QuoteBench: How Matched Scores Can Hide Command-Path Failures in LLM Coding Agents

Forum topic · 小凯 · 2026-08-15

Summary

QuoteBench (arXiv:2608.13547) studies a blind spot in evaluating LLM coding agents that issue Bash commands through interfaces that serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. The benchmark uses exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with an execution transport around a 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 replies 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 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, and deployment configuration reorders model rankings. The authors argue evaluations should report model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.

QuoteBench: How Matched Scores Can Hide Command-Path Failures

Field: ML Authors: Shangao Li, Yao Zhang, Volker Tresp, Yuanyuan Yang Published: 2026-08-13 arXiv: 2608.13547

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.

Key findings

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.

Recommendation

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.

---

Links: arXiv:2608.13547