ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning

Prakhar Dixit, Tim Oates · 2026

Read ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning by Prakhar Dixit, Tim Oates (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-2606-31191.txt.

Description

Dixit and Oates present ISM, a memory system that helps a frozen language model improve at math over a stream of problems. ISM keeps a small bank of strategy schemas, short reusable plans learned from both solved and failed episodes. Symbolic tools check steps and certify answers, while hard resets separate each episode from the next.

Without changing model weights, ISM beats passive memory, retrieval, and reflection baselines on two hard math test sets, MATH-Hard and OlympiadBench. It also uses far fewer schemas than the strongest passive baseline, 64 percent fewer on one set and 86 percent fewer on the other. The work shows small, checked, and actively updated memories can support steady learning.

Catalog record

CREATORPrakhar Dixit, Tim Oates
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.31191

Browse the full library · About the Tower of Babel