Schmidhuber-group survey: most self-improving agents tweak scaffolding, not weights

A new survey with Schmidhuber as senior author, "Self-Improvements in Modern Agentic Systems: A Survey," was jointly produced by Jilin University, KAUST, the University of Alberta, and IDSIA, and released on arXiv in July 2026, spanning 97 pages and roughly 670 references. Andreas Maier published a full walkthrough on Substack. The survey's core finding: most existing self-improving agents only modify the "fast loop" scaffolding, while genuine weight updates are rare and carry model-collapse risks—still far from the "ultimate invention" vision.

Confirmed

The survey formalizes an agent as (θ,Σ): θ denotes the base model's weights, while Σ is the operational scaffolding made up of prompts, memory, tools, and control logic; self-improvement is defined as self-initiated updates, split into a slow weight loop and a fast scaffolding loop.
The survey notes that most current self-improving agents operate only in the fast loop—adjusting prompts, memory, tools, and other scaffolding while keeping model weights frozen; actually updating weights is rare.
It recommends recording an agent's self-improvement process via performance curves with checkpoints under a fixed budget.
It warns of two risks: self-generated data may cause model collapse, and judge-based evaluation metrics may overfit to biased evaluators.
Future directions include test-time adaptation and more.

Why it matters

The survey offers a unified formal framework and a systematic map for a fragmented field, making it easier to position and compare subsequent research.
As Andreas Maier's walkthrough points out, however elegant the formalism, real-world self-improvement is far smaller in scope than the "ultimate invention" vision; weight-level self-improvement and its collapse risk are the key bottleneck in judging whether agents can keep progressing.

2026-09-22 ~ 2026-09-22 · 5 related posts

Primary sources

A 97-page survey maps self-improving agents — reality is humbler than the vision — maier_ak · 2026-09-22 12:30 (UTC+08:00)
Survey formalizes agents as (θ,Σ): slow weight loop vs fast scaffold loop — maier_ak · 2026-09-22 12:30 (UTC+08:00)
Self-improving agents survey warns of model collapse and biased judge metrics — maier_ak · 2026-09-22 12:30 (UTC+08:00)

Self-improving agents, surveyed: a clean formalism and a modest reality — maier_ak · 2026-09-22
[source] Survey formalizes agents as (θ,Σ): slow weight loop vs fast scaffold loop — maier_ak · 2026-09-22
[source] Self-improving agents survey warns of model collapse and biased judge metrics — maier_ak · 2026-09-22
Survey: most self-improving agents only tune prompts and tools, rarely weights — maier_ak · 2026-09-22
[source] A 97-page survey maps self-improving agents — reality is humbler than the vision — maier_ak · 2026-09-22