[论文解读] When Stability Fails: Hidden Failure Modes Of LLMS in Data-Constrained Scientific Decision-Making

Nazia Riasat|arXiv (Cornell University)|Mar 16, 2026

Scientific Computing and Data Management被引用 0

一句话总结

本文提出了一种受控行为评估，将稳定性、正确性、提示敏感性和输出有效性在基于LLM的数据约束科学决策任务中分离，并显示高稳定性并不保证与地面真相的一致性或输出的有效性。

ABSTRACT

Large language models (LLMs) are increasingly used as decision-support tools in data-constrained scientific workflows, where correctness and validity are critical. However, evaluation practices often emphasize stability or reproducibility across repeated runs. While these properties are desirable, stability alone does not guar- antee agreement with statistical ground truth when such references are available. We introduce a controlled behavioral evaluation framework that explicitly sep- arates four dimensions of LLM decision-making: stability, correctness, prompt sensitivity, and output validity under fixed statistical inputs. We evaluate multi- ple LLMs using a statistical gene prioritization task derived from differential ex- pression analysis across prompt regimes involving strict and relaxed significance thresholds, borderline ranking scenarios, and minor wording variations. Our ex- periments show that LLMs can exhibit near-perfect run-to-run stability while sys- tematically diverging from statistical ground truth, over-selecting under relaxed thresholds, responding sharply to minor prompt wording changes, or producing syntactically plausible gene identifiers absent from the input table. Although sta- bility reflects robustness across repeated runs, it does not guarantee agreement with statistical ground truth in structured scientific decision tasks. These findings highlight the importance of explicit ground-truth validation and output validity checks when deploying LLMs in automated or semi-automated scientific work- flows.

研究动机与目标

• Motivate the need for evaluating LLMs in data-constrained scientific workflows beyond stability.
• Introduce a controlled behavioral framework separating four decision-making dimensions: stability, correctness, prompt sensitivity, and output validity.
• Use a fixed differential expression (DE) table as a ground-truth reference to compare LLM outputs.
• Characterize common failure modes in statistical gene prioritization under varied thresholding and prompt wording.

提出的方法

• Provide a fixed DESeq2-derived differential expression table as input and query multiple LLMs (ChatGPT, Gemini, Claude) across regimes.
• Vary thresholds (strict FDR ≤ 0.05, relaxed 0.05 < FDR ≤ 0.10), borderline ranking, and minor prompt wording changes (P7a vs P7b).
• Assess outputs via four metrics: run-to-run stability (Jaccard), agreement with ground truth (Jaccard vs truth), prompt sensitivity (differences across prompts), and output validity (presence of invalid gene identifiers).
• Use deterministic prompts and 10 repeated runs per configuration to isolate model behavior from data variability.
• Provide code and results in a supplementary repository for reproducibility.

实验结果

研究问题

• RQ1 Does high run-to-run stability imply correctness with respect to a statistical ground truth?
• RQ2 How do minor prompt wording changes affect LLM decision outputs under fixed inputs?
• RQ3 What is the impact of relaxing statistical thresholds on LLM-based gene prioritization?
• RQ4 Do LLMs generate invalid or hallucinated gene identifiers even with stable outputs?

主要发现

• LLMs can show near-perfect run-to-run stability while disagreeing with ground truth.
• Small wording differences in prompts can markedly shift prioritization outcomes.
• Relaxed statistical thresholds promote over-selection or collapse rather than reliable sensitivity improvements.
• Models may produce syntactically plausible but invalid gene identifiers not present in the input, indicating output validity issues.
• Stability reflects internal robustness but does not guarantee agreement with deterministic statistical references.
• A four-dimensional evaluation framework is necessary to diagnose LLM behavior in data-constrained scientific workflows.