Schmidhuber-Coauthored Survey: Most Self-Improving Agents Only Tweak Scaffolding, Weight Updates Rare

A new survey, Self-Improvements in Modern Agentic Systems: A Survey, coauthored by Schmidhuber (last author) with Jilin University, KAUST, University of Alberta and IDSIA, was released on arXiv in July 2026 — 97 pages covering roughly 670 references. Andreas Maier published a full walkthrough on Substack. Its core finding: most self-improving agents only modify fast-loop scaffolding; genuine weight updates are rare and carry model-collapse risks.

Confirmed

The survey formalizes agents as (θ,Σ): θ is the base model's weights, Σ the operational scaffolding 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.
Most existing self-improving agents operate only on the fast loop — tuning prompts, memory and tools while weights stay frozen; real weight updates are rare.
The survey recommends recording self-improvement via checkpointed performance curves under fixed budgets.
Two risks are flagged: self-generated data may cause model collapse, and judge-based metrics may overfit biased evaluators.
Future directions include test-time adaptation.

Why it matters

The survey offers a unified formal framework and systematic map for a crowded field, easing comparison across approaches.
As Maier notes, reality lags the 'ultimate invention' vision; weight-level self-improvement and its collapse risk are the key bottleneck for sustained agent progress.

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, 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
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