Recoleta
GitHub
Source note

Beyond pass@k: Redundancy-Aware RLVR for Multi-Sample Code Generation

Authors
Le Bronnec Florian, Alexandre Verine, Rio Yokota, Benjamin Negrevergne
Source
arXiv
Published
Collection
Software Intelligence
Code GenerationRlvrPass At KCode DiversitySoftware Foundation Models
Open arXivMarkdown source

Summary

The paper claims that multi-sample code generation needs to reduce near-duplicate programs, not only raise single-sample correctness. It adds a JPlag-based anti-redundancy reward to RLVR and reports better Pass@k results across code benchmarks.

Problem

  • Pass@k rewards a model when one of k sampled programs passes tests, but it does not measure whether the samples cover distinct implementations.
  • Correctness-only RLVR can make models repeat similar correct programs, which wastes a finite sampling budget when users ask for many candidates.
  • This matters for code intelligence systems because multi-sample generation is common in benchmark evaluation and practical program synthesis workflows.

Approach

  • The paper measures implementation redundancy with JPlag, a code-similarity tool that is less affected by variable renaming, formatting, and comments than lexical overlap.
  • It defines JPlag diversity as 1 minus mean pairwise JPlag similarity within a sampled group.
  • It compares correctness-only RLVR, Pass@k-RLVR, PKPO, and a new JPlag-RLVR objective on Qwen3-4B, Qwen3-8B, and Olmo3-7B.
  • JPlag-RLVR keeps the executable correctness reward and adds a group-level anti-redundancy reward, with leave-one-out advantages that reward samples that make the group less redundant.
  • Evaluation uses 200 sampled generations per prompt on MBPP, Code-Contest, and TACO-Cobalt.

Results

  • Across 2745 prompt-level comparisons, correctness-only Base-RLVR decreased JPlag diversity in 57.2% of cases and had mean JDiv change of -0.046, while Pass@k-RLVR increased JPlag diversity in 66.9% of cases with mean JDiv change of +0.123.
  • In the same aggregate comparison, JPlag-RLVR increased JPlag diversity in 77.4% of cases, with mean JDiv change of +0.298. It also had mean Pass@1, Pass@10, and Pass@100 changes of +16.4, +19.7, and +17.3.
  • On Qwen3-4B MBPP, JPlag-RLVR reached Pass@1 93.2, Pass@10 98.9, Pass@100 99.3, and JDiv 0.822. Base-RLVR scored 72.4, 82.4, 86.8, and 0.301 on the same metrics.
  • On Qwen3-8B MBPP, JPlag-RLVR reached Pass@10 98.5 and Pass@100 99.2, compared with Base-RLVR at 88.0 and 90.8 and PKPO at 97.1 and 99.0.
  • On Qwen3-8B TACO, JPlag-RLVR reached Pass@10 55.5 and Pass@100 68.7, compared with Base-RLVR at 48.2 and 58.5 and PKPO at 51.0 and 64.7.
  • The paper claims the JPlag reward often matches or beats specialized Pass@k-aware objectives while also producing less redundant code samples.

Link