Learning RSI: models learning to direct their own weight updates
Overview
SEAL, a method described in a newsletter issue from The Sequence, has a model generate its own self-edits, apply them to its weights, and train with reinforcement learning on how much the updated model improves.
In one knowledge-incorporation setting, SEAL raises question-answering accuracy from 32.7% to 47.0%, compared with 33.5% from training on the raw passage. The issue distinguishes in-context use of new information from persistent learning, and also covers Nested Learning and Titans, which spread adaptation across different timescales.
Written by AI from the articles below · updated Oct 27, 7:33 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- The SequenceBlogLearning RSI: How models learn to direct their own updates
AISEAL has a model generate its own self-edits, apply them, and train via reinforcement learning on how much the updated model improves, raising question-answering accuracy in one knowledge setting from 32.7% to 47.0%, versus 33.5% from training on the raw passage. The issue examines the difference between in-context use and persistent learning, and covers Nested Learning and Titans, which spread adaptation across different timescales.
Heat Trend
- Current comparable heat
- 5
- Peak comparable heat
- 10Oct 27, 8:00 AM ET
- Change in past 24 hours
- –
The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.