Self-Guided Test-Time Training: Let the Model Pick Its Own Evidence Spans for +15% Long-Context Accuracy

arXiv / HuggingFace Daily Papers·medium signal

This paper (arXiv:2607.09415, submitted July 10, trending on HuggingFace today) shows that test-time training for long-context LLMs is highly sensitive to which spans you train on — random spans actually degrade accuracy because most are irrelevant to the query. Self-Guided TTT (S-TTT) has the model first identify relevant evidence passages, then run language-modeling adaptation only on those, yielding up to a 15% relative accuracy gain on LongBench-v2 and LongBench-Pro with Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct. The takeaway for builders: TTT can be practical for long documents if you gate what the model learns from at inference time.