AesCoder-4B: Code Aesthetics with Agentic Reward Feedback

AesCoder-4B is a specialized language model developed by researchers from Microsoft Research Asia, Shanghai Jiao Tong University, and Peking University, focusing on improving the aesthetic quality of code generated by Large Language Models (LLMs). While LLMs are proficient in traditional coding tasks, they often struggle with the visual aspects of code, such as webpage design.

Key Capabilities & Innovations

Code Aesthetics Optimization: AesCoder-4B is specifically trained to generate code that is not only functional but also visually appealing, addressing a common limitation of general-purpose LLMs.
AesCode-358K Dataset: The model was fine-tuned on a large-scale instruction-tuning dataset, AesCode-358K, which is dedicated to code aesthetics.
Agentic Reward Feedback: It incorporates a novel multi-agent system that evaluates code based on executability, static aesthetics, and interactive aesthetics. This feedback is integrated into a GRPO algorithm for joint optimization.
OpenDesign Benchmark: The model's performance is evaluated using OpenDesign, a new benchmark for assessing code aesthetics, where AesCoder-4B has shown strong results.
Performance: Notably, AesCoder-4B surpasses GPT-4o and GPT-4.1 in code aesthetics and achieves performance comparable to open-source models with 480B-685B parameters, despite being significantly smaller.

Primary Use Case

This model is primarily designed for webpage design and UI component generation, where code aesthetics are crucial. It provides specific system prompts for various categories like Website, Game Development, 3D Design, Data Visualization, and UI Component creation, guiding the model to produce high-quality, visually appealing HTML, CSS, and JavaScript outputs.