AI
LLMs fail at self-correction not due to reasoning limits, but how errors are framed
A new study shows that relabeling errors from internal thoughts to external sources dramatically improves large language models' ability to spot and fix their own mistakes.
Large language models struggle to catch their own errors despite being adept at fixing mistakes pointed out by others. But new research suggests this may not reflect a fundamental cognitive shortcoming—instead, it appears to be rooted in how those errors are presented in the chat template.
The role-relabeling experiment#
Researchers designed a training-free intervention that keeps an erroneous claim completely unchanged while varying only its source label. They tested the same incorrect statement presented in four different ways: as the model's own internal thought, as a message from the user, as a tool response, or as a system memory block. Across 12 different model-domain combinations—including both proprietary APIs and open-source models ranging down from 70-billion-parameter families—a striking pattern emerged.
When researchers relabeled an error from an internal thought to an external role, the rate at which models explicitly corrected it jumped by 23 to 93 percentage points. The effect was statistically significant in 10 of the 12 test scenarios.
Domain-dependent relabeling strategies#
The most effective way to frame an error proved context-dependent. In mathematical reasoning tasks, marking the error as a system memory block worked best. For logical deduction problems, presenting it as a user message produced the strongest results. This variance suggests that models handle different types of information differently depending on its source label.
Implications for model training#
The findings reframe how researchers should think about the self-correction gap in large language models. Rather than pointing to a capability deficit—where models simply cannot detect their own errors—the evidence indicates that instruction tuning and chat template design play a far more direct role. How a claim is labeled in the message structure acts as a key experimental variable that can substantially alter model behavior, opening a more straightforward path to improving self-correction without fundamental architectural changes.
