Science ❯ Computer Science ❯ Artificial Intelligence

Model Training

Reinforcement Learning Data Poisoning Deep Learning Parameter-efficient Fine-tuning Learning Behaviors Synthetic Data Generation Ethical Considerations Simulation Reasoning Traces Data Annotation Conversational Agents Data Challenges

Anthropic-Led Study Finds About 250 Poisoned Documents Can Backdoor LLMs Regardless of Size

Routine safety training can largely neutralize such simple backdoors.