Sam's News — tech-research — 2026-08-17¶
AI¶
6.5 Jais 2: Arabic-centric large language model family released¶
MBZUAI, Cerebras, and Inception jointly released Jais 2, a family of Arabic-centric LLMs with strong performance on Arabic benchmarks.
Sources: arXiv — Computation and Language RSS
6 Blockwise Causal Memory Transformer reduces quadratic attention complexity¶
New Transformer architecture uses blockwise causal memory to reduce self-attention complexity from quadratic to linear with sequence length.
Sources: arXiv — Computation and Language RSS
Research¶
6.5 Hallucination in Legal RAG Systems: Fine-Grained Analysis¶
A study analyzes hallucination in retrieval-augmented generation systems applied to legal domain, where ungrounded answers carry serious consequences.
Sources: arXiv — Computation and Language RSS
6 Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models¶
A paper investigates whether visible reasoning traces in large language models with extended chain-of-thought actually correlate with correct answers or are merely amplified artifacts of training.
Sources: arXiv — Computation and Language RSS
6 ASSERT: Measurement Pipeline for GenAI System Audits¶
ASSERT is a measurement pipeline that audits generative AI systems by assessing compliance rates and deployment gatekeeping in a principled way.
Sources: arXiv — Computation and Language RSS
6 Agentic AI Prototypes Stall Without Production Discipline, Says Info-Tech Research¶
Info-Tech Research Group reports that agentic AI initiatives frequently stall when prototypes lack production-grade discipline and rigor.
Sources: PR Newswire Web Search
Health¶
6.5 Researchers identify protein that may enable heart failure recovery¶
Scientists at Virginia Tech's Fralin Biomedical Research Institute discovered a protein that could help failing hearts recover.
Sources: Virginia Tech News Web Search
AI Research¶
6 Seeing Red, Thinking Bad: Color Bias in Vision Language Models¶
A study demonstrates that vision language models exhibit problematic color biases that may affect decisions in recruitment and recommendation systems.
Sources: arXiv — Computation and Language RSS, arXiv — Computer Vision RSS
5.5 Regime-Conditional Verification: Safety Classifier Adaptation and Monitoring¶
A new arXiv paper proposes methods to adapt and monitor safety classifiers in large language models as deployment policies and traffic evolve.
Sources: arXiv — Computation and Language RSS, arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
5.5 Stable Miscalibration in Large Language Models: High-Confidence Errors¶
A new study examines how large language models produce confident wrong answers that remain stable under small perturbations.
Sources: arXiv — Computation and Language RSS, arXiv — Artificial Intelligence RSS