CONTRA: Discovering and Qualifying Behavior-Changing Questions for Selective Clarification in LLM Code Generation

2026-10-01

AI Summary

This study addresses behavioral deviations in large language model code generation caused by silent assumptions, proposing CONTRA, a novel training-free framework. CONTRA generates candidate questions and combines semantic filtering with execution-based behavioral change detection to precisely identify critical ambiguities. Furthermore, it incorporates interaction history modeling to enable selective clarification, effectively balancing the necessity of clarification with development efficiency. Experimental results demonstrate that CONTRA surpasses the strongest baseline by 13.88 percentage points in F1 score on the ClarifyCodeBench benchmark, significantly outperforming existing tools such as Claude Code.

Abstract

Coding agents can generate code that appears correct but implements behavior the user never intended. This mismatch can arise when an agent silently resolves underspecified requirements through its own assumptions. As subsequent development builds on these assumptions, correcting the resulting behavior can become increasingly costly. Early clarification can help prevent such mismatches, but unnecessary questions can interrupt developers and slow down development. Existing methods struggle to identify key clarification questions while avoiding unnecessary ones. Therefore, we propose CONTRA, a training-free method that combines broad question discovery with semantic and execution-based question qualification. CONTRA first generates candidate questions and filters out those unrelated to required behavior or already resolved by the requirement. For each remaining question, it generates programs conditioned on two plausible answers and checks for stable behavioral differences on shared inputs. It then uses the interaction history to select among qualified questions or stop asking. Experiments on ClarifyCodeBench show that CONTRA achieves the highest F1 with all four coding agents, exceeding the best baseline macro-average F1 by 13.88 percentage points. With the same LLM and evaluation protocol, CONTRA also achieves higher clarification recall and F1 than the coding harnesses Claude Code and OpenHands. To support practical use, we also implement CONTRA as a Claude Code plugin that integrates selective clarification into everyday development.

Problem

• LLM code generation
• selective clarification
• underspecified requirements
• coding agents
• behavior mismatch

Innovation

• Selective Clarification
• Training-free Method
• Execution-based Qualification
• Code Generation
• Behavior-Changing Questions