Measuring the metacognition of AI
arXiv:2603.29693v1 Announce Type: new Abstract: A robust decision-making process must take into account uncertainty, especially when the choice involves inherent risks. Because artificial Intelligence (AI) systems are increasingly integrated into decision-making workflows, managing uncertainty relies more and more on the metacognitive capabilities of these systems; i.e, their ability to assess the reliability of and regulate their own decisions. Hence, it is crucial to employ robust methods to measure the metacognitive abilities of AI. This paper is primarily a methodological contribution arguing for the adoption of the meta-d' framework, or its model-free alternatives, as the gold standard for assessing the metacognitive sensitivity of AIs--the ability to generate confidence ratings that
View PDF HTML (experimental)
Abstract:A robust decision-making process must take into account uncertainty, especially when the choice involves inherent risks. Because artificial Intelligence (AI) systems are increasingly integrated into decision-making workflows, managing uncertainty relies more and more on the metacognitive capabilities of these systems; i.e, their ability to assess the reliability of and regulate their own decisions. Hence, it is crucial to employ robust methods to measure the metacognitive abilities of AI. This paper is primarily a methodological contribution arguing for the adoption of the meta-d' framework, or its model-free alternatives, as the gold standard for assessing the metacognitive sensitivity of AIs--the ability to generate confidence ratings that distinguish correct from incorrect responses. Moreover, we propose to leverage signal detection theory (SDT) to measure the ability of AIs to spontaneously regulate their decisions based on uncertainty and risk. To demonstrate the practical utility of these psychophysical frameworks, we conduct two series of experiments on three large language models (LLMs)--GPT-5, DeepSeek-V3.2-Exp, and Mistral-Medium-2508. In the first experiments, LLMs performed a primary judgment followed by a confidence rating. In the second, LLMs only performed the primary judgment, while we manipulated the risk associated with either response. On the one hand, applying the meta-d' framework allows us to conduct comparisons along three axes: comparing an LLM to optimality, comparing different LLMs on a given task, and comparing the same LLM across different tasks. On the other hand, SDT allows us to assess whether LLMs become more conservative when risks are high.
Comments: 18 pages, 5 figures, 2 tables
Subjects:
Artificial Intelligence (cs.AI)
Cite as: arXiv:2603.29693 [cs.AI]
(or arXiv:2603.29693v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2603.29693
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Richard Servajean [view email] [v1] Tue, 31 Mar 2026 12:48:42 UTC (288 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
mistralmodellanguage model
Development and multi-center evaluation of domain-adapted speech recognition for human-AI teaming in real-world gastrointestinal endoscopy
Automatic speech recognition (ASR) is a critical interface for human-AI interaction in gastrointestinal endoscopy, yet its reliability in real-world clinical settings is limited by domain-specific terminology and complex acoustic conditions. Here, we present EndoASR, a domain-adapted ASR system designed for real-time deployment in endoscopic workflows. We develop a two-stage adaptation strategy based on synthetic endoscopy reports, targeting domain-specific language modeling and noise robustness. In retrospective evaluation across six endoscopists, EndoASR substantially improves both transcrip — Ruijie Yang, Yan Zhu, Peiyao Fu
Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework
Memory emerges as the core module in the large language model (LLM)-based agents for long-horizon complex tasks (e.g., multi-turn dialogue, game playing, scientific discovery), where memory can enable knowledge accumulation, iterative reasoning and self-evolution. A number of memory methods have been proposed in the literature. However, these methods have not been systematically and comprehensively compared under the same experimental settings. In this paper, we first summarize a unified framework that incorporates all the existing agent memory methods from a high-level perspective. We then ex — Yanchen Wu, Tenghui Lin, Yingli Zhou
Human-Guided Reasoning with Large Language Models for Vietnamese Speech Emotion Recognition
Vietnamese Speech Emotion Recognition (SER) remains challenging due to ambiguous acoustic patterns and the lack of reliable annotated data, especially in real-world conditions where emotional boundaries are not clearly separable. To address this problem, this paper proposes a human-machine collaborative framework that integrates human knowledge into the learning process rather than relying solely on data-driven models. The proposed framework is centered around LLM-based reasoning, where acoustic feature-based models are used to provide auxiliary signals such as confidence and feature-level evi — Truc Nguyen, Then Tran, Binh Truong
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Models
Why I Built a Menu Bar App Instead of a Dashboard
Everyone who builds with AI eventually hits the same moment. You're deep in a coding session. Claude is flying. You're feeling productive. Then you open your API dashboard and the number hits you like a bucket of cold water. That happened to me. I don't want to talk about the exact number, but it was enough to make me stop and actually think about what I was doing. The problem wasn't that I was spending money. The problem was that I had no idea I was spending it. The dashboard problem My first instinct was what everyone does: open the Anthropic dashboard. Check the usage graphs. Try to correlate the spikes with what I was working on. But here's the thing about dashboards — they're designed for after-the-fact analysis, not real-time awareness. You go to a dashboard when something's already
Predicting 10 Minutes in 1 Square Meter: The Ultimate AI Boundary?
Can an AI predict everything that will happen in a 1-square-meter space between two human beings over the next 10 minutes? It sounds like a thought experiment from a sci-fi novel, but in the realm of predictive modeling, it is the ultimate stress test for Artificial Intelligence. To achieve this, an algorithm wouldn't just need computing power; it would need to bridge the gap between physics, biology, and the sheer chaos of human consciousness. Here is a breakdown of why this is the final frontier of predictive AI, and what it teaches us about the systems we can currently predict. 1. The Variable Explosion (The Micro Level) To predict 10 minutes of interaction, the AI must process an incomprehensible number of variables simultaneously: Biometric inputs: Heart rate variations, pupil dilatio
Mamba4 Explained: A Faster Alternative to Transformers for Sequential Modeling
Transformers revolutionized AI but struggle with long sequences due to quadratic complexity, leading to high computational and memory costs that limit scalability and real-time use. This creates a need for faster, more efficient alternatives. Mamba4 addresses this using state space models with selective mechanisms, enabling linear-time processing while maintaining strong performance. It suits tasks like [ ] The post Mamba4 Explained: A Faster Alternative to Transformers for Sequential Modeling appeared first on Analytics Vidhya .
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!