Sam's News — tech-research — 2026-09-11¶
AI¶
7 SimpleDesign: joint generative model for protein sequence and structure design¶
Apple Machine Learning Research published SimpleDesign, a generative model that jointly designs protein sequences and 3D structures in a single end-to-end training process. The model captures the intrinsic relationship between amino acid sequences and structures using Transformer-based multimodal architecture trained on over 2 million sequence-structure pairs.
- Single-stage end-to-end training directly in data space, unlike multi-stage autoencoder approaches
- Trained on over 2 million sequence-structure pairs
- Uses Transformer multimodal backbones with discrete cross-entropy loss for sequences and regression for structures
- Achieved competitive performance on co-design and unconditional generation benchmarks
- Published September 2026
Sources: Apple Machine Learning Research AI Web Searched
6 Verbalized Confidence as Robust Soft-Scoring for LLM-as-a-Judge¶
Research shows that explicit confidence statements from LLMs provide more reliable soft-scoring for evaluation tasks than previously assumed.
Sources: arXiv — Computation and Language RSS
6 Fragility Spectrum in Recursive Language Model Training¶
Research documents how repeatedly training on model-generated text causes output diversity collapse and identifies vulnerable data mixtures.
Sources: arXiv — Computation and Language RSS
6 OmniHallu: Unified Hallucination Detection for Multimodal LLMs¶
A framework for detecting hallucinations in both comprehension and generation tasks across multimodal language models.
Sources: arXiv — Computation and Language RSS
5.5 Data-Efficient Language Modeling: Qiushi Engine advances BabyLM 2026 research¶
A research study describes advances in training language models on limited text through improved context usage, generalization, and capability retention.
Sources: arXiv — Computation and Language RSS
5.5 Auto-RecSys: Autonomous Agents for Industry-Scale Recommendation Systems¶
Researchers propose using autonomous research agents to automate hypothesis generation and experiment execution for scaling recommendation models in production environments.
Sources: arXiv — Computation and Language RSS
5.5 Distribution-Aware Language Neuron Identification in Multilingual LLMs¶
A study identifies language-specific neurons in multilingual large language models using distribution-aware techniques.
Sources: arXiv — Computation and Language RSS
Health¶
6 LLM-anchored paralinguistic enrichment improves Alzheimer's disease detection from speech¶
A speech-based method using LLM enrichment of paralinguistic features enables improved automatic early detection of Alzheimer's disease.
Sources: arXiv — Computation and Language RSS
Medical Technology¶
6 Lincoln Laboratory's AI-GUIDE Catheterization Device Wins 2026 Excellence in Technology Transfer Award¶
Lincoln Laboratory and Massachusetts General Hospital's handheld AI-GUIDE catheterization device was awarded the 2026 Excellence in Technology Transfer Award.
Sources: MIT — Artificial Intelligence RSS
AI Research¶
5.5 Research: Language model membership inference through verified duplication counts¶
A new arXiv paper investigates how to reliably detect whether specific sentences appear in language model training data using duplication count analysis.
Sources: arXiv — Computation and Language RSS, arXiv — Cryptography and Security RSS