Source note
ZeroCoder: Can LLMs Improve Code Generation Without Ground-Truth Supervision?
Summary
ZeroCoder trains a model to write code and generate tests at the same time, without human-written tests or reference solutions. It uses execution outcomes between self-generated code and self-generated tests to create rewards, then improves both roles through co-evolution.
Problem
- Standard RL for code generation usually needs human-written unit tests or tests derived from reference solutions, which are costly and hard to scale.
- Self-generated tests alone are often weak, so bad code can still pass them and the reward signal becomes noisy.
- The paper asks whether an LLM can improve code generation with no ground-truth supervision at all, using only code-test execution feedback.
Approach
- ZeroCoder uses one model in two prompted roles: a coder that samples candidate programs and a tester that samples candidate tests for the same problem.
- It executes every sampled program against every sampled test to build a pass/fail matrix, then applies a selector such as MaxPass, CodeT, or B^4 to pick a consensus set of likely good solutions and tests.
- The coder gets reward for being selected into the consensus solution set. The tester gets reward for producing executable tests that both pass on a proxy-good solution and kill mutated versions of that solution, which pushes tests to be discriminative rather than trivial.
- Before RL, ZeroCoder filters out low-information training problems by keeping only instances whose pass/fail matrix has high enough rank, since low-rank matrices give weak learning signals.
- The paper also adds DyB^4, a dynamic version of the Bayesian selector that recalibrates its priors during training using as few as 10 labeled calibration instances to reduce selector drift.
Results
- In the fully label-free setting, on Qwen2.5-Coder-7B-Instruct, ZeroCoder with B^4 improves code generation by 14.5% and test generation by 15.6% relative to the base model.
- On the same model, adding DyB^4 raises the gains to 21.6% for code generation and 24.3% for test generation, and the paper says this is competitive with oracle-supervised training.
- Averaged across 3 model families and 6 benchmarks, ZeroCoder with DyB^4 improves code generation by 18.8% and test generation by 62.7% over base models.
- The method is evaluated on 4 code-generation benchmarks and 2 test-generation benchmarks. Code generation uses Pass@1 under greedy decoding; test generation uses test accuracy and mutation score.
- DyB^4 uses a calibration set of only 10 labeled instances, and the paper reports that this calibration step adds about 2% per-step wall-clock time.