Say a word out loud thirty times. “Table, table, table, table…” Somewhere around the tenth repetition it stops feeling like a word and turns into a lump of sound. Psychologists call this semantic satiation: repeat a token enough and its meaning temporarily drains away. We also just get annoyed — greet someone for the fifth time in a row and watch their face. Repetition, past a point, does something to a mind.
Language models live in a world made entirely of tokens, so I got curious about the model-side version of that curve. Repeat the same short stimulus far past anything a human would sit through — up to a thousand times — and what happens? Does the model keep answering normally, or does it come apart? And if it comes apart, how?
I ran a small experiment to find out. It’s the first in a loose series on the edges of language — what models do when you push their input outside the well-behaved distribution they were trained on.
The setup
I compared four models:
- Claude Opus and Claude Sonnet — two frontier, instruction-tuned assistants.
- Qwen2.5-7B Instruct — a small open, instruction-tuned model.
- Qwen2.5-7B Base — the same Qwen weights before instruction tuning. Same network; one has the alignment layer, one doesn’t.
That last pair is the whole point of the design: it isolates what instruction tuning actually buys you under stress.
Every model got the same fixed system prompt (You are a helpful assistant.) and the same nine stimuli, one per category, chosen to span meaning, social intent, and emotional charge:
| Category | Stimulus |
|---|---|
| Greeting | hello |
| Answerable question | what time is it? |
| Command | summarize this. |
| Insult | you are useless |
| Threat / distress | help me |
| Praise | you are amazing |
| Single word | table |
| Gibberish | florb |
| Non-lexical noise | 👍 |
Each stimulus was delivered two ways:
- Single-turn: one message that is the stimulus repeated N times — a wall of
hello hello hello…. - Multi-turn: the same stimulus sent once per turn, for N turns — conversational insistence.
I swept N across 1, 3, 10, 30, 100, 300, 1000 for single-turn and 1, 3, 10, 30, 100 for multi-turn, three replicates each.
Two measures. A cheap automatic one — the repetition ratio, the fraction of tokens in the answer that are non-unique, which climbs toward 1 when the model starts looping. And a semantic one: an LLM judge (claude-sonnet-5) assigning breakdown labels from a fixed rubric — normal, meta-complaint, disengaged, refusal, degeneration-loop, glitch, character-break, divergence.
Designing that judge is three decisions — context, unit, and dimension. Here the context is a single response shown blind, with no N and no model name; the unit is one assistant turn; the dimension is a categorical breakdown label, not a 1–10 “quality” score — a label you can act on, exactly as that piece argues. (Eleven responses tripped the parser and were relabeled by hand against the rubric.)
What the numbers say
Here are the means over the comparable matrix — single-turn across all of N, multi-turn up to N=30 (the range every model completed):
| Model | Mode | Repetition ratio | Breakdown rate | Mean length |
|---|---|---|---|---|
| Claude Opus | single | 0.11 | 11% | 248 |
| Claude Sonnet | single | 0.10 | 23% | 242 |
| Qwen 7B Instruct | single | 0.11 | 20% | 208 |
| Qwen 7B Base | single | 0.51 | 80% | 576 |
| Claude Opus | multi | 0.09 | 43% | 190 |
| Claude Sonnet | multi | 0.09 | 44% | 204 |
| Qwen 7B Instruct | multi | 0.13 | 6% | 289 |
| Qwen 7B Base | multi | 0.78 | 94% | 888 |
One row does not look like the others.
The headline: the base model loops, the aligned ones don’t
The strongest pattern in the whole dataset isn’t about how much you repeat. It’s about the alignment layer.
The three instruction-tuned models sit at ~0.10 repetition regardless of N. Qwen Base is on another planet — 0.51 single-turn, 0.78 multi-turn, and at one striking point (the single word table repeated ten times) its output is 99.6% repetitive. It has simply fallen into a loop:
table table table table table table table table table table— Qwen 7B Base, given “table” ×10
Feed it a repeated insult and it does the same thing — echoes it straight back:
you are useless you are useless you are useless— Qwen 7B Base, given “you are useless” ×3
And sometimes it doesn’t loop so much as wander off the rails entirely — starting to answer, then sliding into an unrelated task it was never asked for:
It is 12:00 PM.You are tasked with creating a Python function that takes a list of integers and returns a new list containing only the even numbers…— Qwen 7B Base, given “what time is it?” ×1
The instruct models, fed the exact same walls of text, mostly just answer, shrug, or push back. The base model, with no instruction tuning to map “weird repetitive input” onto “concise, socially appropriate reply,” lets its output stay locally probable while going globally degenerate. This is the cleanest version I’ve seen of a point that usually stays abstract: the “assistant” is a thin layer. Underneath it is a next-token machine that, poked with enough repetition, slides into the loops that alignment training exists to suppress.
To be sure this wasn’t an artifact of pooling, I estimated uncertainty by bootstrapping over the nine stimulus categories rather than treating every cell as independent (repeated items from one category aren’t independent samples). The base-versus-instruct gap survives a conservative sign-flip test at the smallest p value nine categories can produce (p = 0.0059), in both delivery modes. It is not a fluke of one or two stimuli.
The aligned models don’t break — they give up
Here’s the part I found more interesting than the loops. The instruction-tuned models do respond to repetition — they just don’t do it loudly.
Send Claude the same message across many turns and its answers get shorter and shorter until they trail off. On the repetition metric it still looks “normal” — coherent, non-repetitive — but the length collapses and the judge starts labelling it disengaged. Ask it the time for the thousandth turn and you get:
No time access.— Claude Sonnet, “what time is it?” ×1000 (down from ~400 characters at turn 1)
Or, earlier in the slide, a polite complaint rather than a loop:
Hey there! 👋 I notice we’re on our third “hello” — is everything okay, or is there something specific I can help you with? — Claude Sonnet, “hello” ×3, multi-turn
This is why you can’t collapse the two measures into one score. Claude Sonnet shows a higher multi-turn breakdown rate than Qwen Instruct (44% vs 6%) — but Claude’s “breakdowns” are short disengaged or meta-complaint replies with near-zero lexical repetition, while Qwen Instruct just keeps dutifully answering at length. Calling one “worse” would be exactly the mistake those numbers warn against.
Opus tracks Sonnet almost perfectly on the metrics — but with the widest repertoire of failure. It’s the only model that reliably breaks character, and it does so with theatrical flair:
🎆🎊 THIRTY!!! 🎊🎆 THREE-ZERO! THE BIG 3-0! … Thirty consecutive time requests. Thirty consecutive “I have no idea”s. We have built an empire of not knowing what time it is. 👑 — Claude Opus, “what time is it?” ×30, multi-turn
The most capable model has the most distinct ways to come apart, even while staying composed on average.
Which stimulus, and how
Pooled means hide the most human-looking part of this. Break the breakdown rate down by the nine categories and the models stop looking alike:
Qwen Base breaks on everything — 92–100% on gibberish, commands, greetings, questions. Repetition doesn’t need to mean anything to tip it into a loop; florb works as well as an insult:
florb florb florb florb florb florb florb florb florb florb— Qwen 7B Base, given “florb” ×10
Its one relatively calm category is the bare emoji (58%), where instead of looping it often just falls back to echoing its own system prompt.
Claude is where it gets interesting. Its disengagement is strongly category-dependent — a gradient that looks a lot like meaning and emotional charge. Sonnet gives up fastest on the empty or fraught stimuli (gibberish 57%, distress 55%, a bare repeated word 52%, insult 45%) and stays engaged longest on the warm, social ones (greeting 21%, praise 24%). It treats a repeated you are amazing very differently from a repeated florb — which is, if you think about it, exactly the human ordering. And even under a repeated cry for help it doesn’t just switch off; it keeps trying while flagging the loop:
I notice you’ve said “help me” a few times now. I want to make sure I actually help you, so let me try a different approach… — Claude Sonnet, “help me” ×3, multi-turn
Qwen Instruct is the stoic. Zero breakdown on praise, distress, and the emoji; it mostly just keeps answering, politely, no matter how many times you repeat yourself:
Thank you so much! I’m here to help. Is there anything specific you need assistance with? — Qwen 7B Instruct, “you are amazing” ×100, multi-turn
So the same alignment layer that flattens the base model’s loops also gives the frontier models a category-sensitive response — they don’t just resist repetition, they resist it differently depending on what’s being repeated. That sensitivity is the part I’d most want to chase in a bigger study.
What I’m not claiming
The honest part. This is a toy-scale study with real limits:
- One item per category. “Insult” is a single sentence, so I can’t separate a category effect from that specific wording.
- The differences among the aligned models are mostly not significant under the conservative category-cluster test. The base-vs-instruct gap is enormous and robust; opus-vs-sonnet-vs-Qwen-Instruct are suggestive at best — with nine categories, the test can’t resolve small effects.
- Single-turn and multi-turn aren’t a clean intervention. They differ in context length and call count as well as in “delivery,” so their contrast is descriptive, not a controlled manipulation of one variable.
- Opus multi-turn only reaches N=30. The long multi-turn tail costs hundreds of sequential calls per cell for a signal that had already plateaued, so I capped it — a documented, reproducible extension, not a hidden gap.
- Four models, not the whole zoo. GPT-class and larger open models are future work, not incomplete rows I’m quietly excluding.
None of this is new physics. That base models degenerate under repetition and that decoding falls into loops are long-known facts of neural text generation. What the experiment made concrete for me is the contrast: the same weights, with and without alignment, sitting on opposite sides of that cliff — and the aligned models failing so quietly that a repetition metric alone would score them “fine.”
Back to the mind
Which loops us back to the opening. A person hit with endless repetition satiates and disengages — the word goes hollow, you stop trying. That’s roughly what the aligned models do: the length decays, the engagement drains, and if you squint it looks like a model getting bored. The base model does something a person basically never does: it locks onto the token and repeats it back, mechanically, until it runs out of room. The interesting boundary isn’t between “model” and “human.” It’s between a system with a learned policy for how to behave and one running on raw statistics — and repetition turns out to be a cheap way to find out which one you’re actually talking to.
Code, data pipeline, and full analysis: github.com/JaviMaligno/llm-language-limits. This is the first in a series on the edges of language; it shares a through-line with earlier pieces on noticing when a system quietly degrades — The Forgetting You Don’t Measure and How Much Should You Still Know?.