# Position Bias

> Position bias is a language model's tendency to use information at the start or end of its context more reliably than the middle (lost in the middle).

Source: https://metavert.io/position-bias  
Published: 2026-10-07  
Updated: 2026-10-07

**Position bias** is the tendency of a language model to use information differently depending on where it sits in the input, most famously the "lost in the middle" effect, in which material at the beginning or end of a long context is used more reliably than material in the middle. For anyone whose content reaches a model through retrieval, it means that being retrieved is not the same as being used: where a passage lands in the context, and where the answer sits inside the passage, both affect whether it shapes the response.

### The Original Finding

The effect was documented by Liu and colleagues in "Lost in the Middle: How Language Models Use Long Contexts" (2023, published in TACL). Testing multi-document question answering and key-value retrieval, they found that "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models." Plotted against position, accuracy forms a U-shaped curve. The result reframed long [context windows](https://metavert.io/context-windows): a model that can accept many documents does not necessarily attend to all of them equally.

### 2026: Attenuating, Not Gone

Newer models are more robust, but the picture is conditional. A May 2026 study (Zhang et al., "Positional Failures in Long-Context LLMs") evaluated nine models on reasoning tasks placed at controlled positions in 8K, 32K and 64K-token contexts with two kinds of filler. Under one filler type at 64K, three of the four newest releases stayed within ±6 percentage points of their end-position accuracy. Under the other filler type, middle-position drops persisted across all four, ranging from 16 to 56 points. The authors also found that 76% of middle-position errors matched an answer belonging to the surrounding filler, against 22% at the end position, which suggests the failure is partly interference from neighbouring content and not simple forgetting.

Small-model evidence agrees in direction. NanoKnow (SIGIR 2026) observed a lost-in-the-middle effect when the answer document was placed between distractors, and found that irrelevant context harms accuracy according to both its position and its quantity. The honest summary is that position bias is model-dependent and shrinking in the best systems under favourable conditions, while remaining easy to reproduce when the surrounding context is distracting.

### Why Retrieval Rank Carries Through

In [retrieval-augmented generation](https://metavert.io/rag) the context is assembled from ranked search results, so rank and position are linked. C-SEO Bench (NeurIPS Datasets and Benchmarks 2025) measured this directly: "making the target document the first one in the LLM context window leads to far greater citation ranking gains in the LLM response than any C-SEO method." In its tests with one model, moving a document to first position improved its rank in the answer by 1.6 to 2.8 places depending on the product category, with the benefit fading by the third position. The study's broader conclusion was that conventional ranking improvements outperform content rewriting aimed at the model.

One caveat: production engines do not publish how they order retrieved passages, how many they include, or whether they reorder them before generation. The link between a page's search rank and its position in a commercial engine's context is a reasonable inference from these experiments, not a documented behaviour. It is nonetheless consistent with the separate finding that being cited and actually influencing the answer are different outcomes, a distinction covered under [AI citations](https://metavert.io/ai-citations).

### Implications for Content Structure

Retrieval systems generally work on passages, not whole pages, so position bias operates at two scales. Across documents, a page competes for an early slot, which is a ranking problem and, with [query fan-out](https://metavert.io/query-fan-out), a problem of ranking for many sub-queries at once. Within a document, the practical response is to write sections that survive being lifted out alone: state the answer in the first sentence under a heading, keep the definition and the key figure together, and avoid burying the substantive claim after several paragraphs of setup. A passage that leads with its conclusion is robust whether it lands first, last or in the middle.

These are design inferences from the research and not tested optimization tactics; no study reviewed here shows that restructuring a page for position produces a measurable gain in a live engine. They cost little, and they also happen to describe clear writing.

## Related Topics

- [Context Windows](https://metavert.io/context-windows) — The space in which position effects occur
- [Retrieval-Augmented Generation](https://metavert.io/rag) — How ranked results become ordered context
- [Attention Mechanism](https://metavert.io/attention-mechanism) — The model component that weighs one position against another
- [Context Engineering](https://metavert.io/context-engineering) — The practice of arranging what a model sees
- [Query Fan-Out](https://metavert.io/query-fan-out) — Why a page must rank for many sub-queries
- [AI Citations](https://metavert.io/ai-citations) — Being selected versus shaping the answer
- [Training Data Frequency](https://metavert.io/training-data-frequency) — The other route by which content reaches an answer
- [Generative Engine Optimization](https://metavert.io/generative-engine-optimization) — Where these findings are applied

## Further Reading

- [Lost in the Middle: How Language Models Use Long Contexts](https://arxiv.org/abs/2307.03172) — Liu et al., TACL, 2023
- [Positional Failures in Long-Context LLMs: A Blind Spot in Reasoning Benchmarks](https://arxiv.org/abs/2605.23170) — Zhang et al., arXiv, May 2026
- [NanoKnow: How to Know What Your Language Model Knows](https://arxiv.org/abs/2602.20122) — Gu, Jedidi and Lin, SIGIR 2026
- [C-SEO Bench: Does Conversational SEO Work?](https://arxiv.org/abs/2506.11097) — Puerto et al., NeurIPS Datasets and Benchmarks 2025
- [Citation selection versus absorption in AI search answers](https://arxiv.org/abs/2604.25707) — Zhang, He and Yao, arXiv, April 2026