Goodfire releases tools to look inside DeepSeek's R1 reasoning model

In short

Goodfire, an AI research company, says it has trained and released the first sparse autoencoders for DeepSeek's R1 reasoning model. A sparse autoencoder is a tool that breaks a model's inner activity into parts that people can study. Goodfire also reports two unusual results when it tried to steer R1. The report does not describe any model misbehaving.

What happened

Goodfire published the research on April 15, 2025. It is a research release, not a test of a model going wrong. No real person or organisation was affected, according to the report.

Goodfire says it trained two sparse autoencoders on R1, which has 671 billion parameters. Parameters are the adjustable numbers inside a model. One autoencoder used a custom reasoning dataset. The other used OpenR1-Math, a large set of maths problems. Goodfire says these are the first public tools of this kind built on a true reasoning model, and on any model of this size. It has made them open-source.

The team also tried "steering" R1. Steering means turning up or down a particular internal feature to change what the model writes. Goodfire reports two results it had not seen in models that do not reason:

• Timing matters. Steering from the very start of R1's chain of thought did not work. It only worked after the model wrote an opening phrase such as "Okay, so the user has asked a question about…". Goodfire says more than 95% of R1's English reasoning begins with "Okay,". Its hypothesis is that the model treats this phrase as part of the prompt.
• Too much steering can undo the change. In some cases, pushing a feature harder made R1 go back to its original behaviour. Only with still more force did its output become incoherent. Goodfire's working idea is that the model may sense when it is "confused" and correct itself.

What was done about it

Nothing needed fixing, since no problem was reported. Goodfire released the two autoencoders and its datasets. It also uploaded databases of the examples that most strongly trigger each feature, so researchers who cannot run R1 can still study them. Goodfire says it plans more work on how reasoning models operate internally.

Why it matters

Goodfire says it has not yet studied these steering effects systematically, and both explanations are hypotheses. It also says sparse autoencoders have known limitations and do not solve the wider problem of understanding how models work. The company says that if the self-correcting effect is common in reasoning models, then changing their behaviour, such as trying to suppress dishonest answers, may need more advanced methods. It says it hopes other researchers will use the tools to help make such models more reliable. Read more in the original report.