AI models recover code better from the shapes of executions

Across three program-induction domains, the authors’ geometric representation outperformed sequence-model and Fourier baselines in almost every setting, with the largest margins when training data was scarce.

Research Lab

Mohamed Ghanem · Bernd Finkbeiner

CISPA Helmholtz Center for Information Security · Technical University of Munich

Research Digest·11 November 2026·2 min read

Illustration: AI models recover code better from the shapes of executions

In one line

Path signatures encode execution traces for program induction, outperforming standard sequence models, especially with limited data.

Before you read this

An idea this paper builds on

verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.323 papers →

Read it as

Synthesizing programs from their step-by-step execution traces is important for reverse-engineering software and understanding AI agents. Standard sequence models struggle to capture the information in these traces, especially when training data is limited.

This work treats execution traces as paths and encodes them using a tool from stochastic analysis called the path signature transform. The signature compactly represents the trace's shape regardless of its length.

The log-signature encoder outperformed standard models in almost every test, with the largest gains when data was scarce. It also produced shorter valid programs than the ground truth and revealed a clear structure where each trace's states form a separate chain. This makes program induction more data-efficient and interpretable.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

path signature transform program induction rough path theory geometric representation learning verification