ModelRefs / Recency Bias — AI Glossary

Recency Bias — AI Glossary

A model's tendency to weight information appearing late in the context more heavily than earlier content.

Overview

Recency bias is a manifestation of positional bias in transformer attention—later tokens receive stronger gradient signal in autoregressive training. Affects multi-turn conversations (overweighting recent turns), RAG (favoring last-inserted chunks), and long-document summarization. Addressed by strategic context ordering and multi-position ensemble prompting.

Reference details

Topic	prompting
Last reviewed	2026-06-24

How it works

Autoregressive training makes every position predict its successor, so nearby context is the strongest predictor and the model learns to weight it heavily.
At inference that shows up as later context exerting more influence than earlier context of equal importance.
In a long prompt an instruction near the end is therefore followed more reliably than the same instruction at the start.
In conversation it means recent turns dominate, so a constraint set twenty turns ago competes poorly with the last thing said.
The mitigation is placement and repetition: put what must hold at the end, and restate it rather than assuming it persists.

Example: Why restating an instruction works

A 20,000-token prompt opens with a formatting constraint, then 19,000 tokens of documents, then the question. The constraint is 19,000 tokens from the generation point and is frequently dropped. Moving it to sit immediately before the question costs nothing — the same tokens, a different position — and compliance rises. Repeating it at both ends costs perhaps 30 tokens. This is the cheapest reliability fix available in prompt design, and it is an artefact of position rather than of the instruction's wording.

Commonly confused with

One particular positional bias: late content outweighs early content. That is why an instruction repeated at the end of a long prompt is often obeyed when the same instruction at the top is not, and why appending retrieved context after the question tends to beat prepending it. It is not the same as lost-in-the-middle, which says the middle loses to *both* ends.

When to use it

Reach for it when:

Placing constraints in a long prompt, where the end is the strongest position.
Diagnosing why an early instruction is ignored while the same instruction later is obeyed.
Designing conversation history handling, where old constraints fade against recent turns.

Reach for something else when:

Short prompts, where there is little positional distance for it to act over.
As a synonym for positional bias, which is the general effect this is one case of.
Explaining a wrong answer with no positional variation tested.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Recency Bias — AI Glossary.

Frequently asked questions

What is Recency Bias?