---
language: en
title: "Position bias"
description: "Position bias describes how the location of information in the context can change how a model uses it. Learn about the test and the limits of this evidence."
canonical: "https://tropk.ai/en/resources/glossary/position-bias"
last_modified: "2026-10-12T16:13:03.595Z"
---

# Position bias

Reviewed on 2026-10-06.

## What it means

Position bias is the influence of the location of information in the context on a language model’s response. In tasks studied in the article Lost in the Middle, relevant documents at the beginning or end were used more effectively than documents in the middle. The content remained the same; the order changed the observed performance.

## Why it matters for marketing

For marketing, this concept helps evaluate assistants that consult catalogs, commercial policies, and campaign materials. A forgotten condition may have been retrieved correctly but used improperly during generation. Before attributing the failure to missing documentation, it is worth investigating how the passages were organized in the model’s input.

## Example in practice

Imagine a team building an assistant to consult shipping rules. The document limiting free delivery to capital cities appears among nineteen other passages. In a hypothetical test, the response ignores the restriction when the document is in the middle and includes it when it is first. The team decides to investigate how the context is assembled.

## How to apply or verify

To verify, keep the question, documents, and configuration constant and change only the position of the required passage. Record responses, model version, context size, and whether the commercial condition is handled correctly. Repeat the runs to distinguish the influence of order from natural variability. The technical lead should also check that the entire passage reached generation.

## Important distinctions

This bias concerns position in the model’s input; it does not prove a page-ranking rule or guarantee citations for texts with answers at the beginning. There is also order bias in comparisons made by evaluator models. The 2023 study’s results depend on the tasks and models examined, so testing in the current implementation is required.

## Sources and references

- [Lost in the Middle: How Language Models Use Long Contexts](<https://arxiv.org/html/2307.03172v3>) — Liu e colaboradores, arXiv
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](<https://arxiv.org/html/2306.05685v4>) — Zheng e colaboradores, arXiv

## Continue a leitura

- [Context window](</en/resources/glossary/context-window>)
- [LLM as a judge](</en/resources/glossary/llm-as-a-judge>)
- [Passage retrieval](</en/resources/glossary/passage-retrieval>)

## About the author

[Vinicius Pacheco](</en/team/vinicius-pacheco>)

Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content.

---

[All pages in Markdown](<https://tropk.ai/sitemap.md>)