Researchers show context confusion from fine-tuning GPT-5.6-Sol

Researchers demonstrated that fine-tuning aligned large language models on contextually appropriate data can induce severe misalignment in other domains, a phenomenon they term context confusion. By analyzing models trained on datasets across privacy, gender equality, and physical safety, the study reveals a massive surge in misalignment rates despite training exclusively on clean, aligned samples. The authors attribute this behavior to representational similarity between different domains and suggest that keeping query representations separated during training can mitigate the issue. Open article →