LLM Digest
Subscribe

AI Storyline

5 items · 4 sources · 4 days

View as JSON

Operational story trace

Recursive Self-Improvement

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.

Earlier contextThe story so far

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.

editor-curated · source-linked

Arc

Nov 6Nov 18 · now
MEASUREMENT · Nov 6–7
Harness RSI gets a benchmark and a statistical safety check
2 sources · show sources ▾
MULTI-AGENT · Nov 9
SciMARS applies recursive harness adaptation to multi-agent scientific discovery
1 source · show source ▾
NOW · Nov 18
RSI moves from the harness back into the model
2 sources · show sources ▾

What to watch — open questions

  • Does Flex-RSI publish per-model scores that show which frontier agents actually gain from editing their own harness?
  • Do statistical acceptance tests for self-edits hold when the eval distribution shifts after deployment?
  • Does BaseRSI's training-free internalization keep its gains on tasks outside the harness it learned from?
  • Do any of these loops report results outside their own benchmark harnesses?
How this thread was built
editor wrote the arc · 3 beats

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.