EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering
Researchers have developed EASE-TTT, a novel framework for improving long-context question answering in smaller language models. This method aligns retrieved evidence chunks with attention mechanisms to guide model adaptation. Experiments on six LongBench QA tasks demonstrated EASE-TTT's superior performance compared to existing retrieval and test-time training approaches. AI
IMPACT Enhances the capabilities of smaller language models for complex, long-context question answering tasks.