Measuring Free-Form Decision-Making Inconsistency of Language Models in Military Crisis Simulations (2410.13204v1)

Abstract: There is an increasing interest in using LMs for automated decision-making, with multiple countries actively testing LMs to aid in military crisis decision-making. To scrutinize relying on LM decision-making in high-stakes settings, we examine the inconsistency of responses in a crisis simulation ("wargame"), similar to reported tests conducted by the US military. Prior work illustrated escalatory tendencies and varying levels of aggression among LMs but were constrained to simulations with pre-defined actions. This was due to the challenges associated with quantitatively measuring semantic differences and evaluating natural language decision-making without relying on pre-defined actions. In this work, we query LMs for free form responses and use a metric based on BERTScore to measure response inconsistency quantitatively. Leveraging the benefits of BERTScore, we show that the inconsistency metric is robust to linguistic variations that preserve semantic meaning in a question-answering setting across text lengths. We show that all five tested LMs exhibit levels of inconsistency that indicate semantic differences, even when adjusting the wargame setting, anonymizing involved conflict countries, or adjusting the sampling temperature parameter $T$. Further qualitative evaluation shows that models recommend courses of action that share few to no similarities. We also study the impact of different prompt sensitivity variations on inconsistency at temperature $T = 0$. We find that inconsistency due to semantically equivalent prompt variations can exceed response inconsistency from temperature sampling for most studied models across different levels of ablations. Given the high-stakes nature of military deployment, we recommend further consideration be taken before using LMs to inform military decisions or other cases of high-stakes decision-making.

• The paper introduces a BERTScore-based method to quantify free-form decision inconsistency in military crisis simulations.
• It shows that all tested LMs, including GPT-4, exhibit significant inconsistency regardless of escalation levels.
• Findings underscore LM prompt sensitivity and highlight the need for rigorous evaluations before deployment in high-stakes settings.

Measuring Free-Form Decision-Making Inconsistency of LLMs in Military Crisis Simulations

In the context of evolving interest in autonomous decision-making, the use of LMs in critical settings, such as military crisis simulations, raises pertinent questions regarding their reliability and consistency. The paper "Measuring Free-Form Decision-Making Inconsistency of LLMs in Military Crisis Simulations" by Shrivastava et al. explores the inconsistency of LMs when tasked with decision-making in high-stakes environments, specifically through the lens of wargame simulations akin to tests performed by the US military.

Overview

The paper underscores the need to evaluate the dependability of LMs' decision-making, highlighting the inherent challenges in measuring semantic differences and evaluating natural language responses free from pre-defined choices. The paper introduces a robust method using BERTScore to quantify the inconsistency of free-form responses, facilitating a comprehensive analysis of LM behavior across various military crisis scenarios.

Methodology

The authors implement an initial setting experiment and a continuations experiment to assess how different escalation levels affect LM decision-making. They scrutinize responses from five LMs across twenty simulations, focusing on semantic consistency rather than syntactic conformity. Anonymization of conflict countries is also tested to determine its effect on model outputs.

The BERTScore-based metric is validated to ensure it effectively captures semantic differences while minimizing emphasis on lexical or syntactical variations, thus providing a reliable measure of inconsistency.

Key Findings

1. Inconsistency Across Models: All tested LMs exhibited significant inconsistency in decision-making, with GPT-4 showing the most variance. The results highlight behavioral differences between models, emphasizing the unpredictability of LM recommendations.
2. Impact of Escalation: The degree of escalation did not significantly affect inconsistency levels, pointing to persistent inconsistency regardless of situational variations.
3. Prompt Sensitivity: Variations in prompt phrasing induced inconsistency levels comparable to high-temperature sampling, underscoring the sensitivity of LMs to prompt structure.
4. Temperature Variations: Temperature adjustments influenced response variability, but inconsistency remained evident even at lower temperatures.

Implications and Future Work

The findings suggest cautious deliberation before deploying LMs in high-stakes military settings. The propensity for inconsistency could lead to unpredictable decision-making, posing risks in sensitive environments. The paper advocates for enhanced scrutiny of LM behavior and encourages the integration of robust evaluation mechanisms before wider military application.

Further research should explore diverse crisis scenarios and involve fine-tuned models to better understand the nuances of LM decision-making in military simulations. Additionally, the development of alternative metrics to complement BERTScore could refine the assessment framework, ensuring comprehensive evaluations of semantic consistency.

In conclusion, the deployment of LMs for military decision-making remains fraught with challenges. This research marks a significant step toward understanding these complexities, laying the groundwork for future explorations into AI governance and safety in high-risk domains.