Sam's News — tech-research — 2026-10-14¶

AI¶

6.5 Study reveals privacy leakage risks in multi-agent LLM dialogue systems¶

Research shows that conversational AI agents may leak sensitive information through information flow in multi-agent dialogue scenarios.

Sources: arXiv — Computation and Language RSS, arXiv — Cryptography and Security RSS

6.5 Georgia Tech and Emory announce $1 billion joint AI research and workforce initiative¶

Georgia Tech and Emory University committed over $1 billion to expand AI research, computing infrastructure, and technology workforce training programs.

Sources: CBS News Web Search

6.5 Geospatial AI agents deployed to address planetary environmental challenges¶

A new system uses geospatial artificial intelligence agents to tackle critical environmental and climate issues facing Earth.

Sources: Google Research Blog RSS

6 Ultra-low-bit quantization preserves language model accuracy at extreme compression¶

Researchers developed FidelityQ to maintain predictive accuracy when quantizing large language models to ultra-low-bit precision for efficient deployment.

Sources: arXiv — Computation and Language RSS, arXiv — Artificial Intelligence RSS

Research¶

6.5 Scaling Laws for Multimodal Data Mixtures in Frontier AI Systems¶

Study investigates scaling relationships and optimal data mixture ratios for training natively multimodal AI systems across text, vision, and audio.

Sources: arXiv — Computation and Language RSS

6 Agentic Data Engineering for LLM Training via Reinforcement Learning¶

Study investigates how AI agents can use experimental feedback to iteratively optimize data selection, rewards, and curricula for LLM training.

Sources: arXiv — Computation and Language RSS

6 Continuous Bitstream Diffusion for Unified Multimodal Generation¶

Researchers demonstrate that continuous bitstream diffusion provides a unified framework for joint and conditional multimodal generation across image, text, and speech.

Sources: arXiv — Computation and Language RSS

LLM Optimization¶

6 Relevance-Filtered Context Compression for Personalized LLMs¶

Long conversation histories can be compressed to 5% of original text while maintaining personalization performance in large language models.

Sources: arXiv — Computation and Language RSS

NLP¶

6 UltraBERT: Modernized Bidirectional Encoder Pre-training¶

UltraBERT applies innovations from causal large language models to improve pre-training recipes for bidirectional encoders.

Sources: arXiv — Computation and Language RSS

AI Safety¶

6 Lumen-Bench: Cross-Lingual Safety Gaps in Tool-Calling LLMs¶

A benchmark reveals divergence between what multilingual LLMs refuse verbally and what they execute when equipped with tools across languages.

Sources: arXiv — Computation and Language RSS