Model Merging Resolves the Alignment-Calibration Trade-off Without Sacrificing Performance

The process of aligning large language models to behave safely and follow human preferences—known as post-training alignment—comes with a well-documented cost: degradation in task performance. But according to a new paper submitted to arXiv, this hidden "alignment tax" extends beyond mere accuracy drops. Alignment training also severely compromises model calibration, making systems produce overconfident predictions, deliver less reliable outputs, and generate less diverse responses.

The Hidden Cost of Alignment#

Researchers found that after alignment training, models systematically become poorly calibrated: their confidence levels don't match actual correctness, they admit uncertainty less often, and their outputs show reduced variety. This reliability problem compounds the accuracy penalty already known from the alignment tax, creating a two-fold penalty on model quality.

A Simple Weight-Interpolation Solution#

Rather than accept this trade-off as inevitable, the team tested a straightforward post-hoc fix: blending model weights between their pre-alignment and post-alignment states. By interpolating along this spectrum—rather than choosing one endpoint or the other—they identified sweet spots that outperform both versions. These Pareto-optimal merged models simultaneously boost accuracy beyond the original aligned model while recovering much of the calibration lost during training.

Practical Advantages#

The approach offers significant computational efficiency: it requires no retraining, no new data, and no architectural changes. It simply mixes the weight matrices of two existing checkpoints in different proportions, making it trivially cheap to apply at scale compared to full retraining or other mitigation strategies.

Implications#

This work reframes the alignment trade-off not as immutable but as navigable through model merging, suggesting that achieving both capability and reliability may be more tractable than previously assumed.