Imagine a world where submitting your coding assignment means getting instant feedback, not from a human TA, but from an AI. This isn't science fiction, it's getting closer to reality thanks to research into how Large Language Models (LLMs) can evaluate code. Traditionally, we test code with pre-written test cases. But what if those tests miss edge cases or the task, like web scraping, is hard to test automatically? Researchers are exploring ways LLMs can step in and act like human code reviewers, looking at code for semantic correctness (does it *actually* do what it’s supposed to), even without test cases. One promising approach is "CodeJudge." CodeJudge guides LLMs through a "slow thinking" process—first analyzing what the code should do based on the instructions, then examining the code itself step by step to identify potential errors, similar to how a human developer would do code review. CodeJudge can even identify common mistakes and categorize them by severity, from minor issues like missing import statements to major logic errors. Experiments show CodeJudge does a surprisingly good job, outperforming traditional methods in most cases. Even smaller, open-source LLMs can achieve decent results with CodeJudge, making the technology more accessible. But CodeJudge isn't perfect. It can struggle with truly complex code, highlighting the challenge of getting AI to understand nuanced logic. And like a strict code reviewer, it can be over-conservative, flagging minor error-handling issues even when the core functionality is right. This research is still in its early days, but it hints at a future where AI can provide helpful code feedback instantly, complementing traditional testing. This could not only help students and beginner coders, but also accelerate AI alignment efforts which often rely on human evaluations of LLM outputs.