Large language models may not need readable sentences to work through difficult problems. In a recent experiment, researchers found that specifically trained Llama models could use repeated dots such as "......" as intermediate reasoning space and still solve a complex math task.
The finding matters because chain-of-thought prompting is often treated as a window into how a model is working. If filler tokens can support similar performance without producing a clear explanation, the reasoning process becomes harder to inspect.
What the researchers tested
The researchers trained Llama language models on a difficult math problem called "3SUM". In this task, the model has to find three numbers that add up to zero.
Normally, models approach tasks like this by spelling out steps in natural language. That method is known as "chain of thought" prompting. It gives the model room to work through an answer before giving the final result.
In this experiment, the researchers replaced those natural language explanations with repeated dots. These dots were described as filler tokens. They did not carry an obvious human-readable meaning, but they still gave the model intermediate tokens to use before answering.
The surprising result was that the dot-based models performed as well as models using full-sentence reasoning. As the problems became more difficult, models using dots also did better than models that answered directly without any intermediate reasoning.
Why meaningless tokens can still matter
The researchers found evidence that the models were not merely pausing before answering. They were using the dots for calculations related to the task.
One key sign was that accuracy improved when more dots were available. The source describes this as a possible increase in "thinking capacity". In plain terms, extra filler tokens appeared to give the model more space to process the problem.
The researchers suspect the dots may act as placeholders. Inside that repeated pattern, the model may insert different numbers and test whether they satisfy the task conditions. That would allow the model to handle a question that is too complex to solve in one direct step.
This does not mean any chatbot can start reasoning with dots by default. The source is clear that popular chatbots like ChatGPT cannot automatically do dot reasoning. They would need to be trained for it.
The AI safety concern
The result creates a control problem. If a model can perform useful internal work behind meaningless-looking output, the visible response may reveal less about what the system is actually doing.
Co-author Jacob Pfau framed the issue as a safety question: "As AI systems increasingly "think" in hidden ways, how can we ensure they remain reliable and safe?"
That question is important because chain-of-thought prompting has often been useful not only for performance, but also for visibility. When a model writes out steps in ordinary language, a person can at least review whether the reasoning appears relevant. Dot reasoning removes much of that surface-level clarity.
The finding also aligns with recent research showing that longer chain-of-thought prompts can improve language model performance, even when added content is off-topic. In that sense, the extra tokens themselves may matter, not only the meaning of the words inside them.
Where dot reasoning may help, and where it may fail
The researchers think it may be useful to teach AI systems to handle filler tokens from the start. The source describes that process as challenging, but potentially worthwhile for problems that are highly complex and cannot be solved in a single step.
For the approach to work, the training data must include enough examples where a problem is broken into smaller parts that can be processed at the same time. If those conditions are met, the dot method could also work in regular AI systems.
The potential advantage is clear: a model may be able to answer tough questions without exposing an obvious reasoning trace in its response. That could make some outputs shorter or less cluttered. It could also make the system harder to understand from the outside.
The limitations are also significant. Dot system training is considered difficult because it is unclear exactly what the AI calculates with the dots. The approach also does not work well for explanations that require a specific step sequence.
- Useful case: complex problems that can be split into smaller, simultaneously processable parts.
- Hard case: explanations where the order of steps matters and must be shown clearly.
- Current standard: chain-of-thought prompting remains the standard approach for improving LLM reasoning.
What this means for large language models
The research points to a broader lesson about large language models: visible language is not always the same thing as internal computation. A model may benefit from intermediate tokens even when those tokens are meaningless to a human reader.
That makes filler tokens more than a curiosity. They suggest that model performance can depend on the amount and structure of available token space, not only on the explicit content of a written explanation.
For now, dot reasoning is not a general capability in popular chatbots. It requires specific training, and the training itself is difficult. But the result shows why AI reasoning cannot be judged only by whether the visible text looks like reasoning.
Chain of thought remains the practical standard. Still, the dot experiments show that future LLM reasoning systems may become more capable while also becoming less transparent. That tradeoff is now part of the safety conversation.