broader range of problems. This is particularly important as the Internet weighs towards finished artifacts (like books, posts, and articles) that have scant few examples of making an erroneous claim and then backtracking and fixing it. Can model problem solving performance improve if it was trained on examples of mistakes and their corrections?

Explore this link on the map →

related reading