🤖 machine learning

When More Becomes Less: Position-Dependent Repetition Effects in Language Models

This paper demonstrates that the impact of token repetition on language model predictions is position-dependent, revealing that while adjacent repetition consistently boosts target probability, displaced repetition produces an inverted-U effect where probability peaks and then declines as repetition increases, challenging the assumption that repetition effects are uniform regardless of readout position.

Original authors: Han-yu Wang

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Han-yu Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Secret Life of Repetition in AI Brains

Imagine you are trying to teach a robot how to tell a story. You give it a prompt, like a sentence starter, and ask it to guess the next word. This is how many modern Artificial Intelligence (AI) systems, known as Large Language Models, work. They are like super-powered autocomplete engines that have read almost everything on the internet. To figure out if they actually "know" something, scientists often use a trick called a "cloze test." It's like a fill-in-the-blank puzzle: you show the AI a sentence with a missing word, and you see if it can guess the right one.

For years, researchers have believed a simple rule about how these AI brains handle repetition. They thought that if you repeat a word over and over again in the prompt, the AI would get more confident about that word, just like a human might get more comfortable with a concept the more they hear it. This is called "priming." If you say "apple, apple, apple," the AI should think, "Okay, this is definitely about apples!" and predict "apple" with high confidence. Scientists assumed this confidence would keep climbing the more copies of the word you added, no matter where in the sentence they asked the AI to make its guess. But what if that assumption is wrong? What if the AI's brain works differently depending on where you ask it to look? This is the question that Han-yu Wang's paper tackles, exploring a strange quirk where "more" repetition actually leads to "less" confidence, but only under specific conditions.

When More Becomes Less: The AI's Forgetful Spot

In this study, the researcher set up a clever little experiment to test how AI models react to repetition. Imagine you have a magic word, let's say "cat." You can show this word to the AI in two different ways, but you keep the beginning of the story exactly the same.

The "Right Next Door" Test (Adjacent Probe):
First, you write a long list of "cat" words right up to the end of the sentence: "cat, cat, cat, cat... [guess here]." You ask the AI to guess the very next word.

What happened: Just like the old rule predicted, the AI got more and more confident. The more "cat" words you added, the more sure it became that the next word would be "cat." It was like a drumbeat getting louder and louder.

The "Fresh Start" Test (Displaced Probe):
Then, the researcher did something sneaky. They wrote the exact same long list of "cat" words at the start, but then they added a full stop and a brand new sentence frame before asking for the guess. It looked like this: "cat, cat, cat, cat... . The sky is blue. [guess here]."

What happened: This is where the magic trick failed. Instead of getting more confident, the AI's confidence did a weird backflip. At first, adding a few "cat" words helped. But as soon as the list got too long (around 5 to 10 copies), the AI started to lose its confidence. By the time you had 30 "cat" words, the AI was actually less likely to guess "cat" than it was with just a few copies.

The researcher calls this shape an "inverted-U." It's like the AI gets a little excited at first, but then gets overwhelmed or confused by the sheer amount of repetition when it has to jump to a new sentence to find the answer.

The Big Discovery: It's All About Where You Look

The most important finding of this paper is that the position of the question matters immensely. The same list of repeated words can make the AI smarter if you ask it immediately after the list, but make it dumber if you ask it a sentence later.

The study tested this on 13 different AI models, ranging from small ones to very large ones. They found that this "inverted-U" pattern happened in every single one of them. It wasn't just a fluke with one specific computer brain. They even tried it in four other languages—Spanish, Chinese, German, and French—and the same weird pattern appeared. This suggests that this isn't a mistake in how the AI speaks English, but a fundamental quirk in how these models process repeated information.

What It's NOT (And What It IS)

The researcher didn't just stop at finding the pattern; they played detective to figure out why it was happening. They wanted to make sure it wasn't just because the sentence got too long, or because the AI was tired, or because it was thinking about words related to "cat" (like "kitten" or "purr").

It's not about length: They tested sentences with the same number of words but random, boring filler words instead of repeating "cat." The weird drop in confidence didn't happen.
It's not about related words: They tried replacing the repeated "cat" with words that mean the same thing or are related. Again, no drop in confidence.
It IS about exact repetition: The drop only happened when the exact same word was repeated over and over.

This means the AI isn't just getting "bored" or "confused" by long text. Something specific happens when the exact same token (the digital building block of a word) is repeated too many times before a new sentence starts.

Inside the AI's Brain

To understand what was going on, the researcher peeked inside the models' "attention" mechanisms. You can think of attention like a spotlight the AI shines on different parts of the sentence.

The Spotlight Shrinks: As the list of "cat" words grew, the amount of attention the AI gave to each individual "cat" word got smaller. This makes sense; if you have a fixed amount of spotlight and you have to shine it on 30 things, each thing gets less light.
The Total Budget Grows (in some models): Surprisingly, in the models that showed the biggest drop in confidence, the total amount of attention the AI gave to the whole block of "cat" words actually increased as the list got longer. It was like the AI was shouting "Look at all these cats!" louder and louder, but somehow, when it had to jump to the new sentence, that shouting made it harder to remember the specific word "cat."

Why Should You Care?

This paper is a warning label for anyone trying to test how smart these AI models are. If you are a scientist trying to see if an AI knows a fact, and you use a method that repeats a word many times to "prime" the AI, you might be getting the opposite of what you expect. If you ask the AI the question a sentence later, you might accidentally make it less likely to give you the right answer.

The study shows that we cannot treat the position of a question as a neutral detail. Where you ask the question changes the answer. It's a reminder that these AI brains are complex and sometimes counter-intuitive: sometimes, hearing a word too many times doesn't help you remember it; it might just make you forget it, especially if you have to wait a moment before answering.

The researcher tested this on 13 models and found the effect in all of them. They checked it in 42 different combinations of models and languages, and the pattern held up every time. While they can't say exactly why the AI's brain works this way (it's a correlation, not a full explanation of the mechanism), they have proven that this "more becomes less" effect is real, consistent, and happens across the board. It's a small but crucial piece of the puzzle in understanding how these digital brains actually think.

Technical Summary: Position-Dependent Repetition Effects in Language Models

Problem Statement

Cloze-style probing studies in language models (LMs) frequently vary the frequency of a target token in the prompt to measure effects on prediction (e.g., semantic priming or in-context learning). These studies implicitly assume that the readout position (where the model's prediction is sampled) is orthogonal to the manipulation: that the effect of repetition on prediction probability remains consistent regardless of whether the prediction slot sits immediately after the repeated block or is displaced into a new sentence frame.

This paper challenges that assumption, demonstrating that the position of the readout slot fundamentally alters how repetition affects retrieval. Specifically, while repetition acts as a monotonic primer when the slot is adjacent to the repeated block, it produces a non-monotonic "inverted-U" effect when the slot is displaced, causing the probability of the target to decline as repetition count increases beyond a certain point.

Methodology

The Two-Probe Design

The author introduces a matched-prefix experimental design to isolate the effect of readout position:

Adjacent Probe: The prediction slot is placed immediately after a prefix containing $N$ copies of a target token (e.g., target × N [MASK]).
Displaced Probe: The prediction slot is moved into a fresh sentence frame following the same prefix (e.g., target × N . <fresh frame> [MASK]).

The study holds the repeated-target prefix fixed while varying only the slot position. The repetition count $N$ ranges from 0 to 30.

Experimental Setup

Models: 13 open-access language models were tested, including 2 masked LMs (MLMs: ModernBERT-base/large) and 11 causal LMs (CLMs) of varying sizes and architectures (Qwen2.5 family, SmolLM2, Phi-3.5, OLMo-2).
Targets: 261 hand-curated single-token nouns, filtered per tokenizer to ensure single-token validity.
Languages: The effect was replicated across English, Spanish, Chinese, German, and French using multilingual checkpoints.
Metrics: The primary statistic is the per-word drop, defined as $(peak_w - P(N=30)_w) / peak_w$ , where $peak_w$ is the maximum probability observed for a word across $N \in \{1, 3, 5, 10, 20, 30\}$ . A positive drop with a bootstrap confidence interval excluding zero indicates the inverted-U effect.

Ablation and Controls

To isolate the mechanism, the author conducted:

Six-condition causal ablation: Replaced the repeated block with truncated repetitions, deletions, random fillers, semantic-neighbour fillers, and unrelated repetitions.
Frame-pragmatics control: Varied the sentence frame to rule out artifacts of the specific frame structure.
Pre-stated answer: Provided the target explicitly at the start of the prompt to test if the effect persists when an alternative retrieval route exists.
Internal analysis: Measured per-target-token attention and total attention budget assigned to the repeated block.

Key Results

1. The Displaced Inverted-U

The core finding is a behavioral dissociation between the two probe positions:

Adjacent Slot: $P(\text{target})$ climbs monotonically with $N$ and plateaus, consistent with standard priming intuition.
Displaced Slot: $P(\text{target})$ rises to an early peak (typically $N \in \{1, 3, 5, 10\}$ ) and then declines as $N$ increases to 30. This "displaced inverted-U" was observed in all 13 models tested, with the per-word drop being statistically significant (CI excluding zero) in every case.

2. Robustness and Generalization

Cross-lingual: The effect replicated across 42 cells (4 models × 4 languages × 3 frames) in Spanish, Chinese, German, and French.
Scale and Architecture: The effect appears across model sizes (360M to 14B parameters) and architectures (MLMs and CLMs). However, the magnitude of the drop is not a monotonic function of parameter count; for instance, within the Qwen2.5 family, the 1.5B and 14B models showed larger drops than the 7B model.
Instruction Tuning: Instruction tuning (e.g., Qwen2.5-7B vs. Qwen2.5-7B-Instruct) did not eliminate the effect; in some cases, it increased the magnitude of the drop.

3. Causal Mechanism

Exact Lexical Repetition: The effect is driven specifically by exact lexical repetition of the target. Semantic-neighbour fillers (which activate the target's semantic field) failed to reproduce the inverted-U, and random fillers/deletions eliminated the effect entirely.
Not Length or Redundancy: The effect is not a function of sequence length or generic redundancy.
Internal Correlates:
- Per-token Attention: The attention received by each individual target token falls as $N$ increases.
- Total Block Budget: In Causal LMs, the total attention budget assigned to the repeated block grows with $N$ (sub-dilution), whereas in the Masked LM, the budget remains roughly flat. This suggests autoregressive training induces a specific dynamic where the model allocates increasing total resources to the repeated block, yet the per-token signal degrades in a displaced context.

Significance and Claims

The paper claims two primary contributions:

Empirical Finding: Current open-access language models fail to aggregate exact lexical repetition additively at displaced slots. Adding copies of a target can lower its probability one clause beyond the block, even when the same prefix successfully primes an adjacent slot. This contradicts the assumption that repetition effects are position-independent.
Methodological Instrument: The "matched-prefix two-probe design" provides a tool to separate local pattern continuation from displaced retrieval. The author argues that any controlled probing study involving repeated content (e.g., in-context learning benchmarks, factual knowledge probes) cannot treat the readout position as orthogonal to the manipulation. A manipulation that appears as monotonic priming at one position may invert at the position actually scored in many benchmarks.

The author remains modest regarding the underlying mechanism, noting that their internal probes are correlational. They suggest the effect may stem from softmax cooperation between repeated keys, attention-sink dynamics, or autoregressive-specific recency priors, but their data does not distinguish between these candidates. They also note limitations, such as the restriction to single-token targets and the lack of a clean explanation for the SmolLM2-1.7B exception regarding mean-curve inversion.