ETCHR: Editing To Clarify and Harness Reasoning

Researchers introduced ETCHR, a two-stage, question-conditioned image editing model designed to make images easier for downstream reasoning.

The system decouples editing from understanding and is trained to clarify visual content rather than answer the question directly.

Across five visual reasoning task families, ETCHR improved Pass@1 scores on multiple multimodal models.

The gains were seen on models including Qwen3-VL-8B, Gemini-3.1-Flash-Lite, and Kimi K2.5, showing broad compatibility.

It offers a training-free way to boost visual reasoning accuracy across different tasks and model families.