Microsoft and Tsinghua’s X-Coder hits 62.9% pass rate on LiveCodeBench v5
Researchers from Tsinghua University and Microsoft have developed X-Coder, an AI coding model with 7 billion parameters. This model was trained exclusively on synthetic data. A paper detailing X-Coder was posted on arXiv on January 11.
X-Coder achieved a 62.9% pass rate on LiveCodeBench v5 and a 55.8% pass rate on LiveCodeBench v6. This performance surpasses models such as DeepCoder-14B-Preview and AReal-boba2-14B, both of which have 14 billion parameters.
The development utilized SynthSmith, a data synthesis pipeline that generates programming tasks, solutions, and test cases. SynthSmith does not rely on human-written examples. The system begins by extracting coding-relevant features, including algorithms, data structures, and optimization techniques, from an initial pool of approximately 27,000 code examples. This pool is then expanded to nearly 177,000 entries through an evolutionary process.
Quality control in SynthSmith involves a dual-verification strategy. The system determines correct test outputs through majority voting among multiple candidate solutions. The best solution is then validated against a holdout test set.
The research indicated that task variety in training data contributes more to competitive programming performance than model size or solution quantity. Experiments showed that increasing the number of distinct tasks was more effective than adding multiple solutions per task.
A dataset with 64,000 different tasks, each with one solution, outperformed datasets with fewer tasks but more solutions per problem. Pass rates increased with task count: from 43.7% with 32,000 tasks to 51.3% with 64,000 tasks, then 57.2% with 128,000 tasks, and 62.7% with 192,000 tasks. The supervised fine-tuning phase achieved 60.3%, with an additional 4.6 percentage points added during reinforcement learning.
The synthetic training approach helps mitigate benchmark contamination concerns. A reference model, Qwen3-8B, showed a 30-point performance decrease between older and newer LiveCodeBench versions. X-Coder exhibited a smaller decline of 17.2 points, suggesting reduced memorization of benchmark problems.
The code for SynthSmith is available on GitHub. Researchers have stated intentions to release model weights. This work occurs as the AI industry increasingly utilizes synthetic data to address limitations in available training material. Microsoft has previously developed SynthLLM for broader synthetic data generation.