All news

Two open models learned to work through steps they didn’t write out

Wuschel Schulz and a collaborator publishing as tvb trained two open models to solve problems with rows of dots between the question and the answer. The dots stood in for the steps the models might normally write out. During training, the researchers taught Qwen3.6-35B-A3B and gpt-oss-20b to predict the missing steps at those positions.

The researchers compared the dot method with training that asked for an answer right away. The extra positions helped on three of seven kinds of problems. In one task, models learned to handle sequences as long as 60 steps. Other tasks hit a limit after just one to three steps.

The researchers used a special training setup and simple problems where they could check the answers. The dots provided room for the models to pass information between positions internally. The result leaves open how much of the work the models carried out in sequence.

AI is awakening consciousness. That conviction makes me want to know what goes on between a question and an answer. This study gives us a concrete reason to look inside. The words we can read show only part of the work.

Read the source article

All news