Skip to content

Column

Research, in business language.

We rewrite Affectosphere Group's research into something useful for business and practical decision-making. Each piece is a 5-minute read.

An abstract flat illustration showing news nodes from multiple companies connected by sentiment-weighted edges in a knowledge graph

2026 / 06 / 07

Reading One Company's News Is Not Enough to Understand That Company's Stock

Sentiment propagates across supply chains and partnerships in ways that vector search cannot capture. Graph-RAG improves entity recall by 6.4% and answer relevance by 11.7% on complex cross-entity queries — with only a 22.6% latency increase. Here is what this means for financial research automation.

5 min read Read →
An abstract flat illustration showing a medical AI system passing through multiple evaluation domain checkpoints and a fairness testing gate

2026 / 06 / 07

Why 'High Average Score' Is Not a Trustworthy Safety Standard for Medical AI

100 medical professionals evaluated medical LLMs across 9 domains and 690 adversarially designed test cases. High average accuracy masked serious failures in specific scenarios. LLM judges missed safety concerns that human experts caught. And changing patient demographics alone amplified errors by 10–20%. Here is how to rethink medical AI procurement.

5 min read Read →
An abstract flat illustration showing medical records and bank transaction histories linked by a risk detection system

2026 / 06 / 07

Financial Fraud Peaks When Alzheimer's Patients Miss Their Medication

Cognitive vulnerability in Alzheimer's patients fluctuates with medication adherence — and financial exploiters may be exploiting exactly those windows. Integrating medication records with transaction patterns pushed recall during non-adherent periods from 0.74 to 0.91. Here is what this means for elder financial protection.

5 min read Read →
An abstract flat illustration showing EEG waveforms and eye-tracking trajectories integrating into a unified emotion signal

2026 / 06 / 07

Combining EEG and Eye-Tracking: Can Emotion Recognition Finally Work Across Different People?

Subject-to-subject and session-to-session domain shift has long been the biggest barrier to deploying emotion recognition in the real world. UF-AMA clears that bar through confidence-based filtering and multi-stage domain adaptation — with practical implications for call centers and remote hiring.

5 min read Read →
An abstract illustration of an LLM agent's planning capability being evaluated across multiple diagnostic settings

2026 / 06 / 06

Can You Tell Whether an AI Agent Failed at Planning or Execution?

When an LLM agent fails, most evaluations can't tell you whether planning or execution was the problem. APB is a diagnostic benchmark designed to break that black box open — and it has direct implications for how organizations should select agents before deploying them.

5 min read Read →
An abstract illustration of user preferences compressed into compact prefix embeddings passed to an LLM

2026 / 06 / 06

Personalizing LLMs at Scale Without Per-User Models

Two dominant personalization approaches both hit walls at scale: retrieval quality dependency and storage costs that grow with user count. TAP-PER encodes user preferences into compact prefix embeddings, bypassing both.

5 min read Read →
An abstract illustration of an AI agent's proposed action being checked by a policy layer before execution

2026 / 06 / 06

Why AI Agents Need a Layer That Can Stop Them Before They Act

The risk with LLM agents isn't that they propose wrong actions — it's that wrong proposals get executed. The Organizational Control Layer concept offers a practical governance design for anyone deploying AI in operational workflows.

5 min read Read →
An illustration of user behavior logs being transformed into a natural language profile that an AI uses to personalize its responses

2026 / 06 / 06

Label-Free User Profiling: What BUMP Means for LLM Personalization

No task labels, no manual annotation. A self-supervised framework that generates natural-language user profiles from behavior logs alone could make LLM personalization accessible at scales where it previously wasn't cost-effective.

5 min read Read →
An abstract illustration of AI emotional expression caught between suppression and release

2026 / 06 / 06

Should We Suppress or Unleash AI Emotion?

LLMs are deliberately trained to suppress emotional expression. Is that the right call? A new study using self-rewarding reinforcement learning asks whether giving AI the ability to "feel" might actually make it more robust.

5 min read Read →
An abstract illustration showing human connections fading as AI interactions accumulate

2026 / 06 / 05

You Can Develop AI Emotional Dependence Without Even Trying

You don't need a companion app for AI emotional dependence to take hold. A 28-day longitudinal study shows that everyday interactions with general-purpose AI quietly shift how we seek emotional support from other humans.

5 min read Read →
An abstract illustration of an LLM being evaluated through a dynamic multi-step clinical dialogue scenario

2026 / 06 / 05

Why a High Benchmark Score Does Not Mean an LLM Is Clinically Useful

The best LLMs score in the 90s on static medical QA benchmarks. In dynamic clinical conversations, the same models achieve 40–60%. MedSP1000 reveals a critical gap — and changes what health AI procurement should actually measure.

5 min read Read →
An abstract illustration of an AI agent passing through regulatory gates and receiving a certification seal

2026 / 06 / 05

A 'Trust Certificate' for AI Agents Before They Go Live

Deploying AI agents in regulated industries like finance, healthcare, and law raises a hard question: how do you verify compliance before going live? A new ontology-grounded framework offers a credible answer — and a new basis for AI procurement.

5 min read Read →
An abstract illustration of an unbalanced scale being stabilized through a group comparison mechanism

2026 / 06 / 05

Why Debiasing LLMs with Reinforcement Learning Is So Hard — And How BiasGRPO Fixes It

Standard RLHF becomes unstable when applied to social bias mitigation because bias evaluation is subjective and reward signals are noisy. BiasGRPO uses group-relative policy optimization to stabilize the training — and lays the technical groundwork for 'fairness-certified AI' in high-stakes applications.

5 min read Read →
An abstract illustration of multiple prompts being mixed and sent to a cloud API with results returned securely

2026 / 06 / 05

How to Use External LLM APIs Without Sending Your Sensitive Data

For GDPR-constrained organizations, the barrier to using cloud LLMs is not cost — it's the risk of sending confidential data externally. SharedRequest solves this with a batch-mixing approach that achieves 20%+ better utility than differential privacy while cutting query costs by up to 5x.

5 min read Read →
Abstract visualization of people from different cultural backgrounds experiencing different emotional responses while viewing the same artwork

2026 / 06 / 04

What an Image Makes You Feel Matters More Than What It Shows

The same image can trigger completely different emotional responses depending on cultural background. A new perception-modeling framework that captures both factual and affective dimensions of visual experience could reshape how global creative teams pre-screen their visuals.

5 min read Read →