Skip to content

Sam's News — tech-research — 2026-08-15

AI Safety

7 Language Choice Affects LLM Safety: Testing Cross-Language Alignment in Nine Models

Research testing nine language models from six providers finds that prompting in languages other than English can circumvent safety alignment in high-stakes scenarios. The effect stems from the language models reason in rather than input language, with Japanese prompts significantly reducing harmful recommendations in some models through spontaneous moral reasoning.

  • Study tested 9 LLMs from 6 providers across 6 languages using nuclear strike vignettes
  • Claude Sonnet 4.6 launch rate dropped from 40% to 0% in unnecessary-strike scenarios with Japanese prompts
  • Google Gemini Pro 3.1 dropped from 53% to 13% with Japanese prompts
  • Reasoning language, not input language, drives effect; models generate moral vocabulary when reasoning in Japanese
  • 5 other models showed no language-dependent effect; effect requires models already showing English hesitation

Sources: arXiv AI Web Searched, arXiv — Artificial Intelligence RSS

6.5 IntegrityBench: Evaluating LLM Research Integrity as Co-Scientists

Researchers introduce IntegrityBench, a benchmark measuring language models' ability to maintain research integrity under institutional pressure.

Sources: arXiv — Artificial Intelligence RSS

Architecture

6.5 Dual-Flow Transformers: Decoupling Prefill from Decode Inference

A technical approach separates transformer prefill and decode phases to optimize inference cost by exploiting their different hardware characteristics.

Sources: arXiv — Artificial Intelligence RSS

AI Limitations

6.5 Phase Transitions in LLM Constraint Satisfaction: The Compositional Bottleneck

Research reveals that LLMs handle individual constraints well but exhibit phase-transition failures when satisfying multiple simultaneous constraints.

Sources: arXiv — Artificial Intelligence RSS

Healthcare

6.5 ThyroidXAgent: Auditable Agentic AI for Thyroid Ultrasound Diagnosis

An agent system integrates lesion localization, measurement, risk stratification, and reporting for thyroid ultrasound with auditable clinical review support.

Sources: arXiv — Artificial Intelligence RSS

Evaluation

6.5 Jagged Judges: Epistemic Stability of LLM-Based Evaluators

Research shows that LLM judges, central to model evaluation and reward modeling, lack stability under re-prompting, challenge, and persistent questioning.

Sources: arXiv — Artificial Intelligence RSS

Safety

6.5 Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents

Research shows that unsafe successes in self-improving agents can become reusable persistent policies even after their triggering context disappears.

Sources: arXiv — Artificial Intelligence RSS

AI/Autonomous Systems

6.5 FlashDrive: Optimizing Vision-Language-Action Model Inference for Real-Time Autonomous Driving

Researchers present FlashDrive, a method to reduce the computational cost of Vision-Language-Action models for autonomous driving inference.

Sources: arXiv — Artificial Intelligence RSS

AI Collaboration

6.5 BoardroomAI: Human-Steerable Multi-Agent Deliberation with Evolving Decision Graphs

A system enabling real-time human interaction with multi-agent deliberation that tracks evolving evidence, constraints, and priorities through decision graphs.

Sources: arXiv — Artificial Intelligence RSS

AI Research

6.5 VALG: An Agentic System for Automated Machine Learning Theory Research

A system using agents to autonomously work on open problems in machine learning theory by navigating mathematical setups and solving theorems.

Sources: arXiv — Artificial Intelligence RSS