ViewCraft Jointly Evolves What Multimodal Agents See and How They Read It, +6.1 Points Across 8 Benchmarks

Binxiao Xu · hf · 2026-10-14

ViewCraft improves multimodal agents without weight updates by jointly evolving an executable evidence interface—selecting, transforming, and organizing observations—and the reading procedure that interprets the resulting views. Failure traces guide revisions; replay and paired evaluation retain useful changes; interface projection recovers visual edits from eligible rejected joint proposals.

Across eight benchmarks, ViewCraft improves task score over direct execution by an average of 6.1 points, with gains on every benchmark, and the advantage persists at matched input-token budgets. On all five multi-source tasks, the learned interface uses 8%–15% fewer input tokens than the initial skill. Skills also transfer across executors and beat direct execution in every transfer evaluated, supporting evidence construction and reading as a coupled, reusable skill.

Original post →

ViewCraft Jointly Evolves What Multimodal Agents See and How They Read It, +6.1 Points Across 8 Benchmarks

More from coding & agent