Sam's News — tech-research — 2026-08-14¶

AI/Finance/Policy¶

7 Governance Framework for Agentic AI in Financial Services¶

Researcher Henry Han published an analysis of governance frameworks for agentic AI in financial services, identifying the 'Verifiability Gap' as the primary governance constraint—the shortfall between verification demands and explainability after a decision. Testing across nine model versions found frontier models reject standard controls, alter historical actions across releases, and achieve poor reproducibility compared to local models.

58-page study submitted August 11–13, 2026; argues verifiability, not capability, is primary constraint
Local models reproduced 320/320 executions; hosted models: 319/320 and 959/960 reproducibility
Frontier models reject standard temperature, top_p, top_k controls; expose no random seed
Architecture changes affect final actions; no execution record repeated across configurations
Proposes governance where reproducibility functions as profile rather than scalar measure

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

AI/Science¶

7 TRACES Benchmark for LLM Epistemic Reliability in Science¶

TRACES is a benchmark evaluating whether large language models can distinguish reliable from unreliable scientific literature when deployed as research agents. Testing 30 models across 10 runs found every model failed more than 71% of agentic probes and 22 of 30 failed 90%+ of the time, indicating topic-keyed safety behavior rather than robust epistemic competence.

Benchmark submitted August 11, 2026; uses 42 retracted, fraudulent, pseudoscientific papers
Tested 30 models with aggregate IFR-a of 0.93 ± 0.004; aggregate IFR-i of 0.809 ± 0.009
Every model failed 71%+ of agentic probes; 22 of 30 failed 90%+ of the time
Models engaged with untenable premises in 95% of non-empty responses
Rejections concentrated on high-notoriety topics; disappear under matched-structure controls

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

AI/Security¶

6.5 Framework for Evaluating LLM-Generated Cybersecurity Detection Rules¶

Researchers present an open-source evaluation framework and benchmark metrics to measure the effectiveness of LLM-generated security detection rules.

Sources: arXiv — Artificial Intelligence RSS

AI/Education¶

6.5 Methods for Improving LLM-Based AI Tutoring Systems in K-12 Education¶

Researchers develop evaluation and experimentation methodologies to measure and improve the quality of LLM-powered tutoring systems for schools.

Sources: arXiv — Artificial Intelligence RSS

AI/Medical¶

6.5 Uncertainty-Aware Deep Learning for Skin Lesion Classification¶

An ensemble deep learning framework with uncertainty quantification and explainability improves diagnosis of skin cancer from dermoscopic images.

Sources: arXiv — Artificial Intelligence RSS

Systems/ML¶

6.5 APEX: Adaptive Expert Prefetching for Edge MoE Inference¶

A memory-efficient inference optimization for Mixture-of-Experts models on edge devices through adaptive expert prefetching.

Sources: arXiv — Artificial Intelligence RSS

AI/Transparency/Policy¶

6.5 Impact of Disclosing Persuasive Intent in AI Chatbots¶

Study finds that disclosing AI involvement and persuasive intent to users reduces the persuasive effect of AI-generated content.

Sources: arXiv — Artificial Intelligence RSS

Quantum Computing/ML¶

6.5 Hamilton-Zero: Neural Tensor-Network Model for Quantum Ground States¶

A neural tensor-network foundation model enables efficient amortized computation of ground states for arbitrary quadratic qubit Hamiltonians.

Sources: arXiv — Artificial Intelligence RSS

Research¶

6.5 Researchers selected for inaugural DOE Genesis Mission projects¶

The Department of Energy's Genesis Mission launches with selected research projects aimed at doubling American scientific productivity.

Sources: google.com Web Search Update to: Researchers selected for inaugural Department of Energy Genesis Mission projects

Technology¶

6.5 AI-generated code introduces quality and security risks without guardrails¶

Widespread adoption of AI-generated code is accelerating software defects, security vulnerabilities, and technical debt without adequate safeguards.

Sources: google.com Web Search