When multiple sources are concatenated into one long input, the model processes beginning and end reliably but loses information from the middle. This is the lost-in-the-middle effect — a U-shaped attention curve that drops to 52% incorporation for middle-positioned content while beginning and end stay above 94%.
The Data
Finding incorporation rates across 200 runs with 10 sources each:
| Position | Incorporation rate |
|---|---|
| Sources 1-2 (beginning) | 96% |
| Sources 3-4 | 78% |
| Sources 5-6 (middle) | 52% |
| Sources 7-8 | 81% |
| Sources 9-10 (end) | 94% |
The pattern is consistent, not random. The same finding placed at different positions shows different incorporation rates. Position determines reliability, not content quality.
Why It Happens
This is an inherent attention distribution pattern in how models process long sequences. The model attends more strongly to the beginning (primacy) and end (recency) of its input. Middle content receives weaker attention.
This is not a capacity problem. Larger context windows do not eliminate the effect — they still show the U-shaped curve. The pattern is about attention distribution, not context size.
Mitigation Strategies
Three techniques, each additive:
| Strategy | Middle source incorporation |
|---|---|
| Raw concatenation | 51% |
| Section headers added | 72% (+21%) |
| Headers + beginning summary | 91% (+19%) |
1. Beginning summary. Place a summary of key findings from all sources at the top of the input. Critical information from middle sources appears in the highest-attention position. The model processes the summary reliably, then has context for the detailed sections.
2. Section headers. Explicit markers like ”## Module 4: Authentication Results” help the model maintain structured attention across the input. Headers create attention anchors that partially override the U-shaped curve.
3. Source tags and cross-reference markers. When findings from different sources relate to each other, tag them explicitly and note the relationships in the beginning summary. “Tool 3 and Tool 11 findings are related” in the summary ensures the connection is processed even when the detail sections are in middle positions.
What Does Not Work
“Process all sections equally” instructions. Attention distribution cannot be overridden by prompt directives. The lost-in-the-middle effect is a fundamental processing pattern, not a behavioral choice.
Random shuffling. Shuffling source order each run distributes the risk probabilistically but does not prevent loss in any single run. A beginning summary provides deterministic protection.
Larger context windows. The U-shape persists regardless of window size. A 200K-token window still shows weaker middle attention than a 100K-token window.
When to Process Separately
For 12+ sources where all findings must be captured and cross-referencing is not needed, processing each source in an individual API call eliminates the attention effect entirely. But this multiplies API costs and loses cross-source context.
The structured mitigation approach (summary + headers + tags) is more cost-effective for most use cases — it raises middle incorporation from 51% to 91% without architectural changes.
When cross-referencing IS needed (tool 3’s finding relates to tool 11’s), the sources must be in the same input. Separate processing prevents correlation. Use the beginning summary to highlight cross-references explicitly.
One-liner: Place a key findings summary at the beginning and add section headers throughout — this raises middle-source incorporation from 51% to 91%, mitigating the U-shaped attention curve without splitting into separate API calls.