Learning from Reasoning Failures via Synthetic Data Generation

Gabriela Ben Melech Stan, Estelle Aflalo, Avinash Madasu, Vasudev Lal, Phillip Howard · 2025

Read Learning from Reasoning Failures via Synthetic Data Generation by Gabriela Ben Melech Stan, Estelle Aflalo, Avinash Madasu, Vasudev Lal, Phillip Howard (2025) 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-2504-14523.txt.

Description

Gabriela Ben Melech Stan and co-authors present a 2025 method for teaching large multimodal models from their own reasoning failures. Such models work with both images and text, but good paired training data are scarce. Making data in bulk does not fix the exact skills a given model lacks.

Instead, the team first finds where a model reasons badly, then creates synthetic training examples aimed at those gaps. The approach mirrors how a tutor targets a student's weak spots rather than assigning general practice. For model builders, the paper offers a way to spend a training budget where it should help most.

Catalog record

CREATORGabriela Ben Melech Stan, Estelle Aflalo, Avinash Madasu, Vasudev Lal, Phillip Howard
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-14523

Browse the full library · About the Tower of Babel