TIPCODER: Reinforcement Learning Boosted Test-time Instruction Proposer for Code Generation

TipCoder explores a different way to obtain better code candidates: improve the guidance given before generation. Its instruction proposer learns from debugging sequences to produce problem-specific tips about constraints, edge cases, or reasoning. When answering a new problem, the system generates both an ordinary solution and a tip-guided solution, then uses a learned scoring model to choose between them. Keeping both candidates lets the selection stage account for cases where extra guidance is unhelpful. The authors report favorable benchmark comparisons under a shared selection procedure, supporting instruction variation as another way to spend additional computation on coding tasks. Study abstract.