Pick the most accurate statement about how irrelevant chunks degrade answer quality

Attempt it

The intuitive picture, the model finds the right chunk and ignores the rest, is wrong. The empirical picture is that irrelevant chunks (distractors) **actively degrade** accuracy even when the right chunk is present in the same context. The mechanism is attention. The model spreads attention across everything in context. Adding distractors gives the model more candidates to attend to, and even if each distractor only gets a small share of attention, that share is taken away from the right chunk. The result is that the model partially grounds in the wrong material and the answer drifts. This happens in three modes. First, **distractor pulling**: the model occasionally picks a fact from a distractor that sounds plausible. Second, **dilution**: the right chunk is still attended to but its influence is diluted, so the answer is fuzzier. Third, **wrong but grounded answers**: the answer cites real sources and looks confident, but it answers a different question than the user asked. This effect persists on large context windows. The lost-in-the-middle curve and distractor dilution compound: not only is the right chunk competing with distractors for attention, it is often in the middle of the window where attention is weakest.

**Irrelevant chunks are not free.** Adding distractors to a context that already contains the right chunk measurably drops accuracy, and the size of the drop is empirically significant, often 5-20 percent on standard RAG benchmarks depending on the number and similarity of distractors. The mechanism is attention dilution. Attention is a soft retrieval mechanism that distributes weight across all tokens in context. Adding distractors gives the model more candidates to attend to and reduces the share given to the right chunk. The model does not cleanly identify and ignore irrelevant material; it weighs everything it sees. The failure modes are three. **Wrong-but-grounded answers** are the most common: the answer cites real retrieved chunks but focuses on the wrong aspect, the wrong entity, or the wrong question. **Distractor pulling** is when the model lifts a specific phrase from a distractor that looked relevant. **Hallucinations** are the rarest of the three; the answer fabricates content that does not appear in any chunk. Option 3 in the question wrongly elevates hallucination to the dominant mode. The effect does not go away with bigger context windows. Long-context benchmarks like RULER and BABILong show the irrelevance penalty persisting on Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro at sizes well beyond 200K tokens. The lost-in-the-middle curve compounds it: distractors near the head and tail of the window get more attention than the right chunk if the right chunk is in the middle. The operational implication is to retrieve narrow and trim aggressively. A 2026 production RAG pipeline reranks and trims to top-3 or top-5 for this reason.

Quick reference

TL;DR

Distractors actively pull the model's attention even when the right chunk is present; accuracy degrades measurably and the effect persists on large-window models.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Picture a kid trying to answer a homework question with the textbook open. If only the right page is visible, they read it and answer. Now put six other open books on the desk, all about adjacent topics. The kid does not cleanly ignore them. Their eyes wander, they grab a phrase from the wrong book that sounded relevant, and the answer comes out muddled even though the right page was right there. A bigger desk does not fix this. It just lets you fit more open books. Models in long-context mode are the kid with the cluttered desk. Distractor chunks are the open books that should not be there.

Key concepts

Context engineering Retrieval-augmented generation Attention

Detailed answer & concept explanation

~5 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

4 minutes: the empirical signature, the attention-dilution mechanism, why big windows do not help, the failure-mode distribution, operational defenses.

Real-world examples

Real products, models, and research that use this idea.

Liu et al.'s 2023 lost-in-the-middle paper formalized the irrelevance penalty and its position-dependence.
RULER (NVIDIA, 2024-2026) and BABILong are the standard long-context evals that quantify irrelevance and distractor effects at 200K-1M context.
Anthropic's Claude Opus 4.7 RAG cookbook explicitly warns against passing top-30 directly into context for this reason and recommends rerank and trim to top-5.
Cohere Rerank 3.5 was designed against benchmarks that include controlled distractor injection precisely because the irrelevance penalty is the production failure mode.
Pinecone and Weaviate both ship distractor-aware eval templates in 2026 that surface the faithfulness-relevance divergence.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow does the irrelevance penalty interact with the lost-in-the-middle curve?

Sign in or go Pro to see the approach for this follow-up.

QHow much accuracy drop should you expect from 5 distractors added to a context with the right chunk?

Sign in or go Pro to see the approach for this follow-up.

QHow would you build an eval that isolates the irrelevance penalty from other RAG failures?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Assuming that retrieving extra context is harmless because the right chunk is still in the pile. The model does not cleanly ignore the rest; it spreads attention across everything you gave it.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.