QuoteBench: How Matched Scores Can Hide Command-Path Failures

2026-08-13

AI Summary

Current evaluations rely solely on execution-match scores, which fail to disentangle errors arising from large language models’ command generation versus those introduced by the execution pipeline. This work proposes QuoteBench, a benchmark that isolates the impact of command generation from execution transfer through end-state validation, command replay, boundary-aware interpolation and escaping, and multi-configuration comparison across 56 single-turn tasks. The study reveals that match-based metrics often obscure substantial performance degradation; incorporating a parser reduces success rates by 55.4–73.2 percentage points; and explicitly exposing the generation contract and execution boundaries recovers 30.4–60.7 percentage points across six configurations, significantly altering model rankings.

Abstract

LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. QuoteBench measures this boundary with exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with the execution transport around one deliberately unescaped added parser. Escaping at the interpolation point reproduces each replayed reply's raw-path outcome, so any recovery under a disclosed boundary must come from the model changing its generation. Across eight same-window configurations, replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points; disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. Raw generation is nearly saturated at the frontier; boundary adaptation is what still separates models. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models: one reversal among 26 comparable pairs is unambiguous and four more sit on single-task margins. Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.

Problem

• command-path failures
• matched execution scores
• generation contract
• execution transport
• final-state validation

Innovation

• QuoteBench
• command-path failure
• generation contract
• execution boundary
• final-state validation