Sam's News — tech-research — 2026-10-15¶
AI¶
6.5 ELARSpeech: 30K-Hour Multilingual Speech Dataset Covering 675 Underrepresented Languages¶
A new large-scale dataset addresses the linguistic long tail by providing speech data for 675 underrepresented languages to improve multilingual speech models.
Sources: arXiv — Computation and Language RSS
6 DecisionRobust: Benchmarking JEV-Style Decision Models for Robustness¶
A benchmark evaluates the robustness of emerging JEV-style decision models across text and vision modalities for automated decision-making.
Sources: arXiv — Computation and Language RSS, The Register RSS
6 LLM-jp-4: Japanese Sovereign Language Model with 8B and 32B-MoE Variants¶
Researchers release LLM-jp-4, a family of transparent and reproducible Japanese language models including an 8B dense model and a 32B mixture-of-experts variant.
Sources: arXiv — Computation and Language RSS
6 Vision as Scalable Anchors in Native Multimodal Pre-Training¶
Research investigates how visual signals shape scaling behavior and enable multimodal capabilities in language models trained from scratch on mixed modalities.
Sources: arXiv — Computation and Language RSS
5.5 Training Voice Agents to Follow Regulated Procedures in Debt Collection¶
A study demonstrates how to train LLM-based voice agents to follow legally required conversational flows in debt collection scenarios.
Sources: arXiv — Computation and Language RSS
5.5 Pre-Trained Language Models for Zero-Shot Social Science Research Tasks¶
Research demonstrates that pre-trained language models can perform record linkage and demographic inference from names without labeled training data.
Sources: arXiv — Computation and Language RSS
5.5 Study of How Language Models Develop In-Context Learning Abilities¶
Research investigates the mechanisms by which language models acquire and develop in-context learning from prompts and demonstrations.
Sources: arXiv — Computation and Language RSS
5.5 Guidance on Reliable Setup Choices for Language Model Benchmarking¶
Research addresses the lack of standardization in LLM evaluation design choices and their impact on measured performance.
Sources: arXiv — Computation and Language RSS
5.5 Titan-Llama: External Memory Modules for Extended LLM Context¶
A method retrofits pre-trained language models with external neural memory modules to reduce KV-cache requirements for long-context inference.
Sources: arXiv — Computation and Language RSS
AI Research/Security¶
5.5 Defense mechanism against identity theft in personalized image generation systems¶
Researchers present a proactive defense method against unauthorized identity-preserving generation using facial reference images in diffusion model-based applications.
Sources: arXiv — Computer Vision RSS, arXiv — Cryptography and Security RSS