PROCESSBENCH evaluates step-by-step errors in AI models, enhancing analysis and oversight in solving complex math problems.

AI Knowledge Circuits: Transformers use internal circuits to manage and update knowledge with targeted, efficient interventions.

EvalGIM evaluates text-to-image models on quality, diversity, and consistency. Modular and flexible, it supports research and business.

Technology 2025: AI and AR transform markets and society. Innovation and regulation must balance to mitigate global risks.

Tech Trends 2025: AI is the cognitive substrate of the digital future, requiring ethics, transparency, and integration.

RARE boosts LLM reasoning with A6, A7 retrieval and RAFS. Open-source models exceed GPT-4. Limits: cost and non-standard metrics.

Machine Learning improves tasks by learning from data, requiring ethical management to reduce biases and social impacts.

AI in Italy: €760M market, 90% large companies’ investments, SMEs 18%. Barriers: skills, costs. Training and incentives needed.

Generative AI is growing, but integrating solutions remains difficult. Trust and speed are keys to success.

Boston Consulting Group: AI maturity The USA and Singapore lead thanks to investments and ethics.

Corporate AI. Only 10% of Italian companies are leaders. Challenges include inadequate infrastructure, fragmented data, and talent gaps.

AI in accounting and finance optimizes processes and decisions but needs governance and quality data for sustainable strategic results.

Multimodal AI combines data for personalized care. Ethical and privacy challenges require innovation and strategic vision for success.

OWASP Top 10 LLM 2025 lists key vulnerabilities and measures for LLMs: input filters, human supervision, and monitoring for safety.

BALROG evaluates LLMs/VLMs in complex environments, testing reasoning and planning. It seeks to address agentic and multimodal limitations.

GenAI in banking transforms services and risk management through gradual adoption. Governance, ethics, and security are crucial for success.

MRJ-Agent uses risk decomposition and psychological induction for effective multi-round attacks against advanced AI model defenses.

BrainBench reveals LLMs predict scientific results with 81.4% accuracy, surpassing human experts at 63.4%.

Configurable Foundational Models: dynamic LLM modules, more scalable, adaptable, and ideal for efficiency and personalization.