Revision pulls human work toward the model
We took 30 real entries to a museum competition in Helsinki and asked a model to revise them. Measured in units of the spread of the model’s own plain answers, the human entries started 13.1 units from the model’s typical answer. One revision brought them to 6.4, about half the way. Eight rounds brought them to 3.8, about 70% of the way.
Distance of human entries from the model’s typical answer
Revised together, ideas collapse
In 432 revisions, ideas revised alongside each other converged round after round, whether one model revised them or several roles took turns. Two kinds of outside material changed that:
- A random fragment per idea per round kept ideas distinct, but a judge rated them far less credible.
- A real fact per idea per round kept them apart and credible.
Even the judge’s verdict depends on its role. Asked as a practice judge, it wanted to pursue the evidence-fed ideas over the random-fed ones 21% of the time; asked as the client, it would build them 100% of the time.
A retelling test
When people retell a story in a chain, it keeps changing. In the same chain task, Claude Sonnet kept the story nearly fixed. Part of that is length: people halved the story at each step, while the model kept about 135 words, and our similarity measure is length-sensitive. An older model, ChatGPT-3, changed stories about as much as people did. So this is about one model without a length instruction, not about language models in general.
What it means
A closed loop is a converging loop. Every step of our workflow that revises an idea brings in new outside material, and we read it against the open loop idea: invention escapes collapse through fresh evidence and selection.
Sources
- Feathers & Wax, drift, flux and evidence runs, code in
research/cllm-divergence. - Fritz Breithaupt et al., story retelling chains, 2024 (116 stories).