Skip to main content
Back

Trinity College's Bayesian-LoRA Cuts Fine-Tuned AI Overconfidence by 84%

A new probabilistic take on LoRA cuts calibration error by up to 84% and log-likelihood by 76% for less than 1M extra parameters.

Paper, Full Access
  • Bayesian-LoRA reformulates LoRA as a probabilistic update inspired by Sparse Gaussian Processes with a normalizing-flow posterior.
  • Deterministic LoRA emerges as a limiting case when the posterior collapses to a point mass.
  • Reports up to 84% ECE reduction and 76% NLL reduction versus standard LoRA fine-tuning.
  • Adds only ~0.42M parameters and ~1.2x training cost; inference can merge to zero-latency mode.
  • Evaluated on LLaMA 2-7B, Qwen2.5-14B, and Qwen3-30B-A3B across reasoning, WikiText-2, and MATH.
  • Best gains appear under distribution shift; post-hoc Laplace still wins for small in-distribution shifts.

Fine-tuned language models have a bad habit: they claim confidence they have not earned. That overconfidence is a systemic calibration problem, and it worsens when you adapt a base model to a small task-specific dataset with LoRA. A team from Trinity College Dublin and University College Dublin has proposed a fix that keeps LoRA's efficiency while giving it a proper sense of doubt.

Their method, called Bayesian-LoRA, reframes the standard low-rank update as a probabilistic object drawn from a Sparse Gaussian Process. The result is a drop-in replacement for LoRA that reduces expected calibration error (ECE) and negative log-likelihood (NLL) without sacrificing accuracy, for both in-distribution and out-of-distribution evaluations, at a cost of roughly 0.42M extra parameters and about 1.2x the training time.

Why fine-tuned LLMs forget what they don't know

Pre-trained LLMs emerge reasonably well calibrated, meaning their predicted probabilities roughly line up with how often they are right. Fine-tuning them for specific domains often degrades that calibration substantially, especially on small datasets, and produces overconfident predictions. That is a serious problem for anything downstream that consumes those probabilities, from medical triage to autonomous driving to routing between models.

The obvious answer is Bayesian inference over model weights, but as the authors note, doing that at LLM scale is prohibitively expensive. Laplace approximations need Hessians that scale with model size. Bayes-by-Backprop style methods (like BLoB) place mean-field Gaussians over the full LoRA parameter space. Deep ensembles multiply memory and inference costs. None of these approaches is cheap enough to be a default for a 30B-parameter model.