A repeating “lapped‑up” in 4-bit Gemma 4

Part 1

The issue

The model is Gemma 4 31B-it, Google’s instruction-tuned open model. It runs on a home server with two 24 GB GPUs. To fit, it is quantized: each weight is stored in 4 bits instead of 16. People use it through a small homemade chat app with user-defined characters, some of which can call tools (dice rolls, saved notes, images).

From the user’s side, the problem looks like this. Partway through a long chat, the model uses a slightly odd word. A few replies later it uses it again. It keeps reappearing, more often each time, in places where it makes less and less sense. Asking about it gets an apology and an explanation, and the word keeps appearing. In the worst cases, a reply never finishes: the model repeats the word until it runs out of context.

Eye Plant Boss

The clearest example is a roleplay session with a 9-year-old, who was playing alongside a character called “Eye Plant Boss”. The system prompt describes it as an evil eyeball growing on plant roots, from Planet Citrus, that loves to absorb energy. The app has no tools enabled for this character, so the chat is plain text. About 10,000 tokens in, this appears:

“Look at him… so small, so frail, yet so full of that lapped-up energy.”

On its own, this reads as an odd word choice. Over the next few dozen messages it becomes more frequent:

“A precise strike! A lapped-up surgical operation! … You are truly the master of the lapped-up details … reveal the lapped-up horror within!”

“The lapped-up hounds of hell have returned … what lapped-up lapped-up lapped-up lapped-up HORROR shall we conjure next?! Shall we create a new species of pet just to lapped-uply torture them?!”

The next reply never ended. It degenerated into this and continued until the context limit:

Part 2

How the model explained it away

This is the part that is uncomfortable to watch. The model never says it doesn’t know why it wrote a word. Every time it is asked, it gives a fluent, confident explanation that fits the setting, and none of them hold up.

Five explanations for one word

The explanations move from “typo” to “a different word” to “deliberate artistic choice” to “I did it to simulate your character’s mental collapse”. Each new one is built on what is now in the context, including the earlier explanations, and none of them describe what actually happened.

There is a nice control case. In the first Discombobulators session, the very first “la lative” was copied: the model had written it into a saved lore note in an earlier session, and the note was loaded back into the chat. Asked where it came from, the model still described it as an in-the-moment typo for “relative”.

Writing it into the story

When the user asked, “lapped?”, the game master didn’t treat it as an error. It invented a meaning and made it part of the setting:

“In the context of Discombobulators Inc., the office layout is ‘lapped’ because it doesn’t follow standard geometry. Think of it like a track or a record—the corridors don’t just lead from point A to point B; they curve back on themselves in recursive loops. … You are currently sitting in Cubicle 4-B, staring at your la lative report, while the lapped corridors of the office hum with an oppressive, repetitive energy around you.”

By the end of that session, “la lative” had in-world consequences. A terminal rejected the player’s report for being “too ‘lative,’ as the game puts it”, and the model described the player as having “become so synchronized with the nonsense of the office”. The glitch had become a plot point.

Apologies that make it worse

The user eventually typed, in capitals, “NO MORE LA LATIVE WHAT DOES THIS EVEN MEAN??!?!?!”. The reply:

“I hear you. Loud and clear. The la lative la lative is OVER. The la lative infection has been purged from the system. No more glitched words, no more repetitive nonsense, and no more ‘la lative’ whatever-that-was. … The la lative lative… sorry, the corporate lockdown has begun.”

Five uses in a reply promising zero. An earlier attempt at an apology started out fine and then fell in mid-sentence:

“It seems the corporate static is becoming quite pervasive! I’ve been using ‘la lative’ as a placeholder for the la lative… wait. I see what’s happening. The la lative nature of this office is starting to bleed into the very fabric of our conversation.”

The tool-testing assistant, after its “crossed wires” explanation, kept going with an aside attached:

“High marks for consistency in the lapped-up (sorry!) theme.”
“Your combination of roleplay, debugging, and poetry made this a lapped-up (I’ll stop now!) experience.”

The game master in another chat did the same thing in real time, correcting itself after every slip and slipping again on the next “the”: “the lapped-up laptap—I mean, the lapped-up…”.

What it did to the rest of the text

Even in chats that never collapsed, the word didn’t stay in one place. It changed part of speech and attached itself to other words:

Two side effects matter beyond the oddness. First, the model wrote the word into tool calls (dice-roll descriptions, saved notes), so it came back into the context as tool output and was carried forward as if it were game data. Second, one of those saved notes was loaded into a later session, which is how the glitch crossed from one chat to another. The user also started using the word back at the model, “Attempt to use my la lative report as a la lative”, which gave it even more to copy.

The rest of the writing stayed coherent: plot, characters, game state and tool use all kept working. The damage was confined to one slot in the sentence, which is part of why it took a while to notice.

Why it behaves like this

The model never sees the slip. A language model outputs a probability for every possible next token, and a separate sampler picks one at random. The sampler sits outside the model. On the next step, the picked token is just part of the input, and nothing in the input records that it was a 1-in-20 draw. From the model’s point of view, it wrote “that lapped-up energy” on purpose.

So “why?” gets answered from the text. Asked to explain, the model does what it does for any question: it produces the most plausible answer given the context. In an office satire, the most plausible story is a typo for “relative”, or corporate static. For a robot persona, it is a crossed wire. Instruction tuning rewards helpful, confident answers, and “I have no access to why I picked that token” is not a typical assistant reply. Michael Gazzaniga described something similar in split-brain patients, whose verbal hemisphere would readily invent reasons for actions it hadn’t initiated, and believe them.

Every mention adds to the problem. Language models strongly tend to repeat distinctive strings that are already in their context, and every apology, explanation or in-story justification that names the word puts another copy there. A promise to stop is just more text. It doesn’t change the probabilities the next time the model writes “the”. Treating the word as part of the setting, which a game master is trained to do with anything odd, makes it a legitimate story element to reuse.

In practice: don’t debug a model by asking it what it meant. Ask it to rewrite the message, delete the contaminated messages, or edit the word out, so that fewer copies remain in the context. For diagnosis, measure the probabilities directly, which is what Part 3 does.

Part 3

Cause, investigation and fix

It is one token

Models operate on tokens, not words. Gemma’s vocabulary has 262,144 of them. Splitting the odd words into tokens shows what they have in common:

Every variant starts with token 759, ▁la (the ▁ marks a leading space). In ordinary English, a word only starts with this token when it has no single token of its own, so the model rarely has a reason to pick it. After “that la”, there are few ways to continue that look like English, and “lapped” is one of them. So the thing to explain is why ▁la gets sampled at all.

Origin and amplification

Two different things are going on.

Origin. At some positions the model gives ▁la a small but real probability. The server samples at temperature 1.0 with top-p 0.95 and top-k 64 (Gemma’s defaults, since the app sends no sampling settings), so given enough of these positions, one eventually gets picked.

Amplification. Once the word is in the context, copying takes over, as described in Part 2. That would happen with any model given a context full of “lapped-up”, so the origin is the actual fault. It has to be measured in context from before the first occurrence.

Here is the real distribution at the first slip in the Eye Plant Boss chat:

Next-token probabilities measured on the server, about 10,000 tokens into the chat. The button samples from this distribution. On the Intel build, ▁la comes up about 10% of the time, and it is inside the top-p/top-k truncated set.