LLMs fail to spot their own mistakes even when they can spot identical errors from users

Large language models have become remarkably capable, yet they still produce errors and pursue flawed reasoning. Self-correction—the ability to recognize and fix mistakes—is critical for deploying these systems in high-stakes domains, but a new study reveals a paradoxical gap in how models handle their own errors versus those of others.

The Self-Correction Blind Spot#

Researchers from arXiv introduced Self-Correction Bench, a controlled testing framework that exposes a fundamental weakness in current language models. The team injected identical errors into prompts in two ways: either framing them as something a user wrote (external error) or as something the model itself generated (internal error). Everything else in the context stayed the same.

Testing 14 open-source models revealed striking results: models corrected the external errors 64.5% more often than they corrected identical mistakes in their own outputs. This wasn't a knowledge gap—the models could fix the errors when attributed to users. The problem is that the capability remains dormant when applied to self-review.

Root Cause: Training Data#

The researchers traced the blind spot to how models are fine-tuned. Supervised fine-tuning datasets typically contain examples of correcting others' errors, but lack sequences where models correct mistakes they themselves made. This imbalance leaves a critical gap in training.

The solution proved surprisingly effective: fine-tuning with just 5,306 error-correction traces where the model fixes its own output reduced the blind spot by 76%. This demonstrates that the correction ability is trainable once the training data aligns with the task.

Mechanical Insights and Quick Fixes#

Beyond data composition, the researchers identified a conversational-role direction embedded in the model's internal representation space—a pathway that gates whether self-correction activates. Manipulating this direction causally affects the model's ability to self-correct.

Even simpler approaches showed promise. Appending a single word—"Wait"—before asking the model to review its output cut the blind spot by 89.3%, with no additional training required. Notably, this "Wait" mechanism operates through a largely independent pathway from the conversational-role direction, suggesting that self-correction is governed by multiple, separable mechanisms rather than a single bottleneck.