Internal "Belief" Signals Determine Whether LLMs Will Correct Their Own Errors

A new study on how language models handle their own mistakes reveals a striking gap: large language models (LLMs) can recognize when they are wrong—but they rarely say so unprompted, even if they become aware of errors shortly after making them. The crucial factor determining whether a model will acknowledge a mistake appears to be what researchers call the model's "momentary belief" about the correctness of its output.

How researchers measured belief#

The team developed a method to peer inside model reasoning by training a linear probe on the models' internal representations. This probe was calibrated using external datasets unrelated to error correction, then applied to see whether models would spontaneously retract falsehoods. The key finding: models only retract answers when their internal signals indicate they doubt the accuracy of what they just generated.

Interestingly, these internal beliefs often diverge sharply from what the models appear to "know" in other contexts—such as when answering straightforward factual questions. This suggests that confidence in an answer depends heavily on the moment-to-moment state of a model's computation, not just its learned knowledge base.

Why belief matters more than parametric knowledge#

When a model's internal representation registers skepticism about an answer, several things happen: the model is more likely to attempt further verification, and its attention patterns shift in ways that promote correction. In contrast, models with high internal confidence rarely walk back their statements, even if the answers are verifiable as incorrect.

The researchers demonstrated that manipulating these belief signals could causally drive whether models retract. By steering the model's internal state to register doubt, they could increase retraction behavior; conversely, boosting confidence suppressed it.

Implications for fine-tuning#

When researchers used supervised fine-tuning to improve how often models correct themselves, the underlying mechanism remained the same: the improvement came largely from teaching models to develop more accurate internal beliefs about their own answers. Rather than introducing entirely new pathways for error correction, fine-tuning essentially refined the existing belief-to-retraction link.

The findings suggest that future work aimed at making LLMs more reliable at acknowledging mistakes should focus on calibrating what the model genuinely "believes" during generation, rather than simply training it to produce corrective language.