ViewCraft evolves what multimodal agents see and how they read it.
ViewCraft evolves evidence interfaces and reading procedures for multimodal agents, using failures and replay to gain 6.1 points on eight benchmarks
Key Takeaways
- ViewCraft is a reusable skill approach for multimodal agents that improves task scores by 6.1 points on average across eight benchmarks without updating model weights.
- It jointly evolves an executable evidence interface and its reading procedure, helping agents turn raw observations into clearer, easier-to-use evidence.
- It learns from failure traces using replay, paired evaluation, and interface projection to keep useful revisions and discard ineffective ones.
- Its gains persist under matched input-token budgets, and learned interfaces use 8%-15% fewer input tokens on five multi-source tasks.
- The learned skills transfer across executors and outperform direct execution in all evaluated transfers.
Meta description: ViewCraft boosts multimodal agents by evolving evidence interfaces and reading procedures together, lifting task scores 6.1 points on eight benchmarks.
Tags: ViewCraft, multimodal agents, reusable skills, evidence interfaces, benchmark scores
Featured image: abstract illustration of linked panels turning raw observations into a clear view for an agent, friendly style, no real people, no logos, no text.
ViewCraft is a reusable skill approach for multimodal agents. It improves over direct execution by an average of 6.1 points in task score across eight benchmarks, without updating model weights. Its effectiveness depends on both the evidence presented and the procedure used to read it.
One skill, two coupled parts
ViewCraft jointly evolves an executable evidence interface and its reading procedure. The interface selects, transforms, and organizes observations. The reader interprets the resulting views. Treating these as one coupled skill helps the agent build evidence that is easier to use.
Keeping useful changes
Failure traces guide revisions. Replay and paired evaluation retain useful changes. Interface projection recovers visual edits from eligible rejected joint proposals. Together, these steps let the method learn from failed attempts and preserve revisions that help.
Tokens and transfer
Gains appear on every benchmark, and the advantage persists at a matched input-token budget. On all five multi-source tasks, the learned interface uses 8%-15% fewer input tokens than the initial skill. The learned skills also transfer across executors and outperform direct execution in every evaluated transfer.
The result is narrow but concrete: evidence construction and reading are presented as a coupled, reusable skill. The reported gains are 6.1 points on average across eight benchmarks, with token savings on five multi-source tasks.

Comments (0)