Sam's News — tech-research — 2026-10-07¶
Health AI¶
6 Fiorillo v0.5: Open Model for Calibrated Randomized Trial Interpretation¶
An open 4-billion-parameter model answers questions about randomized trial outcomes with probability calibration for clinical decision-making.
Sources: arXiv — Computation and Language RSS
6 TIDE 2.0: Open De-identification Engine for Clinical Notes¶
An open, model-agnostic system for removing protected health information from clinical notes with keyed de-identification.
Sources: arXiv — Computation and Language RSS
Interpretability¶
6 Mechanistic Identification of Genuine Introspection in Large Language Models¶
Research into distinguishing authentic self-reflection from confabulation in LLM claims about internal states using mechanistic analysis.
Sources: arXiv — Computation and Language RSS
Architecture¶
6 Recurrent Looped Transformer: Efficient State Tracking Architecture¶
Transformer architecture that reuses layers in a recurrent loop to efficiently handle state updates without fixed-depth constraints.
Sources: arXiv — Computation and Language RSS
Efficiency¶
6 Monte Carlo KV Cache Eviction: Future-Aware Memory Management¶
Method using Monte Carlo estimation to predict which cached key-value memory will be needed during decoding for intelligent cache eviction.
Sources: arXiv — Computation and Language RSS
Model Efficiency¶
5.5 DLoop: Looped Speculative Decoding¶
Researchers introduce DLoop, a technique that accelerates autoregressive generation in large language models through improved speculative decoding.
Sources: arXiv — Computation and Language RSS, arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
AI Research¶
5.5 Zero-shot visualization method enables exploration of text corpora using natural language¶
Researchers introduced a technique for visualizing text document collections by mapping them onto user-specified natural language concepts.
Sources: arXiv — Computation and Language RSS
5.5 Study examines capacity, responsiveness, and alignment in latent structures of language models¶
Research investigates what makes localized structures in language model activation spaces actionable for controlling behavior.
Sources: arXiv — Computation and Language RSS
5.5 Tree navigation without LLM summaries improves hierarchical retrieval for long-document QA¶
A study demonstrates that hierarchical tree-based retrieval without pre-computed summaries improves question-answering over long documents.
Sources: arXiv — Computation and Language RSS
5.5 Condition-anchored distillation stabilizes language models under continual learning¶
A new method preserves language model performance on previously learned tasks during continual adaptation through generative distillation.
Sources: arXiv — Computation and Language RSS