Evolving Programmatic Skill Networks

A continual-learning agent whose skill graph develops through mechanisms borrowed from neural-network training. It acquires, refines, and reuses a growing repertoire of skills in open-ended worlds, without forgetting what came before.

Haochen Shi1,2 · Xingdi Yuan3,* · Bang Liu1,2,4,*

1Université de Montréal · 2Mila · 3Microsoft Research · 4Canada CIFAR AI Chair   *Equal advising

arXiv 2601.03509 Code BibTeX
The shared quest

Building agents for open-ended worlds

Open-ended Minecraft world — the agent's playground.
Embodied agents must continually acquire, refine, and reuse a growing repertoire of skills. The central challenge isn't just learning skills, it's continually reorganizing and improving them in dynamic, open-ended environments, without forgetting what came before.
The proposal

What if programs could learn like neural networks?

Today's LLM agents store skills as flat libraries (no composition) or static hand-built graphs (no learning), neither of which reorganizes as new tasks arrive. We've mastered learning in continuous, parameterized systems, so what if we applied the same optimization principles to a network of discrete, symbolic programs?

Flat library

Skills indexed by similarity. Voyager-style. Cannot compose; new tasks don't decompose into existing skills.

Static graph

Hand-authored library, fixed at design time. ODYSSEY-style (183 skills). No learning, reuses within its fixed repertoire, but cannot grow and adapt to novel tasks.

Evolving network → ours

Skills as programs in a directed graph that continually reorganizes through symbolic updates as the agent learns.

Neural network training
Backpropagation

credit flows along the activation path

Programmatic skill network
Symbolic credit assignment

REFLECT diagnoses each skill on the trace

Neural network training
Learning-rate scheduling

freeze converged layers, keep new ones plastic

Programmatic skill network
Adaptive stabilization

V(s) gates updates per skill

Neural network training
Architecture search

restructure topology to improve capacity

Programmatic skill network
Structural refactoring

merge, extract, prune skills under rollback

The analogy is partial (PSN works over discrete programs and binary success signals), but the algorithmic structure of NN training transfers, and yields testable predictions that the paper verifies: chain-rule-like propagation depth, retention without oscillation, and stair-stepped structural growth.

Results

PSN masters the tech tree faster and more reliably

Tech tree mastery on GPT-5-mini: PSN reaches the diamond pickaxe in 32±10 iterations across 6/6 runs; Voyager 2/6.

On GPT-5-mini, PSN unlocks the diamond tools in 32 ± 10 iterations across 6/6 runs (44 ± 13 on Qwen3-Coder-Next); a flat-library agent (Voyager) reaches them in 1–2/6. The paper reports the full picture: Crafter cumulative reward, per-milestone skill retention, and the complete baseline comparison (ReAct, Reflexion, ADAM, ODYSSEY).

Explore a learned skill network

An interactive view of a PSN run

A real PSN run on GPT-5-mini. Each node is a JavaScript skill; edges are parent → child reuse links; node color encodes norm V. Low-norm-V nodes are skills whose uncertainty term still exceeds estimated success, i.e., legitimate plastic skills. Click any node for code, optimization history, preconditions, and effects.

low norm V high norm V

Loading skill graph data…