arXiv:2508.16771v3 Announce Type: replace-cross Abstract: Code Language Models (CodeLLMs) learn token importance from data correlations, whereas human developers attend selectively to semantically salient code. We present EyeMulator, a model-agnostic method that injects human visual-attention priors into CodeLLM fine-tuning without architectural changes. EyeMulator distills eye-tracking data into semantic salience and gaze-transition priors, then uses them to reweight token-level training losses. Across six backbones, two data regimes, and three CodeXGLUE tasks, the reported configurations yie
The rapid advancement of Code Language Models (CodeLLMs) has highlighted their limitations compared to human understanding, prompting innovative approaches to bridge this gap.
This development proposes a novel, model-agnostic method to significantly improve CodeLLM performance by integrating human cognitive priors, enhancing their utility and reliability.
CodeLLMs can now be fine-tuned to mimic human visual attention, potentially leading to more accurate code generation, analysis, and debugging capabilities.
- · AI developers
- · Software engineering teams
- · Cloud providers
- · AI-powered code platforms
- · Legacy code analysis tools
- · Undifferentiated CodeLLM providers
Improved performance and reliability of AI-assisted coding tools.
Reduced cognitive load and increased efficiency for human developers working with AI-generated code.
Accelerated development cycles for complex software projects, enabling new applications and services.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI