Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SELF-REDRAFT: Eliciting Intrinsic Exploration-Exploitation Balance in Test-Time Scaling for Code Generation

Yixiang Chen, Tianshi Zheng, Shijue Huang, Zhitao He, Yi R. Fung

TL;DR

Self-Redraft tackles the problem of achieving an intrinsic balance between exploration and exploitation in test-time scaling for code generation without execution feedback. It extends Self-Refine by explicitly prompting drafting when flaws are detected, enabling a hybrid search strategy under an execution-free setting. On LiveCodeBench across six models, Self-Redraft yields modest gains over Self-Refine but does not reach the potential suggested by the pass@8 upper bound, highlighting gaps in the model's self-guided exploration. The study identifies bottlenecks in feedback quality and discriminative judgment, reveals model-specific balancing behavior, and establishes a baseline for future work in improving critique, adaptation, and exploration strategies for robust, real-world code generation.

Abstract

Test-time scaling without interpreter feedback is essential for real-world code generation scenarios where test cases are not readily available. While existing paradigms often rely on either greedy exploitation (i.e., iterative refinement) or stochastic exploration (i.e., relying on sample-based voting or reranking mechanisms), the balance between these two dimensions remains underexplored. To investigate the LLM's intrinsic ability to balance exploitation and exploration, we introduce SELF-REDRAFT, a framework built upon Self-Refine that encourages the model to propose new drafts for solutions that are fundamentally flawed. Our results show that SELF-REDRAFT consistently achieves better performance than Self-Refine when converged under the same maximum number of iterations. Still, we observe that significant room for improvement remains, largely due to two core aspects of current self-redraft capabilities: constrained capacity for generating instructive feedback and fragile discriminative judgment. We also find that balancing strategies vary notably across different LLMs, reflecting distinct, model-specific behaviors. Overall, our study establishes a baseline for intrinsic exploration-exploitation balancing in test-time scaling and identifies feedback and discrimination as key areas with potential for future advances.

SELF-REDRAFT: Eliciting Intrinsic Exploration-Exploitation Balance in Test-Time Scaling for Code Generation

TL;DR

Self-Redraft tackles the problem of achieving an intrinsic balance between exploration and exploitation in test-time scaling for code generation without execution feedback. It extends Self-Refine by explicitly prompting drafting when flaws are detected, enabling a hybrid search strategy under an execution-free setting. On LiveCodeBench across six models, Self-Redraft yields modest gains over Self-Refine but does not reach the potential suggested by the pass@8 upper bound, highlighting gaps in the model's self-guided exploration. The study identifies bottlenecks in feedback quality and discriminative judgment, reveals model-specific balancing behavior, and establishes a baseline for future work in improving critique, adaptation, and exploration strategies for robust, real-world code generation.

Abstract

Test-time scaling without interpreter feedback is essential for real-world code generation scenarios where test cases are not readily available. While existing paradigms often rely on either greedy exploitation (i.e., iterative refinement) or stochastic exploration (i.e., relying on sample-based voting or reranking mechanisms), the balance between these two dimensions remains underexplored. To investigate the LLM's intrinsic ability to balance exploitation and exploration, we introduce SELF-REDRAFT, a framework built upon Self-Refine that encourages the model to propose new drafts for solutions that are fundamentally flawed. Our results show that SELF-REDRAFT consistently achieves better performance than Self-Refine when converged under the same maximum number of iterations. Still, we observe that significant room for improvement remains, largely due to two core aspects of current self-redraft capabilities: constrained capacity for generating instructive feedback and fragile discriminative judgment. We also find that balancing strategies vary notably across different LLMs, reflecting distinct, model-specific behaviors. Overall, our study establishes a baseline for intrinsic exploration-exploitation balancing in test-time scaling and identifies feedback and discrimination as key areas with potential for future advances.

Paper Structure

This paper contains 26 sections, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Test-time Scaling with Self-Redraft
  4. Dataset and Models
  5. Dataset.
  6. Models.
  7. Experiments and Analysis
  8. Main Results
  9. Unexploited Potential in Exploration
  10. Insufficient Model Critique for Methodological Correction
  11. Blinded Evaluation.
  12. Fragile Code Discrimination Leading to Deleterious Redrafts
  13. Cross-Model Inconsistency
  14. Conclusion
  15. Self-Redraft
  16. ...and 11 more sections

Figures (8)

  • Figure 1: Our proposed Self-Redraft framework.
  • Figure 2: Detailed benchmark performance of LLMs on LiveCodeBench, evaluated with 16 iterations of Self-Refine and 16 iterations of Self-Redraft.
  • Figure 3: Comparison of Self-Redraft×16 and pass@8 accuracies on LiveCodeBench.
  • Figure 4: Recall on Draft versus absolute improvement of Self-Redraft x16 over Self-Refine x16.
  • Figure 5: Recall on Draft as annotated by three auxiliary models: GPT-5 mini, GLM-4.6 and Grok 4 Fast.
  • ...and 3 more figures