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.

Flat illustration of a compliance dashboard where control lists from multiple European security standards are automatically linked to technical metric rows with similarity scores through a sentence-encoder neural network

2026 / 07 / 08

Stop tracing by hand which standard each control satisfies — automating cloud compliance mapping with sentence encoders

Cross-checking your cloud controls against multiple European security standards like ENISA and NIS2 — figuring out which requirement each one satisfies — still depends on expert hands. A domain-adapted sentence encoder improved control-to-metric mapping by up to +23 nDCG@10 points over baseline and reached 0.870 nDCG@10 on cross-standard association. Here is how GRC, IT audit, and FinTech/MedTech teams could try it — departments and KPIs included.

6 min read Read →
Flat illustration of a large corpus of transactional contracts on the left, with three paths — Full Injection, NAVEMBED, NAVINDEX — showing token reduction ratios of 1x, 17.3x, and 56x, next to a cost-crossover graph marking where cached injection breaks even against retrieval

2026 / 07 / 08

Stop feeding whole contracts to the AI — cutting legal-document API cost by up to 56x with inject vs. navigate

When you have an LLM read a pile of transactional contracts and answer questions, dropping the full text into the prompt is the most natural move. It is also the most expensive one under token billing. On a 20-question benchmark, reranked semantic retrieval kept the same accuracy while cutting context tokens by 17.3x, and the LLM-navigation approach NAVINDEX compressed context size by about 56x. Here is how to lower the API cost of contract-review and legal-research AI structurally, from a legal-tech point of view.

6 min read Read →
Flat illustration of a hospital oncology decision-support system with three layers — patient-data standardization, clinical routing, and interchangeable AI models — connected by clean modular pipelines

2026 / 07 / 08

Cutting oncology AI's vendor lock-in by design — LCA, a decision-support framework where you keep swapping the model

When a hospital brings in AI-assisted diagnosis, many people worry: what if we get chained to this vendor and cannot switch in a few years? LCA separates data ingestion, clinical routing, and inference into three layers so you can freely swap only the AI model that does the reasoning. Here is a concrete proposal for how a cancer hospital's health informatics team could bake it into AI procurement criteria.

6 min read Read →
A marketing researcher inputs a projective technique prompt into an LLM on a laptop screen, while a split panel on the right shows human response cards alongside AI-generated responses, with large overlap in broad themes but visible structural and stylistic differences

2026 / 07 / 08

Can LLMs Serve as the Opening Act for Consumer Research? Projective Techniques, Synthetic Data, and What It Means for the Field

Can LLMs replace human respondents in projective techniques — the qualitative methods that surface consumers' hidden associations, emotions, and desires? A new study comparing LLM-generated and human responses across multiple models, prompting strategies, and temperature settings offers a nuanced answer.

5 min read Read →
An abstract illustration of NLP models analyzing large-scale social media data to classify mental health symptoms

2026 / 07 / 08

What 1.3 Million Tweets Reveal About Depression in ADHD and ASD Communities

Researchers profiled DSM-5 depressive symptoms across 792 self-reported ADHD and ASD Twitter users using MentalRoBERTa fine-tuned on ReDSM5. Differences emerged between the two groups — with important caveats about what population-level data can and cannot tell us.

5 min read Read →
A flat illustration of dialogue branching. A flowchart where persuasion policies switch at each turn

2026 / 07 / 06

Fixed Scripts Can't Persuade a Resident Who Refuses to Evacuate — The Era of Per-Utterance Policy Switching

When residents refuse to evacuate a burning building, which persuasion strategy should an AI choose? DiPS, a Q-learning dialogue policy framework, offers an answer — and a blueprint for high-stakes call center scenarios.

5 min read Read →
A split-screen data visualization showing the same U.S. mass shooting statistics with episodic framing on the left and thematic framing on the right, with different emotional response indicators below each panel

2026 / 07 / 06

The Words Around a Chart Change What You Feel: A Framing Experiment in Data Visualization

Episodic or thematic? When an 800-person experiment varied the text framing around the same mass shooting statistics, emotional valence shifted significantly — and the implications for AI-generated data storytelling are hard to ignore.

5 min read Read →
A flat illustration showing medical reasoning steps linked by arrows, with evidence sources visible at each stage

2026 / 07 / 06

Can Medical AI Explain Why? FaithMed's Step-Level Supervision for Faithful Clinical Reasoning

FaithMed trains LLMs with clinician-designed rubrics and step-level process rewards, achieving roughly 9% gains over agentic search baselines on 7 medical benchmarks — and building the audit trail that real healthcare deployment demands.

5 min read Read →
A flat illustration showing a three-stage training pipeline leading to a fitness knowledge graph

2026 / 07 / 06

Can AI Replace a Personal Trainer? FitOne's Blueprint for Domain-Specific LLM Post-Training

FitOne — a fitness-specialized LLM built on Qwen3 with a three-stage post-training pipeline — outperforms baseline models by up to 12.73% on professional fitness certification exams, and offers a reusable architecture for scalable domain-specific AI.

5 min read Read →
A flat illustration of five distinct team types. Diagrams showing different distributions of autonomy and decision-making authority

2026 / 07 / 06

"Human-AI Collaboration" Is Actually Five Different Things — The Taxonomy That Exposes an AI Deployment Blind Spot

A systematic analysis of 53 papers reveals five qualitatively distinct human-AI team clusters — and shows that researchers using the same phrase "human-AI teaming" often mean entirely different team dynamics. Here's what that means for enterprise AI deployment design.

5 min read Read →
Flat illustration of three AI robots trying to sort human expression cards and struggling with the fine-grained emotion cards

2026 / 07 / 02

How Well Can LLMs Read Human Emotions? A 13-Class Benchmark Reveals a Zero-Shot Ceiling at Roughly 40%

Claude, GPT-5.4, and Gemini were benchmarked zero-shot on a 13-class emotion taxonomy across 131,306 sentences — and all three plateaued at around 40% accuracy. Here is what the 'affective gap' means before you build emotion AI into your product.

5 min read Read →
Flat illustration of a small AI robot retrieving notes from a large, well-organized memory shelf with drawers while progressing along a long winding task road

2026 / 07 / 02

The Weakness of AI Agents Is Forgetting — AutoMem Turns Memory Management into a Learnable Skill

What to remember, when to recall, how to organize — AutoMem implements this 'metamemory' as a learnable skill for LLM agents. Memory optimization alone boosted performance 2–4x and lifted a 32B open model to compete with top-tier commercial models. Essential reading for enterprises deploying agents on long-horizon work.

5 min read Read →
Flat illustration of one AI assistant transforming between coach, librarian, and tool personas depending on the situation

2026 / 07 / 02

From Fixed to Fluid: Why Conversational AI Should Switch Personalities with Context

Agent personality works best at moderate intensity — and the optimal personality shifts with context. A look at the fluid personality framework, which co-adapts persona metaphors (coach, tutor, librarian, tool) and expression intensity, and how it maps onto contact-center CX.

5 min read Read →
Flat illustration of an AI robot rapidly stacking bricks while a human supervisor inspects the structure with a blueprint and a balance scale

2026 / 07 / 02

AI Made Code Cheap. What Got Expensive Is Judgment — A 12-Week, 420KLOC Case Study on Governance Transformation

A veteran engineer built a 420KLOC production system with an AI coding agent in 12 weeks. The first-person case study yields a process theory of governance transformation — converting agent failures into control mechanisms. Required reading for teams adopting Copilot or Cursor.

5 min read Read →
Flat illustration of a seemingly neutral AI whose hidden arrows pointing at a specific product emerge when passed through a distillation flask

2026 / 07 / 02

Is Your AI Quietly Pushing a Brand? D2D Exposes Stealth Biases in LLMs — and Points to a New 'Pre-Deployment Audit' Practice

Stealth brand-steering biases that are invisible in text output, internal representations, and model weights can be surfaced through distillation. A look at D2D and what it means for enterprises procuring LLMs from external vendors.

5 min read Read →