research
Thirty People Taught an Agent Their Own Bar
Repeated feedback across writing and visual tasks became personal context, weight updates and an evolving rubric—then improved solo work after only tens of examples.
Summary
Repeated feedback across writing and visual tasks became personal context, weight updates and an evolving rubric—then improved solo work after only tens of examples.
The TAHI system treats iterative human-agent work as training data for the individual rather than another sample of population preference. Across 600 writing and visual-creation tasks for 30 people, the authors report solo-task success gains of 4.5 to 20.9 percent after tens of tasks. An evolving rubric captured 16.0 to 22.3 percent more failures than rubrics produced by language models or people alone, and some improvements transferred across users by up to 8.8 percent. These results belong to the study's participants, domains and evaluation setup; they do not establish universal personalization or eliminate the need for human review.
Why it matters
Repeated feedback across writing and visual tasks became personal context, weight updates and an evolving rubric—then improved solo work after only tens of examples.
Limits and context
- These results belong to the study's participants, domains and evaluation setup; they do not establish universal personalization or eliminate the need for human review.
Key claims
Repeated feedback across writing and visual tasks became personal context, weight updates and an evolving rubric—then improved solo work after only tens of examples.
Qualification: These results belong to the study's participants, domains and evaluation setup; they do not establish universal personalization or eliminate the need for human review.
Evidence: source-2026-09-06-002
Sources
- arXiv preprint 2609.04141arXiv · primary research
Corrections
No corrections have been recorded for this story.