LoRA-BAM: Input Filtering for Fine-tuned LLMs via Boxed Abstraction Monitors over LoRA Layers

2025/06/01 by Changshun Wu, Wu, Changshun, Tianyi Duan +5
Computer Science · Engineering · #Advanced Adaptive Filtering Techniques #Advanced Wireless Communication Techniques #Error Correcting Code Techniques #FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · pdf · doi:10.48550/arxiv.2506.00998

openalex publication_date 2025/06/01 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Fine-tuning large language models (LLMs) improves performance on domain-specific tasks but can lead to overfitting, making them unreliable on out-of-distribution (OoD) queries. We propose LoRA-BAM - a method that adds OoD detection monitors to the LoRA layer using boxed abstraction to filter questions beyond the model's competence. Feature vectors from the fine-tuning data are extracted via the LLM and clustered. Clusters are enclosed in boxes; a question is flagged as OoD if its feature vector falls outside all boxes. To improve interpretability and robustness, we introduce a regularization loss during fine-tuning that encourages paraphrased questions to stay close in the feature space, and the enlargement of the decision boundary is based on the feature variance within a cluster. Our method complements existing defenses by providing lightweight and interpretable OoD detection.

Citations

• Out-of-Distribution Detection with Prototypical Outlier Proxy
• BAM: Box Abstraction Monitors for Real-time OoD Detection in Object Detection
• Retrieval-Augmented Generation for Large Language Models: A Survey
• A Survey of Confidence Estimation and Calibration in Large Language Models
• Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement
• Detection of out-of-distribution samples using binary neuron activation patterns
• Out-of-Distribution Detection with Deep Nearest Neighbors
• VOS: Learning What You Don't Know by Virtual Outlier Synthesis
• Measuring Massive Multitask Language Understanding
• Unsupervised K-Means Clustering Algorithm
• A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks
• Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks

Related

• Your Finetuned Large Language Model is Already a Powerful Out-of-distribution Detector
• LoraRetriever: Input-Aware LoRA Retrieval and Composition for Mixed Tasks in the Wild
• LoRA ensembles for large language model fine-tuning
• Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment
• DLP-LoRA: Efficient Task-Specific LoRA Fusion with a Dynamic, Lightweight Plugin for Large Language Models