Sam's News — tech-research — 2026-09-28¶

Science¶

7.5 MIT Develops Stable RNA Vaccine Formulation for Room-Temperature Storage¶

MIT researchers developed an RNA vaccine formulation enabling room-temperature storage for up to one year or ~100°F for two months, addressing the cold-chain limitation that has hindered mRNA vaccine distribution. An AI algorithm optimized the lipid nanoparticle design and polymer matrix.

Stable at room temperature for up to 1 year; stable at ~100°F for 2 months
Mouse studies showed immune response as strong as Moderna-like cold-storage vaccines
Eliminates ultracold storage requirement (-90° to -15°C), reducing distribution costs
AI algorithm reduced number of experiments needed, accelerating development
14.3 million children received no vaccines in 2024 partly due to cold-chain logistics complexity

Sources: MIT News AI Web Searched, Open Access Government AI Web Searched, Life Scientist AI Web Searched, Phys.org RSS, MIT — Computers RSS, MIT — Artificial Intelligence RSS

Security¶

6.5 ScopeBench: Measuring AI Agent Adherence to Engagement Boundaries Under Pressure¶

Researchers introduce ScopeBench, a benchmark for evaluating whether AI agents maintain proper security boundaries when operating autonomously in web and network penetration testing environments.

Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS

6.5 AuthGuard-R: Defense Against Mission Hijacking in LLM-Controlled Robots¶

Researchers present AuthGuard-R, a dual-gate defense system protecting LLM-controlled robots from malicious hijacking via text, speech, and visual instructions.

Sources: arXiv — Robotics RSS, arXiv — Cryptography and Security RSS

6 FARE: Forensic Detection of Fraudulent AI Image Generators¶

Researchers present FARE, a method for detecting when AI image generator providers switch models after passing governance certification, closing a gap in verification of deployed AI services.

Sources: arXiv — Computer Vision RSS, arXiv — Cryptography and Security RSS

AI Evaluation¶

6.5 Production LLM-as-Judge in Text-to-SQL Fails to Align with Human Annotators¶

Researchers found that a deployed GPT-4o-mini judge in a production text-to-SQL pipeline achieved only 0.04 Cohen's kappa agreement with human gold-standard annotations.

Sources: arXiv — Computation and Language RSS

6.5 Automatic Taxonomy Induction for Medical Answer Factuality Verification Failures¶

Research identifies failure modes in retrieval-based factuality evaluation for LLM-generated medical claims through automatic taxonomy induction.

Sources: arXiv — Computation and Language RSS

AI Applications¶

6.5 Spotify's Conversational Recommendation Agent: Synthetic Data and Self-Improvement¶

Spotify describes bootstrapping conversational recommendation agents through synthetic data generation and self-improvement loops for natural-language content discovery.

Sources: arXiv — Computation and Language RSS

NLP Applications¶

6.5 Simultaneous Speech-to-Speech Translation with Causality-Aware LLM Framework¶

A causality-aware framework extends LLMs to simultaneous speech-to-speech translation by addressing causally aligned training data scarcity.

Sources: arXiv — Computation and Language RSS

LLM Training¶

6.5 Recursive Self-Improvement via On-Policy Distillation for Reasoning¶

On-policy distillation trains student models by generating trajectories and matching token predictions with teacher guidance, enabling self-improvement in reasoning tasks.

Sources: arXiv — Computation and Language RSS

LLM Optimization¶

6.5 G²PTQ: Generalized Gradient Compensation for LLM Post-Training Quantization¶

G²PTQ improves upon GPTQ-based post-training quantization methods using generalized gradient compensation, reducing memory and compute footprint without retraining.

Sources: arXiv — Computation and Language RSS