VIEWCRAFT: Evolving What Multimodal Agents See and How They Read It
ViewCraft treats the visual evidence an agent sees and the procedure for reading it as one reusable skill.
The paper says the method evolves both an executable evidence interface and a reader, using failure traces, replay, paired evaluation, and interface projection to keep useful changes. Across eight benchmarks, it reports an average 6.1-point task-score gain over direct execution, with improvements on every benchmark. The gains remain under a matched input-token budget, and on five multi-source tasks the learned interface uses 8% to 15% fewer input tokens than the initial skill. The authors also report transfer across executors, arguing that evidence construction and reading can be reused together. HF Daily Papers' note
The paper says the method evolves both an executable evidence interface and a reader, using failure traces, replay, paired evaluation, and interface projection to keep useful changes. Across eight benchmarks, it reports an average 6.1-point task-score gain over direct execution, with improvements on every benchmark. The gains remain under a matched input-token budget, and on five multi-source tasks the learned interface uses 8% to 15% fewer input tokens than the initial skill. The authors also report transfer across executors, arguing that evidence construction and reading can be reused together. HF Daily Papers' note
score 5