A user attempted to reproduce the "lost in the middle" phenomenon in current large language models, specifically Anthropic's Haiku 4.5 and Sonnet 5. The user found that with up to 30 documents, the models did not exhibit significant position sensitivity in retrieving a single fact. The only observed variation was a change in the format of the answer (e.g., "3" vs. "three"), which the user did not count as a factual flip. The user acknowledges this is a simplified test case and the phenomenon might still occur with longer documents or more complex tasks. AI
IMPACT Suggests current LLMs may be more robust to position bias in retrieval than previously thought for simple tasks.
RANK_REASON User experiment and commentary on LLM behavior, not a primary release or research paper.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
