New Training Method Boosts LLM Consistency Across Rephrased Prompts

Large language models notoriously struggle with a fundamental problem: they often give wildly different answers depending on how you phrase the question, even when the underlying query is identical. Researchers have now developed a training approach designed to tackle this fragility.

The Problem with Prompt Sensitivity#

Prompt perturbations—subtle rewording of the same question—can cause LLMs to flip between correct and incorrect responses, or to shift their outputs significantly. This inconsistency undermines trust in the models and limits their real-world deployment, especially where reliability matters. Until now, most solutions have required labeled data, which is expensive and time-consuming to produce.

How Flip-Flop Consistency Works#

The new method, called Flip-Flop Consistency ($F^2C$), operates without any manual labeling. It works by:

• Consensus Cross-Entropy (CCE): The system generates multiple variations of the same prompt and tallies which answer the majority of them produce. This majority vote becomes a pseudo-label—a stand-in ground truth—that guides training.
• Representation alignment: A secondary loss function pushes the model's internal representations toward agreement, pulling low-confidence and minority predictions into alignment with the high-confidence majority consensus.

Experimental Results#

Testing across 11 datasets covering four natural language processing tasks, each with 4–15 prompt variations, $F^2C$ demonstrated measurable gains:

• Agreement between different prompt phrasings improved by 11.62% on average
• Mean $F_1$ scores (a standard measure of accuracy) climbed by 8.94%
• Performance variance across different phrasings dropped by 3.29%

The method also generalized well to unseen domains and held up even when trained on only a subset of prompt variations and tested on completely new phrasings.

Why This Matters#

By improving consistency without requiring human annotation, $F^2C$ addresses a genuine pain point in making LLMs production-ready. The unsupervised nature of the approach also makes it practical to apply to new domains where labeled data is scarce. Code for the method is publicly available for researchers to build on.