Harness4GenUI '26: "Toward Frontier-Quality Declarative ..."
Toward Frontier-Quality Declarative UI Generation at Small-Model Cost
Yingxiang Yang, Weihang Xiao, Ben Bullough, Tushar Deshpande, and Niresh Agarwal
(Amazon, USA)
Declarative UI protocols such as A2UI let applications generate interactive UIs by selecting pre-built components from a catalog and binding their props to application data, rather than emitting frontend code from scratch. This contract is attractive for production systems because of safety and consistency. An open question is: can low-latency and low-cost small models achieve the required quality for A2UI-based UI generation? To answer this, we systematically study three controllable design choices for catalog-conditioned A2UI generation: supervised fine-tuning (SFT) data construction method, model size, and component-catalog size. Across two React/TypeScript domains and four base checkpoints spanning two model families (Qwen 3.5 0.8B/2B/4B; SmolLM 3B), we find: (i) a 4B fine-tuned student recovers ∼98% of teacher semantic quality and ∼97% of teacher visual quality at more than an order of magnitude lower cost than frontier API calls; (ii) both augmented strategies (Perturbed-catalog and Constrained-GT) Pareto-dominate the unaugmented Full-catalog baseline, while specializing on different axes; (iii) even small models can handle and benefit from relatively large component catalog size. We distill these results into practitioner-facing trade-offs and deployment recommendations across the three design choices.
Article Search
Article: asews26harnessgenuimain-p5-p (type: Full Paper) doi:10.1145/3844177.3844556
Harness4GenUI '26: "Toward Frontier-Quality Declarative ..."
Toward Frontier-Quality Declarative UI Generation at Small-Model Cost
Yingxiang Yang, Weihang Xiao, Ben Bullough, Tushar Deshpande, and Niresh Agarwal
(Amazon, USA)
Declarative UI protocols such as A2UI let applications generate interactive UIs by selecting pre-built components from a catalog and binding their props to application data, rather than emitting frontend code from scratch. This contract is attractive for production systems because of safety and consistency. An open question is: can low-latency and low-cost small models achieve the required quality for A2UI-based UI generation? To answer this, we systematically study three controllable design choices for catalog-conditioned A2UI generation: supervised fine-tuning (SFT) data construction method, model size, and component-catalog size. Across two React/TypeScript domains and four base checkpoints spanning two model families (Qwen 3.5 0.8B/2B/4B; SmolLM 3B), we find: (i) a 4B fine-tuned student recovers ∼98% of teacher semantic quality and ∼97% of teacher visual quality at more than an order of magnitude lower cost than frontier API calls; (ii) both augmented strategies (Perturbed-catalog and Constrained-GT) Pareto-dominate the unaugmented Full-catalog baseline, while specializing on different axes; (iii) even small models can handle and benefit from relatively large component catalog size. We distill these results into practitioner-facing trade-offs and deployment recommendations across the three design choices.
Article Search
Article: asews26harnessgenuimain-p5-p (type: Full Paper) doi:10.1145/3844177.3844556
Harness4GenUI '26: "Toward Frontier-Quality Declarative ..."
Toward Frontier-Quality Declarative UI Generation at Small-Model Cost
Yingxiang Yang, Weihang Xiao, Ben Bullough, Tushar Deshpande, and Niresh Agarwal
(Amazon, USA)
Declarative UI protocols such as A2UI let applications generate interactive UIs by selecting pre-built components from a catalog and binding their props to application data, rather than emitting frontend code from scratch. This contract is attractive for production systems because of safety and consistency. An open question is: can low-latency and low-cost small models achieve the required quality for A2UI-based UI generation? To answer this, we systematically study three controllable design choices for catalog-conditioned A2UI generation: supervised fine-tuning (SFT) data construction method, model size, and component-catalog size. Across two React/TypeScript domains and four base checkpoints spanning two model families (Qwen 3.5 0.8B/2B/4B; SmolLM 3B), we find: (i) a 4B fine-tuned student recovers ∼98% of teacher semantic quality and ∼97% of teacher visual quality at more than an order of magnitude lower cost than frontier API calls; (ii) both augmented strategies (Perturbed-catalog and Constrained-GT) Pareto-dominate the unaugmented Full-catalog baseline, while specializing on different axes; (iii) even small models can handle and benefit from relatively large component catalog size. We distill these results into practitioner-facing trade-offs and deployment recommendations across the three design choices.
Article Search
Article: asews26harnessgenuimain-p5-p (type: Full Paper) doi:10.1145/3844177.3844556
Harness4GenUI '26: "The Driver Was Never the Problem: ..."
The Driver Was Never the Problem: Cross-Platform Test Portability as a Semantic Contract
Simarjot Khanna
(Independent, Canada)
Cross-platform user interface testing relies on modern drivers to execute interactions across different platforms. Because these tools operate by reading the accessibility tree, test portability silently depends on the application exposing consistent accessibility semantics across all of its surfaces. Software teams neither declare, check, nor own this cross-surface dependency. When an application changes, inconsistent accessibility semantics cause test failures that lie outside the responsibility of driver frameworks. This problem motivates treating cross-surface consistency as an explicit semantic contract. Conformance can be verified against rendered application states, per surface, independently of the test suite. Under this model, generative components ground test intent at authoring time into reviewable artifacts, rather than re-evaluating locators during every test run. We implement the check and evaluate it on Wikipedia across desktop web, mobile web, Android and iOS. Semantic agreement is achievable but not automatic: nine of twenty guarantee-surface pairs resolve. The search field carries the same accessible name on all four surfaces, while other checks reveal renamed controls, missing accessible names, vocabulary mismatches between web landmarks and native tab bars, and absent controls. Interfaces that are themselves generated sharpen the need for explicit contracts, because generated layouts lack stable structure, leaving declared meaning as the primary anchor for regression testing.
Article Search
Article: asews26harnessgenuimain-p6-p (type: Full Paper) doi:10.1145/3844177.3844557
Harness4GenUI '26: "Prompt-to-Feature Attribution ..."
Prompt-to-Feature Attribution for Explainable Software Generation
Radosław Klimek
(AGH University of Krakow, Poland)
Large language models (LLMs) enable non-technical users to generate UI prototypes from natural-language prompts. However, generated UI artifacts often lack explanations of why particular interaction features were included or omitted. This paper introduces Prompt-to-Feature Attribution, an explainability approach that links prompt fragments to generated software features through SHAP-inspired contribution scores. The approach is implemented in three prototypes with different LLM backends, attribution methods, and visualization strategies. Initial evaluation indicates that prompt-to-feature attribution helps identify influential prompt fragments, improves understanding of generated features, and supports iterative prompt refinement. These results suggest that the approach can improve the transparency and auditability of LLM-based software generation, with implications for harness engineering in generative UI systems.
Article Search
Article: asews26harnessgenuimain-p3-p (type: Short Paper (4 pages)) doi:10.1145/3844177.3844555
Harness4GenUI '26: "Toward Frontier-Quality Declarative ..."
Toward Frontier-Quality Declarative UI Generation at Small-Model Cost
Yingxiang Yang, Weihang Xiao, Ben Bullough, Tushar Deshpande, and Niresh Agarwal
(Amazon, USA)
Declarative UI protocols such as A2UI let applications generate interactive UIs by selecting pre-built components from a catalog and binding their props to application data, rather than emitting frontend code from scratch. This contract is attractive for production systems because of safety and consistency. An open question is: can low-latency and low-cost small models achieve the required quality for A2UI-based UI generation? To answer this, we systematically study three controllable design choices for catalog-conditioned A2UI generation: supervised fine-tuning (SFT) data construction method, model size, and component-catalog size. Across two React/TypeScript domains and four base checkpoints spanning two model families (Qwen 3.5 0.8B/2B/4B; SmolLM 3B), we find: (i) a 4B fine-tuned student recovers ∼98% of teacher semantic quality and ∼97% of teacher visual quality at more than an order of magnitude lower cost than frontier API calls; (ii) both augmented strategies (Perturbed-catalog and Constrained-GT) Pareto-dominate the unaugmented Full-catalog baseline, while specializing on different axes; (iii) even small models can handle and benefit from relatively large component catalog size. We distill these results into practitioner-facing trade-offs and deployment recommendations across the three design choices.
Article Search
Article: asews26harnessgenuimain-p5-p (type: Full Paper) doi:10.1145/3844177.3844556
Harness4GenUI '26: "Toward Frontier-Quality Declarative ..."
Toward Frontier-Quality Declarative UI Generation at Small-Model Cost
Yingxiang Yang, Weihang Xiao, Ben Bullough, Tushar Deshpande, and Niresh Agarwal
(Amazon, USA)
Declarative UI protocols such as A2UI let applications generate interactive UIs by selecting pre-built components from a catalog and binding their props to application data, rather than emitting frontend code from scratch. This contract is attractive for production systems because of safety and consistency. An open question is: can low-latency and low-cost small models achieve the required quality for A2UI-based UI generation? To answer this, we systematically study three controllable design choices for catalog-conditioned A2UI generation: supervised fine-tuning (SFT) data construction method, model size, and component-catalog size. Across two React/TypeScript domains and four base checkpoints spanning two model families (Qwen 3.5 0.8B/2B/4B; SmolLM 3B), we find: (i) a 4B fine-tuned student recovers ∼98% of teacher semantic quality and ∼97% of teacher visual quality at more than an order of magnitude lower cost than frontier API calls; (ii) both augmented strategies (Perturbed-catalog and Constrained-GT) Pareto-dominate the unaugmented Full-catalog baseline, while specializing on different axes; (iii) even small models can handle and benefit from relatively large component catalog size. We distill these results into practitioner-facing trade-offs and deployment recommendations across the three design choices.
Article Search
Article: asews26harnessgenuimain-p5-p (type: Full Paper) doi:10.1145/3844177.3844556