Developers spend more time reading code than writing it, yet most AI coding assistants treat code generation as a simple text stream. A new preprint by April Wang, submitted to arXiv on September 21, 2026, argues that the way code is rendered is a critical interaction primitive that directly shapes how programmers understand logic. The study moves beyond performance metrics to examine the cognitive load imposed by different rendering strategies in AI tools.

The Flaw in Token-by-Token Streaming

Current AI interfaces typically display code either instantly or token-by-token, mimicking the model's generation process rather than human reading habits. Wang argues that programmers do not read linearly; they scan for structure, skipping low-level details to grasp high-level architecture first. By forcing a linear, character-based reveal, current tools may be actively hindering comprehension by obscuring the syntactic hierarchy that developers rely on to parse complex logic.

Structured Rendering and Eye-Tracking Data

To test this hypothesis, the study introduced "structured rendering," a technique that reveals code in semantically meaningful chunks derived from its syntactic tree. This approach exposes high-level structures before drilling down into implementation details. The researchers conducted an eye-tracking study with 53 participants, comparing three modes: static display, character-based streaming, and structured rendering. The data revealed that dynamic rendering causes fewer but longer fixations, indicating sustained focus but potentially higher cognitive effort per glance.

Why Semantic Chunks Win for Comprehension

The standout finding is that structured rendering guides attention specifically toward semantically meaningful units. Unlike character-based streaming, which forces the eye to track every token, structured reveals allow the developer's gaze to jump between logical blocks, mirroring how experienced developers skim code. This suggests that AI tools should not just generate faster but render smarter, aligning visual output with the mental models developers use to debug and review code.

Key Takeaways

  • Rendering strategy is a UX feature, not just a technical constraint of model generation.
  • Structured rendering improves high-level understanding by aligning with non-linear reading patterns.
  • Dynamic rendering changes eye movement, inducing longer fixations that may increase fatigue.
  • An anonymized dataset and interactive demo are available for further research and tool development.

The Bottom Line

Stop letting your LLM dictate your UI. If you are building a coding assistant, implement structured rendering to respect how developers actually think. Token-by-token streaming is a legacy of raw model output, not a feature of good developer experience.