TipCoder Scales Test-Time Compute in Instruction Space Rather Than Sampling More Programs

arXiv 2609.03309·medium signalXBlueskyLinkedInCopy link

arXiv 2609.03309 notes that test-time scaling for code generation usually samples multiple programs from one fixed instruction, and proposes the complementary axis: many coding failures come from missing constraints, overlooked edge cases, or misleading reasoning paths induced by the original prompt. TipCoder generates problem-specific auxiliary tips before code synthesis, distilling multi-turn debugging trajectories into proactive guidance and optimizing the proposer with RL under a marginal-utility reward. At inference it produces both a base solution and a tip-guided solution and uses a reward model for post-hoc selection, an exploration-selection design that lets tips expose extra candidate potential while limiting regressions when guidance is unnecessary, comparing favorably with stochastic sampling and generic prompt optimization under a shared selection protocol.