QR-LoRA: QR-Based Low-Rank Adaptation for Efficient Fine-Tuning of Large Language Models

The growing scale of Large Language Models (LLMs) has necessitated the development of parameter-efficient fine-tuning techniques. Low-Rank Adaptation (LoRA) has emerged as a promising approach, reducing the number of trainable parameters by applying low-rank updates to pretrained weights. While standard LoRA learns both update factors directly, several recent variants first initialize those matrices via an SVD of the pretrained weights -- an operation that can be expensive on large models and yields singular vectors that are not always easy to interpret. In this work, we extract an orthonormal basis from the pretrained weight matrix using QR decomposition with column pivoting, and then express the LoRA update as a linear combination of these basis vectors -- training only the scalar coefficients, which imposes clear structure on adaptation and drastically reduces parameter count. Experiments across GLUE tasks show that QR-LoRA matches or exceeds the performance of full fine-tuning, standard LoRA, and SVD-LoRA (LoRA with update matrices initialized via singular value decomposition) with as few as 601 parameters -- a reduction of over 1000x compared to full fine-tuning and 77x fewer than typical LoRA setups.

Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models

Language models serve as proxies for human preference judgements in alignment and evaluation, yet they exhibit systematic miscalibration, prioritizing superficial patterns over substantive qualities. This bias manifests as overreliance on features like length, structure, and style, leading to issues like reward hacking and unreliable evaluations. Evidence suggests these biases originate in artifacts in human training data. In this work, we systematically investigate the relationship between training data biases and preference model miscalibration across five idiosyncratic features of language model generations: length, structure, jargon, sycophancy and vagueness. Using controlled counterfactual pairs, we first quantify the extent to which preference models favor responses with magnified biases (skew), finding this preference occurs in >60% of instances, and model preferences show high miscalibration (~40%) compared to human preferences. Notably, bias features only show mild negative correlations to human preference labels (mean r_human = -0.12) but show moderately strong positive correlations with labels from a strong reward model (mean r_model = +0.36), suggesting that models may overrely on spurious cues. To mitigate these issues, we propose a simple post-training method based on counterfactual data augmentation (CDA) using synthesized contrastive examples. Finetuning models with CDA reduces average miscalibration from 39.4% to 32.5% and average absolute skew difference from 20.5% to 10.0%, while maintaining overall RewardBench performance, showing that targeted debiasing is effective for building reliable preference models.