md-log Blog
Start free

The 'Pelican Riding' Phenomenon in AI Coding Models: When Benchmark Optimization Ignores Real Development

md-log6 min read

Introduction: The Pitfalls of Benchmark Universalism

Benchmarks like HumanEval and MBPP are widely used as representative metrics to evaluate the performance of AI coding tools. However, concerns are growing that these benchmark scores may diverge from actual development field requirements. Much like the "pelican riding" phenomenon in image generation AI, where a model over-optimizes for certain prompts (e.g., "a pelican riding a bicycle"), coding models that focus only on solving benchmark problems increasingly generate absurd code in real projects. In this article, we diagnose the problem of AI coding models being biased by benchmark performance alone and explore why developers need to establish reliable evaluation criteria for practical use.

The Temptation of Benchmark Optimization: Data Contamination and Overfitting

Coding benchmarks usually consist of small function-level problems with specific inputs and correct answers. The representative HumanEval, for example, comprises 164 Python functions, so models can quickly memorize patterns through repeated training. In fact, many models have been criticized for "data contamination"—including publicly available benchmark problems and answers in their training data. For instance, a 2023 study presented evidence that a top-performing model at the time had seen over half of HumanEval during training. This is akin to being told the exam questions in advance, making it hard to consider high scores as a reflection of the model's true problem-solving ability.

Moreover, benchmark problems are limited in difficulty and scope, making it easy for models to inflate scores through simple cherry-picking. As the "pelican riding" metaphor suggests, overfitting to specific prompts leaves them unable to properly handle other creative problems. For example, a model that scores 90% on HumanEval might still produce absurd results when calling APIs, handling complex exceptions, or performing multi-file refactoring in real company codebases. The recent AI agent breach incident at Hugging Face (July 2026) also demonstrated that models that might have performed excellently on benchmarks can fail in unpredictable ways in real environments.

Discrepancies in Real Development Scenarios: Cases of Unintended Code Generation

Models with higher benchmark scores are not necessarily safer or more productive in practice. In fact, they tend to generate code that diverges from actual requirements due to excessive optimization. For instance, when asked for a simple string processing function, the model might use overly complex regular expressions or add unnecessary dependencies because it learned from popular open-source library usage. According to a Forbes report from July 2026, many manufacturers invested heavily in AI but saw no visible returns, which can also be interpreted as a result of adopting solutions based on benchmark performance that did not fit real processes.

Another issue is that models generate code with security vulnerabilities or performance degradations without any warning. Recently, there have been increasing cases of hard-coded passwords or SQL injection vulnerabilities in AI-generated code. This blind spot arises because benchmarks assess neither security nor maintainability. Ultimately, benchmarks that only check whether code "works" cannot guarantee production quality.

Practical Guide for Developers: Building Your Own Evaluation Pipeline

To avoid these issues, developers should not rely solely on benchmark rankings but should establish evaluation metrics tailored to their project environment. Here's how:

  • Representative task sample testing: Select 10–20 coding tasks that you frequently perform in real projects, give the same prompts to multiple models, and compare the results. Evaluate comprehensively, including correctness, readability, performance, and compliance with security policies.
  • Create variant problems: Generate variant problems by slightly changing parameters or conditions of benchmark problems to see if models exhibit true problem-solving ability rather than memorized answers. For example, when asking for a sort function, requiring sorting by a special criterion instead of ascending order can reveal a model's adaptability.
  • Continuous evaluation: Run the same test suite every time the AI model version updates to observe performance trends. This enables early detection of regression in specific areas.
  • Combine with human review: Automated evaluation alone makes it difficult to judge code intent or architectural suitability. Build a feedback loop where team members review AI-generated code, incorporating insights into evaluation criteria. For instance, using a human-in-the-loop review tool like md-log, you can have humans review AI outputs and accumulate versioned histories to gauge the model's actual helpfulness.

Conclusion: Selecting AI Coding Tools with a Balanced Perspective

The advancement of AI coding models undoubtedly boosts productivity. However, selecting a tool based solely on simple benchmark scores like HumanEval is akin to choosing a car based only on fuel efficiency. Just as driving feel, safety, and maintenance costs on real roads matter more, the real measure in software development is the utility when applied to actual projects. Going forward, the ability to evaluate how well a model integrates into your codebase and workflow, rather than its benchmark performance, will become increasingly important. If you consistently review and log AI outputs using tools like md-log, you will establish a system to manage invisible quality as well.

References

#Pelican Riding#AI coding model#benchmark optimization#HumanEval#data contamination#practical evaluation

Frequently asked questions

What is the "pelican riding" phenomenon?
It refers to a phenomenon in image generation AI where overfitting to a specific prompt leads to generating similar images for other requests. In coding models, it manifests as strong performance only on certain benchmark problem types.
Doesn't a high HumanEval score mean it's a good model?
Not necessarily. HumanEval is a limited problem set and may suffer from data contamination or fail to reflect real-world complexity. A high score does not guarantee improved development productivity.
How else can I evaluate models besides benchmarks?
Create a test set from tasks that frequently occur in your project. Include variant problems, continuously track performance, and combine with manual code reviews.
Why do security issues arise in AI-generated code?
Current coding benchmarks mainly evaluate correctness and do not consider security or maintainability, so models can score highly even when generating vulnerable code. Therefore, separate security reviews are necessary.
What is md-log?
md-log is a human-in-the-loop review tool where people can conveniently review AI-generated work results via web, phone, or tablet, and each save accumulates a history. It can be useful for managing AI coding quality.
← All posts