A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models

Chain-of-Thought reasoning improves some aspects of trustworthiness but introduces new vulnerabilities: While CoT techniques help reduce hallucinations and improve interpretability, the paper finds that "cutting-edge reasoning models themselves often suffer from comparable or even greater vulnerabilities in safety, robustness, and privacy," suggesting that making AI reasoning more transparent doesn't automatically make it more trustworthy overall.

Trustworthiness in AI reasoning requires monitoring across five critical dimensions: The survey identifies truthfulness, safety, robustness, fairness, and privacy as core pillars for evaluating whether reasoning-based language models can be trusted, indicating that responsible AI development must address multiple interconnected concerns rather than focusing on a single metric.

Reasoning techniques show promise for mitigating specific harms but aren't a complete solution: The paper notes that reasoning approaches hold potential for "hallucination mitigation, harmful content detection, and robustness improvement," but these benefits are inconsistent and don't guarantee trustworthiness across all use cases or contexts.

The field lacks comprehensive understanding of how reasoning affects overall trustworthiness: The authors emphasize that "a comprehensive understanding of how CoT-based reasoning affects the trustworthiness of language models remains underdeveloped," highlighting a significant gap in current AI safety research that needs urgent attention as these systems become more widely deployed.