LLM-VA: Large Language Model Vector Alignment

This repository contains the code for the paper "LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment" (ACL 2026 Main Conference).

Setup

conda create -n llmva python=3.12.8 -y
conda activate llmva

pip install -r requirements.txt

"flash-attn==2.8.2" needs to be installed separately in Flash-Attn.

Usage

Setup server:

python src/server_answer.py

In another terminal, run the client (Use CUDA_VISIBLE_DEVICES to specify which GPUs to use):

python src/run/llmva_run.py