What is Recency Bias in AI models?

Recency bias is the tendency of decoder-only transformer models to place the strongest attention weights on the tokens closest to where generation begins—at the very end of the prompt.

Why do AI models forget instructions placed at the top of a long prompt?

When 3,000+ words of text sit between your initial instructions and the generation point, the model's self-attention degrades (often called the 'Lost in the Middle' phenomenon).

Where should you place critical formatting rules in a long prompt?

Always place your background text or documents first, and anchor your final output rules and schemas at the very bottom, right before the model starts typing.

Recency Bias: Putting Instructions at the Bottom

Have you ever pasted a 10-page contract or article into an AI, written "Output as 3 bullet points" at the very top, and then watched the AI produce 8 rambling paragraphs?

It's infuriating. You feel like the AI just ignored you. But the AI didn't ignore you—it suffered from Recency Bias.

The Transformer Attention Budget

Large language models predict tokens from left to right. As text grows longer, the model must distribute its mathematical attention across every single word. Empirical research (notably Liu et al.'s 'Lost in the Middle') proves that attention is not uniform:

🧠 THE ATTENTION CURVE

Attention is strongest at the very beginning (Primacy) and the very end (Recency). Instructions buried in the middle or separated from the finish line by thousands of words suffer severe degradation.

The Golden Layout for Long Prompts

Top: Role & overall objective.
Middle: Raw documents, background facts, or data (inside boundary tags).
Bottom (The Anchor): Exact formatting rules, schema, and direct output commands.

Recency Bias: Putting Instructions at the Bottom

The Transformer Attention Budget

The Golden Layout for Long Prompts

Topic 04 Scenario Challenge.