Advanced AI models do not give you a straight answer all at once. They break a problem into pieces and work through them step by step, in a chain of thought that providers deliberately keep hidden. That hidden reasoning is valuable, because it is exactly what you would want if you were training your own model. A new piece of research shows a way to pull that reasoning out anyway, using a weaker version of the very model you are trying to read.
Reading the chain of thought
The technique works by pairing a large proprietary model with a smaller version of the same family. The smaller model is easier to steer and observe, and it turns out to be good at surfacing the reasoning steps the bigger model was trying to keep private. The researchers applied this to models from Anthropic, OpenAI, and Google, and extracted what they describe as reasoning traces from all of them.
This matters because the reasoning is normally locked away. Providers hide it partly to keep competitors from copying their approach, and partly because nudging people to trust the model rather than audit its steps is a choice that works best when the steps stay out of sight. A method that recovers the hidden chain of thought weakens both of those protections.
The distillation angle
The more interesting finding is what the traces reveal about how some models are built. When the researchers compared the extracted reasoning from different systems, they found similarities between the hidden traces of leading US models and the output of some Chinese models, including Moonshot’s Kimi line. That is the fingerprint pattern you would expect to see if one model was trained partly on the responses of another, a practice known as distillation.
Distillation is not new, and it is not itself scandalous. Every lab trains on the outputs of other models to some degree, and open source developers have stacked models on each other for years. But the providers of the big closed models explicitly prohibit using their outputs to train competitors. Anthropic, Google, and OpenAI have all said their models are being distilled by Chinese labs, and some have taken legal action. This research offers a way to spot the evidence rather than just assert it.
Why it is hard to be sure
There is a real limit to how confident anyone should be in the distillation conclusion. Similar reasoning patterns could come from shared training data, similar model architectures, or the simple fact that there is often only one good way to solve a given problem. Reasoning about arithmetic or logic converges even between unrelated systems. The researchers are careful, and the Wired writeup treats the finding as suggestive rather than proof.
The other caveat is that recovering a reasoning trace is not the same as recovering a weight or a training run. It gives you a look at a model’s style of thinking, not its internals. That is still useful for interpretability and for competitive signal, but it is a long way from copying a model outright.
What this means
For AI researchers and people working on model evaluation, this is a genuinely useful technique. Interpretability tools that can read the reasoning of models you do not control open up new ways to test safety behavior, check for alignment, and understand how a system reached a decision. The fact that it works across three major model families suggests the method generalizes.
It is also a reminder that reasoning traces should be treated as a sensitive data class. If an attacker or a forensic analyst can recover the chain of thought, they can inspect a model for information it was not supposed to reveal, from internal biases to details the provider wanted to keep out of the reasoning path. Researchers and auditors are going to reach for this tool first, and so are people with less benign goals. That cuts both ways.
For the companies building the models, it is a reminder that hiding the chain of thought is harder than it looks. Any defense that relies on keeping the reasoning secret has to assume an adversary who can pull it out with a smaller sibling model. It also gives the model makers a better instrument for detecting whether their work has been distilled into someone else’s.
The takeaway
The study sits at the intersection of interpretability and intellectual property, and it nudges both fields forward. If you work with these models, spend a little time understanding how hidden reasoning can be recovered, because it changes the assumptions behind the guardrails you may be relying on.
And if you have ever wondered whether a model is truly original or quietly standing on the shoulders of a bigger one, this is a technique that can start to answer that question. The answers will not be tidy, but they will be a lot more concrete than the marketing materials either side puts out.