Research Library
This library curates research on text-to-app (vibe code), evaluation of multimodal LLMs and related tools, security, and extended workflows such as MCP servers.
The paper introduces app.build, an open-source framework that enables reliable large-scale AI-driven application generation through validation pipelines and environment scaffolding.
A focused evaluation of LLM agents’ web-search abilities using a standardized DuckDuckGo interface, a carefully constructed multihop question set, and an exact-match scoring scheme that isolates the final answer from explanatory text.
DesignBench is a comprehensive benchmark for evaluating Multimodal Large Language Models (MLLMs) in automated front-end engineering across multiple frameworks and tasks, including design generation, editing, and repair.
EvoAgentX is an open-source platform that automatically generates, executes, and evolves multi-agent workflows, integrating three optimization algorithms to refine prompts, tools, and topology and delivering sizable gains across reasoning, coding, math, and real-world agent tasks.
A large-scale, privacy-preserving analysis of consumer ChatGPT usage (May 2024–July 2025) showing non-work use now dominates (≈73%) and that value is delivered primarily through decision support and writing assistance.
The paper introduces SPEC, a structured intermediate representation for UI design that externalizes designer intent and enables controllable, iterative generation. Built on SPEC, the SpecifyUI system extracts specs from references, supports targeted edits, and renders high fidelity UIs with a multi agent pipeline.
A systematic review of 38 peer-reviewed studies (2022–2025) mapping how LLMs are integrated across the UI/UX lifecycle, the best practices that enable effective use, and the limitations that still constrain reliability, creativity, and adoption.
The paper proposes a method to synthesize large-scale user instructions (explicit + implicit) for GUI screens using GPT-4o and uses it to train vision-language models for GUI instruction grounding, improving performance on challenging benchmarks.
UXAgent is an LLM-agent-based framework that automates usability testing for web designs, enabling researchers to simulate thousands of user personas and collect multimodal data to iterate on study designs before human-subject studies.
ViMo is a multimodal GUI world model that predicts future mobile app interfaces as images, improving agent planning compared to text-only approaches.
The paper introduces WebGen-Bench, a new benchmark to evaluate an LLM-based agent's ability to create functional, multi-file websites from scratch, and WebGen-Instruct, a corresponding training dataset.
This paper introduces EvalGen, a mixed-initiative system that aligns LLM-assisted evaluations with human preferences by combining automated assertion generation with human grading feedback.
This paper introduces Codex, a GPT-based model fine-tuned on public GitHub code, and evaluates its ability to generate Python functions from natural language docstrings. The authors propose a new benchmark (HumanEval) and show Codex significantly outperforms GPT-3 and other models on functional correctness.
The paper introduces SPoC, a framework and dataset for translating human-written pseudocode into functionally correct C++ programs via search guided by compilation-error–driven credit assignment.