Report
Forgetting proposed as a fix for coding agents that stop improving early in autoresearch
The paper’s author says coding agents can run for days on one problem but sometimes stop improving early.
TLDR
The author introducing “Scaling Discovery Requires Forgetting” says autoresearch with coding agents can run for days on one problem, but they have seen an increasing number of cases where the agents stop improving early. The paper proposes forgetting to address the issue. A researcher commenting on the work says long runs can lock agents into taking small steps in the same direction.
Combined views
904
2 Sources, first seen ago
