Recoleta
GitHub
Source note

BACE: LLM-based Code Generation through Bayesian Anchored Co-Evolution of Code and Test Populations

Authors
Kaushitha Silva, Srinath Perera
Source
arXiv
Published
Collection
Software Intelligence
LLM Code GenerationProgram SynthesisBayesian InferenceTest GenerationCo Evolution
Open arXivMarkdown source

Summary

BACE improves LLM code generation by treating generated tests as noisy evidence instead of ground truth. It co-evolves code and test populations with Bayesian belief updates, while keeping the search tied to public example cases.

Problem

  • LLM-generated code often has logic errors that single-pass prompting misses, so systems use feedback loops with generated tests.
  • Generated tests are unreliable: wrong code can pass weak tests, and correct code can be pushed into wrong edits by faulty assertions.
  • Recent systems such as MapCoder and CodeSIM avoid test generation because of this failure mode, which leaves useful execution signal on the table.

Approach

  • BACE keeps a population of candidate programs and a population of generated tests instead of relying on one code sample and one test suite.
  • Each code candidate and each test gets a Bayesian belief score: the probability it is correct or valid. Pass/fail outcomes update those beliefs as noisy observations rather than hard truth.
  • The noise model uses three pass probabilities for misleading cases: false pass for valid code on broken tests (α\alpha), accidental pass for wrong code on valid tests (β\beta), and coincidental pass for wrong code on broken tests (γ\gamma).
  • Public input/output examples from the prompt are fixed as high-confidence anchors. Code that fails an anchor gets belief driven to zero; code that passes anchors gets a strong positive update.
  • BACE alternates evolution of tests and code, and keeps diverse high-belief individuals through behavior-based elitism and differential testing.

Results

  • On LiveCodeBench v6 (post-March 2025), BACE outperforms CodeSIM by 3.8% with GPT-5-Mini.
  • On the same benchmark, BACE outperforms CodeSIM by 5.0% with Qwen2.5-Coder-7B.
  • The paper claims better performance than leading multi-agent frameworks across both proprietary models and open-weight small language models.
  • The excerpt does not provide full absolute scores, pass@k values, variance, or a complete baseline table.
  • The authors also state that AgentCoder underperformed direct prompting on medium problems with GPT-5-Mini, but the excerpt gives no number for that gap.

Link