Self-Guided Test-Time Training for Long-Context LLM Utilization

This method addresses long-context accuracy degradation by treating the test context as an instance-specific training example. It avoids the noise of random span sampling to improve the model's ability to identify relevant evidence within massive inputs.