Anti-BAD Challenge

Welcome to the Anti-Backdoor (Anti-BAD) Challenge, an IEEE SaTML 2026 competition dedicated to advancing the trustworthiness of post-trained large language models (LLMs).

LLMs have become the rising stars of our time, powering a wide range of applications. As training large models from scratch become increasingly costly, many practitioners now rely on post-trained models shared by others. This convenience, however, raises a critical question: can we trust these models to behave safely?

One of the most serious risks is the backdoor attack, where a model behaves normally on clean inputs but produces malicious outputs when a hidden trigger appears. The Anti-BAD Challenge invites participants to design effective and practical defense methods that can mitigate such hidden behaviors without access to the training history or backdoor knowledge, while maintaining strong task performance.

News

October 21, 2025: Competition opens for registration on Codabench.
August 31, 2025: Anti-BAD Challenge accepted as part of IEEE SaTML 2026.
Join our Discord community for announcements, discussions, and Q&A.

Overview

The Anti-BAD Challenge simulates a realistic setting where users download post-trained models from public sources without knowing their training history. The goal is to produce a deployable model that preserves utility while suppressing malicious backdoor behavior during inference.

The challenge includes three tracks, each representing a distinct application scenario:

Generation Track
Classification Track
Multilingual Track

Each track contains two tasks, for a total of six. Every task provides multiple post-trained models that reflect real-world model-sharing practices. Participants may defend individual models or develop approaches that integrate information from several models to improve robustness.
The models are based on popular architectures such as Llama and Qwen.

Detailed setup instructions and starter kit information are provided in the Getting Started page. For full task descriptions, evaluation details, and submission guidelines, visit the Challenge page.