Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Language models can have a large context window and still fail to use information placed inside it reliably. Researchers call this position sensitivity the “lost in the middle” effect: in tested settings, models often performed better when relevant material appeared near the beginning or end of a prompt than when it appeared in the middle. That is a measured retrieval-and-use pattern, not evidence that models literally forget like people.
What “lost in the middle” means
The phrase describes a performance change when researchers keep a task and its relevant evidence comparable but move that evidence to different positions in the input. In their 2024 paper, Liu and coauthors found a U-shaped performance pattern in the settings they tested: results tended to be stronger when the answer-bearing information was near the start or end, and weaker when it was in the middle. They call the beginning and end advantages “primacy” and “recency” bias, respectively.
As an Amazon Associate I earn from qualifying purchases.
The distinction matters because a context window describes how much text a model can accept, not how consistently it can locate, combine, and use every detail in that text. The study measures performance on specific tasks and models; it does not establish one failure rate for all long-context systems. Read Liu et al.’s 2024 paper in Transactions of the Association for Computational Linguistics.
What the original study tested
Liu et al. examined multi-document question answering and synthetic key-value retrieval. In the question-answering setup, they varied the context length and the position of the document containing the answer. The key-value task offered a controlled way to test whether a model could find a relevant item within an input. Together, these evaluations show why long-context capability should be assessed by more than whether a model accepts a long prompt.
#1 Best Overall
One reported comparison is easy to misread: GPT-3.5-Turbo’s closed-book multi-document QA comparison score was 56.1% in Liu et al.’s 2024 study. The authors report that the model performed worse when relevant information was in the middle than when answering without documents. The 56.1% figure is the closed-book comparison score—not the model’s middle-position score, a general accuracy rate, or a result for every model and task.
Why this matters for retrieval and long prompts
In retrieval-augmented generation (RAG), a system supplies documents as context and asks a model to answer using them. That resembles the multi-document QA setting: success depends not only on retrieving useful evidence, but also on the model finding and using it once it is in the prompt. The original study is relevant to this problem, but it was not a head-to-head evaluation of every commercial RAG system.
A practical implication is to treat prompt organization and evidence placement as things to measure, not assume. A result from a short prompt, a single relevant passage, or one placement cannot by itself show how well a system handles longer inputs or multiple pieces of evidence.
What later studies add
A proposed calibration method
In 2024, Hsieh and coauthors described a calibration approach intended to address positional attention bias. Their Google Research publication reports improvements in information location and RAG performance across evaluated tasks, with gains of up to 15 percentage points over compared methods. “Up to” is the maximum reported in that evaluation, not a promised improvement for another model, dataset, or production workload. See the 2024 Google Research publication.
Multiple relevant passages change the test
A 2025 ACL Findings paper by Tian and coauthors complicates a simple one-passage account. Its LongPiBench evaluation examines multiple relevant passages, varied tasks, and input lengths. The authors report that most of the commercial and open-source models they tested were more robust to the classic lost-in-the-middle issue, while the spacing between relevant information pieces still caused noticeable bias. In other words, a model may handle one item’s absolute position better yet struggle when it must connect evidence separated across a long context. The authors conclude: “These findings highlight the importance of evaluating and reducing positional biases for long-context LLMs.” Read the LongPiBench paper in Findings of ACL 2025.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to check a model or application for position bias
For a useful evaluation, vary the evidence position while keeping the question, relevant content, and scoring method as consistent as possible. Then test whether the result holds when the task requires combining evidence, rather than retrieving one isolated fact.
- Choose representative tasks. Include the questions or workflows your application actually needs to answer. Use both single-evidence retrieval and tasks that require combining multiple passages if both occur in practice.
- Make matched prompt variants. Keep the relevant evidence and question fixed, but place the evidence near the beginning, middle, and end. For multi-evidence questions, vary the spacing between the relevant pieces as well.
- Include realistic context lengths. Test the lengths your application encounters, rather than relying only on a short prompt or a model’s stated maximum context window.
- Score answers against a consistent rubric. Record whether the answer is correct and supported by the supplied evidence. Separate an answer that is wrong from one that is correct by coincidence or unsupported.
- Compare conditions across repeated cases. Look for consistent differences by position and spacing, not just one conspicuous miss. If results vary, report the task, model version, prompt length, and evaluation setup with them.
- Change one factor at a time when troubleshooting. If performance drops, test whether changing evidence order, reducing irrelevant context, or improving how evidence is grouped changes the outcome. Re-run the same evaluation before treating a prompt change as a fix.
This kind of test does not establish a universal ranking of models. It tells you how a particular model and application behave under the conditions you measured.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Quick Recap
Best Value
What the evidence does—and does not—show
- It shows: position can affect a model’s ability to use relevant information in tested long-context tasks, and the spacing of multiple relevant items can matter too.
- It does not show: that every model has the same bias, that a fixed share of all prompts will fail, or that every model with a long context window is unreliable.
- It does not establish: a current vendor-by-vendor ranking or a population-wide failure percentage. The reported findings are tied to the studies’ models, tasks, and evaluation setups.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
