Failed attempts pollute the context and bias output in undesirable ways, making it more likely for the LLM to get stuck in a loop of uncorrectable bad responses.