Primal Generation, Dual Judgment: Self-Training from Test-Time Scaling

2026-05-11

AI Summary

This work addresses the limitation of traditional code generation methods that rely solely on sparse execution feedback (e.g., pass/fail signals) and overlook the relative correctness among multiple candidate programs sampled at test time. The authors propose DuST, a novel framework that, for the first time, leverages the dual judgment space formed by these candidates to generate discriminative training signals. By employing GRPO-based reinforcement learning within a sandbox environment, DuST enables self-ranking and self-training over model-generated candidates, thereby jointly enhancing both judgment and generation capabilities. Evaluated on LiveCodeBench, the approach improves Qwen3-30B-Thinking’s judgment quality (NDCG) by 6.2, pass@1 accuracy by 3.1, and Best-of-4 accuracy by 4.1, achieving baseline Best-of-4 performance with a single generation.

Abstract

Code generation is typically trained in the primal space of programs: a model produces a candidate solution and receives sparse execution feedback, often a single pass/fail bit. Test-time scaling enriches the inference procedure by sampling multiple candidates and judging among them, but the comparative information this process reveals is discarded after inference. We argue that this information defines a dual judgment space that provides a far richer training signal: the model learns not from an isolated success or failure, but from the relative correctness structure across its own plausible attempts, identifying which succeed, which fail, and what distinguishes them. We introduce DuST (Dual Self-Training), a framework for self-training from the dual judgment space. DuST samples candidate programs from the model's own distribution, labels them through sandbox execution, retains groups containing both successes and failures, and trains the model to rank candidates by execution correctness using GRPO. The objective is purely discriminative: the model is never directly rewarded for generating correct programs. Dual self-training improves both judgment and generation. Across five models spanning two families and three scales (4B to 30B), DuST consistently improves Best-of-4 test-time scaling on LiveCodeBench. For Qwen3-30B-Thinking on LiveCodeBench v6, judgment quality improves by +6.2 NDCG, single-sample pass@1 improves by +3.1, and Best-of-4 accuracy improves by +4.1. The trained model's single rollout matches the base model's Best-of-4 performance. SFT on the same ranking data improves judgment without improving generation, confirming that on-policy RL is the mechanism that transfers dual-space learning back into primal generation.

Problem

• code generation
• test-time scaling
• self-training
• dual judgment
• execution feedback

Innovation

• Dual Self-Training
• Test-Time Scaling
• GRPO
• Code Generation
• Relative Correctness Ranking