Science – Page 156 – Investment Policy

Can AI Write Secure Code? A New Benchmark Evaluates LLM Performance.

30.11.2025 by Ray Dalio

DualGauge establishes a benchmarking system designed to rigorously evaluate the security of code generated by automated systems, providing a unified framework for assessing vulnerabilities and robustness.

Researchers have developed an automated system to rigorously assess both the security and functional correctness of code generated by large language models.

The Limits of Efficient Algorithms for Identifying Code

30.11.2025 by Ray Dalio

The construction of graphGG relies on a specific set of tools and components, meticulously assembled to facilitate its creation and functionality.

A new proof demonstrates that the long-studied Identifying Code problem is unlikely to have a fast, practical solution, resolving a key question in parameterized complexity.

Seeing Through Camouflage: A Reference-Free Detection Approach

30.11.2025 by Ray Dalio

RefOnce distinguishes itself from existing reference-based approaches by retrieving contextual information directly from a prototype memory, enabling inference without explicit reference images - a departure that addresses limitations inherent in methods relying on pre-defined correspondences.

Researchers have developed a new framework that eliminates the need for reference images in detecting camouflaged objects, relying instead on distilled category knowledge and adaptive feature alignment.

When Wordplay Breaks AI: What Language Models Struggle to Spell

30.11.2025 by Ray Dalio

$Model performance on puzzle solving demonstrates a capacity-dependent sensitivity to difficulty, with smaller models-like Qwen-4B, exhibiting a $19\times$ performance drop from easy to hard puzzles-showing far greater variation than larger models such as GPT-5-mini ($2.5\times$), though calibration strength remains modest across all models-ranging from $r=0.24$ to $0.38$-with proprietary models displaying slightly better correlation.$

New research explores how large language models perform on tasks requiring strict adherence to orthographic rules, revealing surprising weaknesses despite their fluency.

Can Language Models Handle a Reworded Question?

30.11.2025 by Ray Dalio

The system employs a defined prompting strategy to elicit paraphrased questions, ensuring consistent formatting of the output and facilitating controlled linguistic variation.

New research introduces a benchmark and metric to test how well AI systems maintain consistent answers when faced with paraphrased queries.

Pinpointing Faults in a Wind Farm Network

30.11.2025 by Ray Dalio

A wind farm’s electrical infrastructure integrates multiple individually-controlled, bidirectional renewable energy sources - known as IBRs - along radial feeders to efficiently collect and distribute generated power.

New research offers a streamlined approach to accurately locate short circuits in wind farms increasingly reliant on inverter-based resources.

Network Resilience Under Uncertainty

30.11.2025 by Ray Dalio

The study demonstrates that incorporating distributional robustness into cascading risk assessment-particularly in networks with zero time delay-significantly mitigates risk compared to traditional methods that do not account for parameter uncertainty, effectively reducing it to the level of single-agent risk.

New research provides tools to quantify and mitigate the risk of cascading failures in multi-agent systems grappling with unpredictable delays and network conditions.

The Limits of Efficient Error Correction

30.11.2025 by Ray Dalio

New research establishes fundamental constraints on how compactly data can be encoded for reliable, localized decoding.

Mirroring Security: Digital Twins Fortify IoT Networks

30.11.2025 by Ray Dalio

Digital twins facilitate service selection within SAGIN-enabled IoT networks, operating not as prescriptive tools but as evolving models anticipating inevitable systemic failures and necessitating continuous adaptation to maintain functionality.

A new framework uses digital twin technology and game theory to enhance secure service selection in the complex landscape of space-air-ground integrated networks.

Securing Medical AI: A New Approach to Collaborative Learning

30.11.2025 by Ray Dalio

Clients employ a privacy-preserving federated learning approach by extracting $768$-dimensional $[CLS]$ tokens from a Vision Transformer, encrypting them with the CKKS scheme, and enabling the server to aggregate information across numerous clients while performing encrypted inference.

A novel framework combines the power of vision transformers with lightweight encryption to enable privacy-preserving medical image analysis across distributed datasets.