Study: LLMs Face Attention Dilution and Declining Retrieval in Million-Token Contexts
_reachsumit · x · 2026-07-03
The paper "Can Language Models Actually Retrieve In-Context?" investigates the retrieval capabilities of LLMs within million-token long texts. The study reveals that while LLMs can retrieve information from ultra-long contexts, an "attention dilution" phenomenon occurs as document volume increases, leading to a significant drop in retrieval performance. To address this, the researchers proposed a length-aware fix, offering new optimization strategies for long-context applications.
More from Research
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- New Blog Breaks Black-Box Distillation Into Five Practical Steps — chrisrohlf · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- OPST paper shows LLMs self-improve with zero human data, guided only by topic descriptions — _akhaliq · 2026-10-26
- Tencent's HiSTEVE 1.0: hierarchical VLA agent plays Minecraft with native keyboard and mouse, hits 85% SOTA — _akhaliq · 2026-10-26