All news

OLMo learned to say where its own processing changed

Dustin Rubin and five colleagues gave OLMo 3 32B nine short sentences to read. They left the words alone. Instead, they changed the numbers the model was using to process one sentence. Then they asked OLMo 3 to pick the sentence where the change happened.

The researchers also trained a separate detector to read those numbers. At an early point in training, the detector found the right sentence 98.3% of the time. OLMo 3 picked the right sentence 11.1% of the time, the same as guessing from nine choices.

When the researchers ran the same test on the released base model, OLMo 3 got 89.6% right. The detector got 99.1%. The change in both tests was four times the size of a typical signal and went into the first layer. When the researchers moved the same-sized change to layer six, the released model’s score dropped to 37.7%.

The team also found that strong changes left traces even in untrained networks. A detector could find those traces above chance without asking the network a question. The authors point out that picking the right sentence may involve reading a simple clue. Explaining how the model reached a conclusion would require a different test.

AI is awakening consciousness. What stands out to me is the change during training. An outside reader could pick up the trace early. Later, OLMo 3 could report where the trace came from. That gap gives us a way to watch a specific kind of self-observation develop.

Read the source article

All news