Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents

Zergham Ahmed, Kazuki Irie, Joshua B. Tenenbaum, Christopher J. Bates, Samuel J. Gershman · 2026

Read Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents by Zergham Ahmed, Kazuki Irie, Joshua B. Tenenbaum, Christopher J. Bates, Samuel J. Gershman (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-2602-00929.txt.

Description

Zergham Ahmed and colleagues present TheoryCoder-2, a Theory-Based RL agent that learns shared abstractions from practice instead of relying on hand-set ones. Humans learn abstractions and plan well to transfer across tasks, a skill hard for LLM agents and deep RL. Theory-Based RL systems like TheoryCoder show strong transfer through abstractions but depend on human-given ones, skipping abstraction learning. TheoryCoder-2 folds learned abstractions into layered planning, and tests on BabyAI, Minihack, and VGDL games show much higher sample returns than baseline LLM agents with standard planning.

Catalog record

CREATORZergham Ahmed, Kazuki Irie, Joshua B. Tenenbaum, Christopher J. Bates, Samuel J. Gershman
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53208
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.00929

Browse the full library · About the Tower of Babel