MPSelectTune - Prompt-type Selection for Fine-tuning improves Concept Unlearning in LLMs

Sep 21, 2026

Uploaded by Asher

LLMs can be adapted to many tasks via prompts with few-shot examples, but biased or harmful concepts in pretrained models can still cause unsafe outputs across different prompt types. Existing unlearning methods often neglect how prompt variation affects concept removal. This paper proposes MPSelectTune, a two-stage adversarial fine-tuning method using worst prompt-type concept prediction and a multi-task loss to improve unlearning robustness, achieving 2–15% gains on main tasks and up to 17% reduction on worst-case concept accuracy across four benchmarks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

AccessFree to download

Introduction

Related Works

arXiv :2607 .03932v 1 [ cs .LG] 4 Jul 2026 MPSelectTune: Prompt-type Selection for Fine-tuning improves Concept Unlearning in LLMs Shubhadip Nag∗ Srinjoy Das∗ Agniva Saha∗ Anushree Ghosh∗ Soumi Das† Tarun Kumar‡ Suparna Bhattacharya‡ Sourangshu Bhattacharya∗ ∗IIT Kharagpur †MPI-SWS ‡HPE Labs Abstract LLMs can be conveniently adapted to a diverse set of tasks, e.g, prediction, questionanswering tasks, etc, using appropriate prompts with few-shot examples. Biased or harmful concepts, e.g. gender or bio-weapons, present in pre-trained LLMs can lead to unsafe or unethical responses for many such prompts. Removing such undesirable concepts robustly across different prompt types remains a challenging problem, since existing unlearning methods typically ignore the impact of prompt variation. In this paper, we explore a novel adversarial approach to use a joint prompt for the main task and concept task prediction. We show that fine-tuning using the “worst prompt type” for concept prediction (with the highest concept accuracy) improves the average unlearning performance over a fine-tuning method that uses a combination of all prompt types. Our proposed method, MPSelectTune, is a two-stage approach that minimizes the concept accuracy of the highest accuracyprompt type, after fine-tuning using a novel multi-task loss using multiple prompt types. Experimental results on four benchmarks show 2 − 15% main task accuracy improvements over recent baselines and while reducing the worst-case concept accuracy by up to 17% compared to recent baselines. 1 Introduction LLM unlearning Yao et al. [2023] has emerged as an important component of overall LLM safety and compliance objectives in many organizations. The LLM unlearning objective can be broadly divided into two types: (1) information unlearning (IU) Pawelczyk et al. [2024], which erases personally identifiable information from the model, and (2) concept unlearning (CU) Gandikota et al. [2024] . Concept unlearning attempts to erase the effect of a biased or harmful concept (usually in the context of a task) from the LLM, e.g. gender removal in the context of profession prediction De-Arteaga et al.[2019] or toxicity prediction Sahoo et al. [2022], removal of information about biological weapons in the context of scientific question answering Li et al. [2024], etc. The concept to be unlearned is specified as a dataset called the forget set, while retain set Liu et al. [2024a] provides information tobe preserved (related to the main task) in the model. In this paper, we focus on concept unlearning. Concept erasure in the representation learning setup Ravfogel et al. [2022a], Belrose et al. [2024] assumes that the concept can be represented using a linear subspace of the output representation of the features of the examples. However, for LLMs, zero-shot prompting techniques Wei et al. [2022], Kojima et al. [2022], and few-shot prompting techniques involving in-context learning Dong et al.[2024] enable various predictive tasks. In this prompt-based predictive model setup, the representation unlearning techniques are not directly applicable for two reasons: (1) the predictive performance of the model depends strongly on the type of prompt used to elicit concept labels, unlike in the representation learning setup, and (2) the relationship between LLM representations and predictive performance remains unclear. Preprint. MPSelectTune : Prompt Selection for Fine-tuning (a) Flow diagram of the proposed framework showing the main components of each stage. (b) An illustrative example showing that fine-tuning using worst prompt type leads to better concept unlearning and task prediction across multiple prompt types. Figure 1: Overview of the proposed MPSelectTune framework. In this paper, we propose to use joint task and concept prediction prompt types, for unlearning concepts from LLMs. Fig. 1 (Left) shows the flow of our method. We generate multiple jointprediction prompts for ea

FAQs

What problem does MPSelectTune address in LLM unlearning?

It targets concept unlearning robustness across different prompt types, since prior methods largely ignore prompt variation and its effect on concept removal.

How does MPSelectTune work in two stages?

Stage 1 performs multi-prompt tuning with multiple prompt types and a multi-task loss, using a novel format loss to control output formatting. Stage 2 fine-tunes using the worst concept predictor prompt type to minimize the highest-accuracy concept predictor.

What improvements are reported in the experiments?

Across four benchmarks, MPSelectTune improves main-task accuracy by 2–15% over recent baselines and reduces worst-case concept accuracy by up to 17%.

Sep 21, 2026

Uploaded by Asher