When Do LLM Agents Treat Surface Noise Differently from Semantic Noise? A 68-Cell Measurement Study with a Held-Out Trace-Level Validation
Liyun Zhang; Jiayi Guo · 2026 · arXiv
WASTE classifies this as Failed Experiment Report · AI classification, approximate
An experimental approach did not work as intended — learn what to avoid before investing the same effort.
Abstract (excerpt)
We document an empirical phenomenon in chain-of-thought and ReAct agents driven by ten large language models from seven architecture families: meaning-bearing perturbations (e.g., paraphrase, synonym) alter final answers more often than presentation perturbations (e.g., formatting, reordering) of comparable severity. Across 68 cells spanning GSM8K, MATH, and HotpotQA (1,530 originals and $\sim$11,150 variants), the inconsistency gap averages +19.69 pp after severity matching (paired $t=9.58$, $p<0.0001$), with 64/68 cells positive. The gap survives four severity-proxy audits and remains signif
Excerpt shown for reference under fair use — read the full paper at the publisher.
About to run something similar?
Run an AI Precheck on your own design to catch failure modes like this one before you spend the time. Your first desk check is free.
Related failures
Leakage and the reproducibility crisis in machine-learning-based science
Negative / Null Result ReportDefining and detecting quantum speedup
Negative / Null Result ReportService robots in hotels: understanding the service quality perceptions of human-robot interaction
Negative / Null Result ReportBoosting methods for multi-class imbalanced data classification: an experimental review
Negative / Null Result ReportFINANCIAL DEVELOPMENT AND ECONOMIC GROWTH: A META‐ANALYSIS
Negative / Null Result ReportThe impact of site-specific digital histology signatures on deep learning model accuracy and bias
WASTE indexes this work — it does not host or republish it. Failure-type classification is automated and approximate.
Metadata source: arXiv