Informazioni
September 2026 Harness4GenUI paper studying whether small language models can generate production-oriented A2UI from approved component catalogs. Across task-management and cloud-console domains, the authors compare training-data strategies, model sizes, catalog sizes, cost, latency, binding correctness, and rendered quality.
Sintesi
Why This Paper Adds New Signal:
1. Tests a deployable contract: Generates structured A2UI with catalog membership and resolvable data bindings instead of arbitrary frontend code
2. Quantifies the cost-quality trade-off: A fine-tuned 4B student recovers about 98% of teacher semantic quality and 97% of visual quality at over an order of magnitude lower API cost
3. Compares training recipes: Catalog perturbation improves overall quality, while constrained targets produce stronger data-binding reliability
4. Challenges catalog-size intuition: Models from 0.8B to 4B improve as the available catalog grows from roughly 10 to 86 components
5. Includes cross-domain evidence: Replicates the main strategy trade-offs with a different 3B base model and cloud-management interface catalog
1. Tests a deployable contract: Generates structured A2UI with catalog membership and resolvable data bindings instead of arbitrary frontend code
2. Quantifies the cost-quality trade-off: A fine-tuned 4B student recovers about 98% of teacher semantic quality and 97% of visual quality at over an order of magnitude lower API cost
3. Compares training recipes: Catalog perturbation improves overall quality, while constrained targets produce stronger data-binding reliability
4. Challenges catalog-size intuition: Models from 0.8B to 4B improve as the available catalog grows from roughly 10 to 86 components
5. Includes cross-domain evidence: Replicates the main strategy trade-offs with a different 3B base model and cloud-management interface catalog
Tag
a2uismall-language-modelssupervised-fine-tuningcomponent-catalogsdeployment-cost