Science – Page 119

Spotting AI Lies in Finance: A New Approach Cuts Errors by 92%

04.12.2025 by qfx

ECLIPSE demonstrates a substantial reduction in hallucination rates as coverage increases, achieving a 92% improvement at 30% coverage-reducing errors from 43.3% to 3.3% relative to entropy-only detection-and suggesting an inherent resilience to spurious outputs as system awareness expands.

A novel information-theoretic method dramatically improves the detection of inaccurate statements generated by artificial intelligence in financial contexts.

Uncovering Hidden Fraud with Quantum Graph Networks

04.12.2025 by qfx

The QTGNN framework establishes a pipeline-from graph preprocessing to fraud detection-that leverages graph neural networks to identify fraudulent activities, effectively translating relational data into actionable insights.

A new approach combines the power of quantum computing and topological data analysis to detect subtle patterns of financial crime.

Detecting Financial Fraud with the Power of Language

04.12.2025 by qfx

A new approach combines the reasoning abilities of large language models with statistical anomaly detection to uncover hidden patterns in accounting data.

Cleaning the Data Pipeline: Copyright and AI Training

04.12.2025 by qfx

The system dissects the challenge of AI-driven copyright infringement by funneling potential violations through layers of preventative measures, effectively establishing a framework to mitigate illicit content generation.

As artificial intelligence models grow in complexity, ensuring the legality of the data used to train them is becoming a critical challenge.

Taming the Chaos: Adaptive Federated Learning for Unreliable Data

03.12.2025 by qfx

$The study demonstrates that testing accuracy on the CIFAR10 dataset within a heterogeneous system is significantly impacted by both link probability ($qq$) and the proportion of compromised clients ($ \varrho$); specifically, performance degrades under LF corruption modeled via an Erdős-Rényi graph, highlighting the vulnerability of federated learning to network topology and adversarial participation.$

A new approach dynamically adjusts learning to shield decentralized systems from the impact of malicious or poorly-behaved clients.

Navigating the AI Regulation Maze

03.12.2025 by qfx

A new review charts the rapidly evolving global landscape of artificial intelligence governance, offering critical insights for organizations seeking to comply with emerging rules.

Tracing the Origins of AI Images

03.12.2025 by qfx

A multimodal large language model leverages domain-specific ontologies to extract semantic triples from images, constructing per-image knowledge graphs and quantifying semantic similarity through node overlap analysis.

A new approach links generative AI outputs to their training data, offering insights into stylistic influences and potential copyright concerns.

Can AI Really Manage Your Money?

03.12.2025 by qfx

The TAC architecture presents a framework for agent behavior, emphasizing a holistic approach to system design where interconnected components work in concert to achieve desired outcomes, as detailed in [xu2024theagentcompany].

A new benchmark assesses how well artificial intelligence agents can handle real-world financial tasks, revealing key limitations and pathways to improvement.

Beyond Open Source: DeepSeek-V3.2 Challenges Closed-World AI

03.12.2025 by qfx

DeepSeek-V3.2’s performance is benchmarked against existing models using the HMMT 2025 February competition data and the text-only subset of the HLE evaluation, establishing a comparative baseline for assessing its capabilities.

A new open-source large language model, DeepSeek-V3.2, is closing the performance gap with proprietary systems through innovations in model architecture and training techniques.

Can AI Traders Be Fooled? New Research Exposes Hidden Risks

03.12.2025 by qfx

An agent leveraging large language models for trading necessitates careful consideration of inherent vulnerabilities - encompassing risks from prompt injection and data poisoning - which demand mitigation strategies focused on robust input validation and adversarial training to ensure reliable and secure financial decision-making.

A new framework reveals that artificial intelligence-powered trading agents are surprisingly vulnerable to manipulation, potentially jeopardizing investment portfolios.