Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents
Reframes RSI around the harness and adds a benchmark for frontier agents improving it.
5 items · 4 sources · 4 days
Operational story trace
Follow in this browser to see new updates on your Live feed.
Latest change
Two Nov 18 papers push RSI back into the weights: BaseRSI internalizes harness gains into a base vision-language model without training, and Piccolo co-evolves 0.8B-27B decision models with their own training data.
September's recursive self-improvement (RSI) papers retrained models on their own output; November's wave instead has agents rewrite their harness (prompts, tools, memory, control loop) with the base model frozen. Early November added a benchmark for that loop, a statistical-guarantee method to keep it trustworthy, and a multi-agent version for closed-loop science.
Arc
Reframes RSI around the harness and adds a benchmark for frontier agents improving it.
Adds statistical guarantees so harness self-edits are accepted only with evidence.
Extends harness RSI from a single agent to a multi-agent scientific-discovery loop.
Moves harness gains into a base vision-language model without a training run.
Co-evolves model weights and training data, with a model family spanning 0.8B to 27B.
Reframes RSI around the harness and adds a benchmark for frontier agents improving it.
Adds statistical guarantees so harness self-edits are accepted only with evidence.
Extends harness RSI from a single agent to a multi-agent scientific-discovery loop.
Moves harness gains into a base vision-language model without a training run.
Co-evolves model weights and training data, with a model family spanning 0.8B to 27B.
What to watch — open questions
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.