**Lost-in-the-middle** is a behavioral pattern in large language models, first documented by Liu et al. in 2023, where the model's ability to use information in a long context depends strongly on where that information sits in the prompt.
The shape of the dependence is a **U-curve**. Plot retrieval accuracy on the y-axis and the position of the target chunk in the context on the x-axis, and you get a curve that is high at the start, low in the middle, and high again at the end.
The two high regions correspond to **primacy** (the start of the context, where the model attends reliably) and **recency** (the end of the context, where attention is also strong because of how decoder attention naturally weights the most recent tokens).
The low region in the middle is where the model often skips information even when that information is directly relevant to answering the question. This effect persists in 2026 even on frontier long-context models like Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, although the depth of the trough varies.
The practical context-engineering move is to position important content at the head or tail of the retrieved block, not in the middle. The most important chunk goes first, the second-most goes last, and less critical content fills the middle where the U-curve is low.
The **lost-in-the-middle** effect produces a U-shaped accuracy curve when you plot retrieval accuracy against the position of the target document inside the context window. The two ends of the curve are high (primacy at the start, recency at the end) and the middle is significantly lower.
The finding comes from Liu et al. (2023), "Lost in the Middle: How Language Models Use Long Contexts," which evaluated 10 documents at different positions on multi-document QA. Accuracy was around 75% when the target document was first, dropped to around 50% when it was in the middle, and rose back to 70% when it was last. The shape held across models and task variants.
In 2026 the U-shape persists but the trough is shallower on frontier models. RULER measurements on Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro show high accuracy across positions up to roughly 32K context length, with the U-curve becoming more pronounced past that. The takeaway is not that the effect has been solved, but that it kicks in later than it used to.
The mechanism is rooted in attention dynamics. Causal attention naturally weights recent tokens (recency), and the start of context tokens get strong attention because they appear in every attention computation as anchors (primacy). The middle competes for attention against everything else and tends to lose.
The practical move is **head and tail positioning** for retrieved content: place the highest-priority chunks at the start and end of the retrieved block, leave the middle for context that supports rather than answers the query. Adding a question restatement just before the answer-generation cue is a complementary trick that exploits recency directly.
Quick reference
TL;DR
Lost-in-the-middle produces a U-shaped accuracy curve: models attend reliably to the start (primacy) and the end (recency) of a long context but skip content placed in the interior.
Imagine a teacher reading a long list of names and later asking which names were on it. Students remember the first few (they were paying full attention at the start) and the last few (those are still ringing in their ears). The middle names get fuzzy. Language models do something similar with long context: the names at the start and end stick, the names in the middle blur. Plot what they remember against where the name was in the list, and you get a U-shape: high on both ends, low in the middle. The trick for context engineering is to put the important stuff where the U is high, not where the U is low.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Short contexts (under 4K tokens) where position effects are negligible. The U-curve is a long-context phenomenon. For short prompts the effect is small enough that head and tail tricks add complexity without measurable gain.
Tasks where document ordering encodes meaning (sequential narratives, time-ordered logs). Reordering chunks to exploit the U-curve destroys narrative order. For these tasks you must preserve sequence and use other tricks (summarization, retrieval pruning) to fight the middle effect.
Highly-tuned RAG systems where the reranker already produces small candidate sets. If you are passing 3 chunks rather than 30, the middle is small enough that position tricks add little. Save the U-curve engineering for systems that genuinely span the middle.
Quick scan · 2–4 min
Everything important, quickly.
Identify the U-shape. Name primacy and recency. Explain the attention mechanics behind each. Cite the Liu et al. (2023) paper. Note how the curve has shifted on 2026 frontier models without disappearing. Close with the head and tail positioning trick that operationalizes the finding.
Real-world examples
Real products, models, and research that use this idea.
1Liu et al. (2023) Lost in the Middle paper plotted the original U-curve on multi-document QA
2RULER (NVIDIA, 2024) replicated the U-shape across modern long-context models at varying lengths
3BABILong (2024) extended the finding to multi-step reasoning tasks at lengths up to 1M tokens
4Anthropic's long-context prompting guide recommends head and tail placement based on the U-curve
5Greg Kamradt's Needle-in-a-Haystack visualizations made the position vs accuracy heatmap famous and widely reproduced
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat does an inverted U-shape look like, and when do you see it in practice?
Sign in or go Pro to see the approach for this follow-up.
QHas the U-shape changed for frontier models in 2026?
Sign in or go Pro to see the approach for this follow-up.
QHow should the U-curve influence retrieval-set ordering?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Expecting accuracy to decay monotonically with position and missing that the start is also a privileged location, not just the end.