Verification Trap: Understanding Test-Time Selection Failures under False Premises in Code Generation

2026-10-04

AI Summary

This study addresses the "verification trap" in code generation, a phenomenon wherein generators and verifiers sharing erroneous premises cause test-time selection mechanisms to fail. This work is the first to formally define and quantify this issue, revealing the limitations of coupled evidence and proposing decoupled evidence as a core mitigation strategy. The approach is systematically validated through multi-benchmark evaluations, comparisons across five code models, and a lightweight predictor that operates without gold-standard labels. Results demonstrate that the verification trap significantly degrades generation accuracy, while the proposed predictor achieves an AUROC of 0.846. Furthermore, introducing premise-agnostic auditors effectively recovers oracle-level performance, establishing a new paradigm for enhancing the robustness of code generation systems.

Abstract

Test-time compute has become a central way to improve code generation: systems sample multiple candidate programs and use verifier-visible evidence to select the final output. This paradigm implicitly assumes that the verifier provides a corrective signal independent from the generator. We challenge this assumption under misleading task premises. When the generator and verifier share a false premise, they become coupled through a mistaken belief: the generator produces premise-consistent shortcuts, while the verifier supplies evidence that fails to expose them. Consequently, the selector may choose a hidden-test-wrong candidate even when a hidden-test-correct program exists in the pool. We call this failure mode Verification Trap. Across three code-generation benchmarks and five code models, false premises consistently degrade first-sample correctness, reduce selector-chosen correctness after 64-sample test-time selection, and amplify recoverable mis-selection. Mechanistically, verifier-written tests inherit the premise-level blind spot, reshaping verifier-visible candidate space away from hidden-test correctness. These traces make Verification Trap predictable before hidden execution: a lightweight gold-free predictor using verifier-visible features reaches 0.846 AUROC. Our results identify decoupled evidence as a key mitigation axis: coupled scaling provides limited recovery, whereas premise-agnostic robustness auditors recover substantial oracle headroom.

Problem

• code generation
• test-time selection
• verification trap
• false premises
• verifier-generator coupling

Innovation

• Verification Trap
• Test-Time Selection
• Code Generation
• False Premises
• Decoupled Evidence