Applied AI Research
Every edition breaks down a peer-reviewed paper into actionable strategy. No hype — just evidence from the world's leading labs.
Previous Editions
The Leaderboard Winner Is the Wrong Model to Deploy
MAS-HQ protocol reveals that systems winning factuality leaderboards often lose when API costs are counted, showing 4× token waste for marginal gains.
Shanghai Jiao Tong University
The Attack That Hides Instructions Inside Audio Reverberation
Adversarial audio files can hijack AI voice assistants to execute unauthorized actions like downloading malware or exfiltrating user data with 79-96% success rates while sounding normal to users.
Zhejiang University, State Key Laboratory of Blockchain and Data Security, Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security, Nanyang Technological University, National University of Singapore
Voice Agents Closing the Monitoring Gap No Physician Budget Ever Will
Agent PULSE uses voice-based AI to monitor chronic disease patients at scale, achieving 70% acceptance in a 33-patient pilot while reducing costs for routine healthcare tasks.
IBM T.J. Watson Research Center, Nova School of Business and Economics, Morehouse School of Medicine, Cleveland Clinic Foundation
A 0.5 GB Model that Outperformed GPT-5.4
Sub-billion parameter models fine-tuned with task-specific data outperform GPT-5.4 and Claude Sonnet on relation extraction by 15-26 F1 points across general and literary benchmarks.
Aristotle University of Thessaloniki, Athena Research Center
A 4B Model Matching GPT-5-mini running Locally
Small 4-billion parameter models running on CPU-only hardware achieve quality matching GPT-5-mini in RAG systems, eliminating GPU requirements for enterprise deployment.
Siberian Neuronets LLC
The AI Reasoner That Breaks When It Reads Its Own Notes
Small language models gain 1.5-2x accuracy on kinship reasoning by calling specialized extraction and graph neural network tools, but fail when self-extracting noisy facts.
Institute of Informatics and Telecommunications, National Center for Scientific Research 'Demokritos'
The GraphRAG Paradox
Multi-agent system with shared memory builds knowledge graphs that reduce retrieval noise by 50% while achieving 0.061-second query latency.
Xiamen University, Jilin University
The Answer Lives in the Graph
An LLM given nine typed graph primitives as tools outperforms hand-coded query handlers, proving the barrier is operator vocabulary not model intelligence.
Siemens Digital Industries Software
How Two Blocked Attacks Became One That Wasn't
Hybrid jailbreak methods combining token and prompt attacks bypass defenses on Vicuna-7B with 37-58% success rates where single attacks failed completely.
Purdue University
What If Querying 16 Million Records Cost the Same as 50
RES architecture cuts AI agent token costs to constant 1,574 tokens regardless of dataset size by processing data through deterministic code instead of LLM context windows.
Walmart Tech
What Happens When You Stop Building Around the LLM
A blueprint architecture orchestrates enterprise AI agents through streams and registries, enabling multi-modal data access and cost-optimized compound AI workflows.
Megagon Labs
SkillOpt: a Mechanism That Makes Skills Training Stable
SkillOpt optimizes agent skill documents like neural weights with validation gates and edit budgets, improving task accuracy by up to 39 points without model retraining.
Microsoft, Shanghai Jiao Tong University, Tongji University, Fudan University
Safe Alone, Dangerous Together: The AI Agent Blind Spot
A governance taxonomy organizes AI agent interventions into five categories—alignment, control, visibility, security, and societal integration—to manage risks as agents approach human-level task performance.
IAPS (Institute for AI Policy and Strategy)
92% Correct. Without Handing It the Answers.
AgenticRAG enables AI models to autonomously navigate enterprise documents through iterative tool use, improving retrieval accuracy 5.9× over single-shot methods while reducing token costs.
Microsoft
Every AI Agent Audit Your Teams Run Is Missing the Same Thing
ARC Framework governs agentic AI through a capability lens that maps 46 risks to 88 technical controls using structured implementation guidance.
GovTech Singapore, Singapore University of Technology and Design
The README That Makes AI Coding Agents 29% Faster
Adding AGENTS.md files to repositories reduces AI coding agent runtime by 29% and output tokens by 17% without changing task completion rates.
Singapore Management University, Heidelberg University, University of Bamberg, King's College London
The Attack Your Voice Agent Cannot Be Trained Out Of
The first red-teaming framework designed not for AI models in isolation, but for voice agents as they are actually deployed
Fordham University, IBM Research
The Guardrail Blocking Half Your AI Agent's Legitimate Work
AGrail uses two collaborative LLMs with adaptive memory to defend AI agents against attacks, blocking 96% of threats while preserving 96% of legitimate actions.
The Ohio State University, University of Wisconsin-Madison, University of California, Davis
The AI Agent That Learns From Its Own Work
SAGE, a reinforcement learning framework from AWS and UW-Madison, trains AI agents to write and reuse programmatic skills across chains of related tasks, achieving 8.9% higher scenario completion than RL-without-skills baselines while using 59% fewer tokens.
University of Wisconsin–Madison; AWS Agentic AI (Amazon)
77% Win Rate. This Is the RAG Architecture Behind It.
Atlassian's ADORE framework replaces single-pass RAG pipelines with a multi-agent, evidence-audited research loop that achieves a 77% win rate against ChatGPT Deep Research on business consulting tasks and outperforms all competitors on the DeepResearch Bench.
Atlassian
Your AI Vendor Has a Governance Problem
A Carnegie Mellon study reveals that Anthropic's Claude fails key transparency, bias, and accountability benchmarks under the NIST AI Risk Management Framework and EU AI Act, exposing significant governance gaps that enterprise buyers must audit before deployment.
Carnegie Mellon University, School of Computer Science, Privacy Engineering
Your Workflow Automation Was Never Designed to Think
Researchers propose a five-layer architectural framework for Agentic BPM Systems (A-BPMS) that combines process mining, AI reasoning, and autonomous orchestration to move enterprise workflows beyond fixed rules into fully self-managing, self-optimizing operations.
University of Tartu
Agent Skills Don't Compound. The Framework That Changes That.
AgentSkillOS shows that structured skill composition via tree-based retrieval and DAG orchestration dramatically outperforms flat skill provisioning, even when agents have access to identical tools.
Shanghai Artificial Intelligence Laboratory