Lost In The Middle

Definition

A position effect where a model's attention is strongest at the start and end of the context, so information buried in the middle of a long context is recalled less reliably.

Key points

• Attention is strongest at the beginning and end of the context; the middle degrades ("lost in the middle").
• Manifests as degraded quality deep in a conversation — attention dilution over a long window.
• Fix: keep critical constraints in the system prompt (a high-attention position), not buried mid-conversation.
• Distinct from Context Rot (total-volume degradation): this is about token position, not total count.
• Progressive summarisation compounds the problem by pushing detail loss into the middle.

Why it matters for the exam

• The diagnosis for "deep-context decay" in the context-diagnosis map: degraded quality deep in a conversation → lost in the middle.

Common gotchas

• Diagnosis map: inconsistent answers → system prompt (missing/incomplete); misrouting → tool descriptions; deep-context decay → lost in the middle. Don't confuse the three.
• Putting must-follow constraints at the end of a long transcript instead of the system prompt.

See also

• Context Rot
• Progressive Summarisation Risk
• Scratchpad Pattern
• Effective Tool Descriptions

Sources

• netlight-study-guide
• certsafari