Skip to content
conftrace_
2026 EMNLP EMNLP 2026

AnInter: Anti-Interference Training for Long Context Retrieval

AbstractDense retrievers in Retrieval-Augmented Generation (RAG) have achieved substantial progress in semantic matching, but they remain limited in long context retrieval. In long context retrieval, target evidence often occupies only a small local span, while the remaining content acts as noise. We aim to improve long-document retrieval by enhancing retrievers’ anti-interference capability. Our contributions are threefold: (1) we construct the Anti-Interference Capability Benchmark (AICB) to evaluate retrievers’ anti-interference capability under controlled noise levels; (2) we propose anti-interference training as a short-text training path for improving long context retrieval; and (3) we train AnInter, a series of embedding models that are more robust to noisy long documents. Evaluation on LongEmbed shows that AnInter improves over its backbone by 16.4 points on average and achieves state-of-the-art performance on this challenging long context embedding benchmark. These results show that interference-aware training is a practical direction for improving long context retrieval.

More from EMNLP 2026

All EMNLP 2026 papers →