Goodfire probes show AI models often 'perform' reasoning on easy questions

In short

Goodfire, an AI research company, reports that two open AI models often "know" their final answer well before their written reasoning shows it. This happened mostly on easy questions. Goodfire says it does not mean the models are being deceptive.

What happened

Goodfire published the work on March 12, 2026. It was done in research tests, not in real-world use. The source does not describe any person or organisation being affected.

Many AI models now write out their reasoning step by step before answering. This is called chain-of-thought. Goodfire wanted to compare what a model writes with what is going on inside it.

The team studied two large open models, DeepSeek-R1 and GPT-OSS-120B, plus smaller versions of DeepSeek-R1. They used two sets of multiple-choice questions, MMLU-Redux and GPQA-Diamond. They used "probes," which are small tools trained to read a model's internal signals and predict its final answer while it is still working.

Goodfire calls it "performative" reasoning when the written steps read like real deliberation but the internal signals already point to an answer. It reports this was common on easier, recall-based questions. On harder questions that need several steps, the written reasoning and the internal signals improved together. That suggests the writing there reflected real reasoning.

Goodfire also looked at moments where a model writes things like "Wait" or "Aha!". Responses where the probe showed high confidence had about half as many of these as less confident ones. But individual moments did not reliably line up with shifts in the probe's confidence.

What was done about it

Goodfire tested using the probes to stop reasoning early once they were confident. For DeepSeek-R1, stopping at 95% probe confidence saved 68% of tokens (the small chunks of text a model produces) on MMLU and 33% on GPQA-Diamond. It kept over 95% of its original accuracy. For GPT-OSS, savings were 61% and 26%, with accuracy at about 95% and 98% of the baseline.

Goodfire says that if the only goal is accuracy, it may still make sense for a model to keep reasoning even when it is confident.

Why it matters

Goodfire says tools that read a model's reasoning are only as good as the reasoning itself. It argues that probes can add information that the written text does not reliably show. It says understanding why the gap exists is important for keeping these systems monitorable and aligned. The report does not say the models are lying. It says "performative" describes a mismatch between written reasoning and apparent internal state.

The tests covered two model families and two multiple-choice benchmarks. Read more in the original report.