Closed-Loop Graph Algorithm Execution with Small Language Models: Step Accuracy and Rollout Reliability

Michal Podstawski · 2026

Read Closed-Loop Graph Algorithm Execution with Small Language Models: Step Accuracy and Rollout Reliability by Michal Podstawski (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-24980.txt.

Description

Michal Podstawski studies whether small language models can execute graph algorithms step by step without drifting off course. The setting is closed loop: the model repeatedly picks each next action from the current graph and its algorithmic state. Each choice changes what it sees next, so small errors can compound. His framework covers several classical graph procedures and multiple synthetic graph families, with disjoint training, validation, and test data. Both local decision quality and global behaviour are measured.

Metrics include exact rollout accuracy, constraint checks, partial credit, prefix survival, and intervention probes. Adaptation produces dependable policies for structural procedures like traversal and colouring. Weighted algorithms behave differently: they stay far more sensitive to error build up. Podstawski's broader point is about method. Strong next step prediction does not guarantee reliable solo running. Models should be judged on complete closed loop rollouts, not isolated decisions.

Catalog record

CREATORMichal Podstawski
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2606.24980

Browse the full library · About the Tower of Babel