Anti-BAD Challenge

Welcome to the Anti-BAD Challenge, an official IEEE SaTML 2026 competition.

This competition addresses growing concerns around backdoor attacks in Large Language Models (LLMs), especially when models are shared in post-trained form without transparency about their training process. Anti-BAD invites participants to explore effective defenses that restore model integrity while preserving utility—under practical constraints such as limited data access or trigger knowledge.

News

• August 31, 2025: Anti-BAD Challenge accepted and initiated for IEEE SaTML 2026.

Overview

The Anti-BAD Challenge provides a benchmark for evaluating post-training backdoor defenses across three representative tasks:

Generation Track – instruction-following models
Classification Track – standard classification tasks
Multilingual Track – cross-lingual generalization and robustness

Participants are encouraged to submit defenses that:

Remove or mitigate backdoor behavior in post-trained LLMs
Preserve model utility on clean inputs
Operate under deployment constraints, without access to training data or large-scale retraining

We welcome: