Sam's News — tech-research — 2026-09-09¶
AI Research¶
6.5 LLM-Based Phenotyping Improves Opioid Use Disorder Detection in Electronic Health Records¶
A rubric-guided language model approach improves identification of opioid use disorder cases from clinical narratives buried in electronic health records.
Sources: arXiv — Computation and Language RSS
6 Agent Governance Frameworks for Autonomous AI Systems¶
Research synthesizes design and evaluation frameworks for governed autotelic AI agent organizations operating within safety guardrails.
Sources: arXiv — Computation and Language RSS
6 Magnetic Vector Organization Reveals Linguistic Structure Within Language Model Layers¶
Research identifies special token vectors in language models that organize surrounding tokens by attraction or repulsion, revealing internal linguistic structure.
Sources: arXiv — Computation and Language RSS
AI Safety¶
6.5 Safety Monitors in LLMs: Recall Alone Doesn't Prevent Harm¶
A study finds that safety monitors for large language models are evaluated by recall against harmfulness labels, but preventing a flagged request requires the model would have otherwise complied.
Sources: arXiv — Computation and Language RSS
AI Ethics¶
6.5 Clinical Framing Improves Suicide Risk Measurement Beyond Policy-Based Content Moderation¶
Content moderation APIs built to flag violations fail to measure clinical risk; clinical framing reveals graded risk beyond binary detection.
Sources: arXiv — Computation and Language RSS
6 LLMs Sacrifice Individual Distinctiveness When Demographically Conditioning for Cultural Adaptation¶
Demographic conditioning in LLMs for cultural adaptation encourages stereotyping rather than serving individual user preferences.
Sources: arXiv — Computation and Language RSS
6 LLMs Skew Male When Generating Media in Local Languages Despite Reflecting Country Gender Patterns in Text¶
LLMs accurately reflect country-specific gender patterns when answering questions but exhibit male bias when generating long-form media in local languages.
Sources: arXiv — Computation and Language RSS
ML Research¶
6 Multi-Turn LLM Degradation: How Previous Responses Bias Later Behavior¶
A study examines how an LLM's own previous responses become context that affects its behavior in multi-turn conversations.
Sources: arXiv — Computation and Language RSS
6 Neuron-Guided Fine-Tuning: Efficient LLM Alignment Without Catastrophic Forgetting¶
A fine-tuning method targets specific neurons to align large language models while preventing catastrophic forgetting and reducing parameter redundancy.
Sources: arXiv — Computation and Language RSS
AI Alignment¶
5.5 Spillover-Aware Multi-Value Steering for Pluralistic LLM Alignment¶
Researchers propose a new activation steering method that handles multiple concepts simultaneously to enable pluralistic alignment in large language models where different stakeholders have different value priorities.
Sources: arXiv — Computation and Language RSS, arXiv — Artificial Intelligence RSS