Imagine teaching an AI to write better code by letting it create its own practice tests and solutions. That’s the core idea behind a new technique called DSTC, or Direct Preference Learning with Only Self-Generated Tests and Code. Researchers are exploring how to improve large language models (LLMs) for code generation, particularly focusing on enhancing accuracy—that is, getting the AI to produce correct code on the first try. Traditionally, improving code LLMs involved feeding them lots of human-labeled data or using complex reinforcement learning techniques. But these methods can be expensive and resource-intensive. DSTC offers a clever alternative. The technique works by having the LLM generate multiple code solutions and accompanying tests for a given coding prompt. Then, it employs a 'minimax' selection process: it chooses the code solution that passes the most tests (the 'most correct') and the test that the fewest code solutions pass (the 'most difficult'). This creates a more reliable training pair than random selection, as it emphasizes the most successful code and the most discerning test. The chosen code and test are then combined into a single input to further refine the LLM's learning process, guiding it toward the desired behavior. This approach allows the model to learn from its own generated examples, creating a more effective and autonomous learning loop. Experiments with DSTC show impressive results. When combined with existing preference learning methods like DPO and KTO, it consistently improves the accuracy of code generated by LLMs of varying sizes. The gains are particularly noticeable in benchmarks involving more challenging tasks, demonstrating DSTC's potential to enhance performance on complex coding problems. While DSTC offers a promising new approach, it's still early days. Researchers are exploring how to further refine the minimax selection process and improve the quality of the self-generated tests. The next steps also include expanding DSTC to handle a wider range of coding tasks and programming languages. As these methods improve, we can expect code LLMs to become even more accurate and reliable, opening new doors for automation and collaboration between humans and AI in software development.