X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests

Abstract: Competitive programming presents great challenges for Code LLMs due to its intensive reasoning demands and high logical complexity. However, current Code LLMs still rely heavily on real-world data, which limits their scalability. In this paper, we explore a fully synthetic approach: training Code LLMs with entirely generated tasks, solutions, and test cases, to empower code reasoning models without relying on real-world data. To support this, we leverage feature-based synthesis to propose a novel data synthesis pipeline called SynthSmith. SynthSmith shows strong potential in producing diverse and challenging tasks, along with verified solutions and tests, supporting both supervised fine-tuning and reinforcement learning. Based on the proposed synthetic SFT and RL datasets, we introduce the X-Coder model series, which achieves a notable pass rate of 62.9 avg@8 on LiveCodeBench v5 and 55.8 on v6, outperforming DeepCoder-14B-Preview and AReal-boba2-14B despite having only 7B parameters. In-depth analysis reveals that scaling laws hold on our synthetic dataset, and we explore which dimensions are more effective to scale. We further provide insights into code-centric reinforcement learning and highlight the key factors that shape performance through detailed ablations and analysis. Our findings demonstrate that scaling high-quality synthetic data and adopting staged training can greatly advance code reasoning, while mitigating reliance on real-world coding data.

• The paper introduces the SynthSmith pipeline that fully synthesizes competitive programming tasks, solutions, and tests without real-world data.
• The study demonstrates that synthetic-data-trained models outperform larger RL baselines with significant improvements in pass rates and generalization.
• Dual-verification and long chain-of-thought techniques are highlighted as key factors for enhancing code LLM performance in rigorous reasoning tasks.

Fully Synthetic Data for Advancing Code Reasoning: The X-Coder Framework

Introduction

The paper "X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests" (2601.06953) addresses the critical bottleneck faced by modern code LLMs in the domain of competitive programming: the scarcity of large-scale, challenging, and diverse reasoning-focused datasets. The reliance of current state-of-the-art models on real-world or evolved-from-real data fundamentally limits their scalability, diversity, and generalization capacity. This work departs radically from the conventional paradigm by demonstrating that code-centric LLMs can achieve highly competitive results using entirely synthetic data: tasks, solutions, and comprehensive test suites generated and verified without recourse to any real coding benchmarks or human-written seeds.

The SynthSmith Data Synthesis Pipeline

A pivotal contribution is the SynthSmith pipeline, designed for systematic, feature-driven synthesis of competitive programming tasks, verified solutions, and robust test cases. SynthSmith incorporates a multi-stage pipeline:

1. Feature Extraction and Evolution: Competitive programming-relevant features (algorithms, data structures, implementation logic, etc.) are initially extracted using LLMs from curated datasets (e.g., TACO). Features are diversified and refined through evolution along both breadth and depth dimensions, resulting in a large, hierarchical feature set tailored for competitive scenarios.
2. Task Generation: Rather than composing problems from broad or generic concepts, SynthSmith selects mutually consistent feature subsets (via an explicit selection stage) and integrates them into task scenarios matching various competitive programming styles (Codeforces, AtCoder, LeetCode). A two-stage process ensures that tasks remain non-trivial and genuinely require sophisticated multi-step reasoning.
3. Test Case Generation: The framework leverages both prompting-based and tool-based strategies (using CYaRon) to automatically generate input cases spanning standard, edge, and stress conditions.
4. Solution Synthesis and Dual-Verification: Multiple candidate solutions per task are generated with state-of-the-art LLMs. A dual-verification protocol is then applied: consensus voting yields reliable test outputs, and solutions are filtered using a hold-out validation split with difficult test cases specially weighted, ensuring both robustness and coverage.

SynthSmith is specifically engineered for scalability and high fidelity, yielding datasets suitable for both supervised fine-tuning (SFT) and reinforcement learning (RL) stages.

The X-Coder Model Series and Empirical Results

Leveraging the SynthSmith pipeline, the authors introduce X-Coder, a model series trained exclusively with synthetic data under an SFT-then-RL regime. The models use robust open-source backbones (Qwen2.5-Coder-7B-Instruct, Qwen3-8B-Base), and employ both SFT on long chain-of-thought (CoT) synthetic trajectories and RL with group-based relative policy optimization (GRPO).

Benchmark Performance

• On the LiveCodeBench v5 and v6 benchmarks, X-Coder-7B achieves an avg@8 pass rate of 62.9 (v5) and 55.8 (v6), significantly outperforming larger RL baselines trained on real or mixed data (e.g., DeepCoder-Preview-14B and AReal-boba2-14B, pass rates <59).
• X-Coder consistently outperforms SFT or RL baselines at equivalent or larger parameter scales, and robustly generalizes across different competitive programming styles and difficulty levels.
• Scaling experiments show strong scaling laws with synthetic SFT data: increasing the diversity of unique tasks yields larger generalization gains than merely augmenting the number of solutions per task. A 21% gain over comparable synthetic approaches (e.g., EpiCoder) is demonstrated in competitive settings.

RL Insights

• The RL phase provides an additional 4-5% performance over strong SFT initializers, with a pronounced "good-gets-better" effect: superior SFT initialization directly correlates with higher RL ceilings.
• RL on synthetic (and thus noisier) test data remains effective, suggesting RL can be decoupled from the need for perfectly clean human-curated evaluation signals.

Ablation Studies and Analyses

A comprehensive set of ablations and analyses isolate the critical factors influencing code reasoning model quality:

• Dual-Verification: Models trained on dual-verified solutions substantially outperform those relying on non-verified data (raw solution pass rates are significantly lower).
• Reasoning Length: Long-CoT samples (requiring a deep sequence of intermediate reasoning steps) yield much stronger code LLMs than short-CoT analogs (17%+ absolute improvement).
• Task Style and Diversity: While AtCoder-style tasks present slightly higher benefit, overall, task diversity resulting from multi-style generation and feature-driven construction dominates gains.
• Test Case Generation: Tool-based test generation (CYaRon) outperforms prompt-based methods across reliability, challenging-case coverage, and discriminative power.
• Data Selection: Selection schemes prioritizing tasks that induce longer reasoning chains (rather than merely high-difficulty as rated by external classifiers) yield better downstream performance.

Failure Modes and Cognitive Observations

Detailed analysis of errors reveals that the key bottleneck in code LLMs remains deep, multi-step reasoning. Mistakes are primarily "Wrong Answer" outputs, with additional weaknesses in efficiency (Time Limit Exceeded) and output completeness (truncated solutions under context window constraints). Notably, longer reasoning traces correlate with greater task difficulty but lower pass rates, confirming the trade-off between problem depth and current model reasoning capacity.

Case studies illustrate that X-Coder exhibits advanced cognitive behaviors (planning, backtracking, reflection) distilled during SFT but not fundamentally improved in diversity or robustness by the RL phase. Reward hacking and retrieval-style artifacts appear as model scale and RL training intensity increase, indicating the need for future mitigation techniques.

Implications and Future Directions

This work demonstrates that fully synthetic data is viable for scaling rigorous code reasoning abilities in LLMs without reliance on real-world or contaminated data. The practical upshot is a reduction in data leakage risks and a path toward scalable, updatable, and custom-tailorable model training protocols decoupled from the limitations of existing public problem corpora. Furthermore, the tools and pipelines introduced (especially SynthSmith) set a new standard for data-centric development in code model research and can be applied or extended to other structured reasoning domains.

Future directions include improvements to synthetic generation fidelity (harder tasks, better error detection, adversarial scenario generation), scaling to larger architectures with stronger base capabilities, and tighter integration of reward modeling to further reduce susceptibility to spurious failure modes or reward hacking. Enhancing the diversity and sophistication of test-case synthesis, along with fine-grained performance attribution, could further close remaining gaps between synthetic and real-world code reasoning.

Conclusion

"X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests" (2601.06953) provides a rigorous demonstration that high-level code reasoning models can be trained and robustly evaluated entirely with synthetic corpora. Through sophisticated feature-driven synthesis, robust dual-verification of solutions, and systematic ablation studies, the authors establish new best practices for data-centric scaling of code LLMs and lay a robust foundation for the evolution of competitive-programming expertise without legacy data dependencies.