Agentic AI for Straight-Through Underwriting
An arXiv study proposes an agentic pipeline for straight-through underwriting that compares a single-LLM approach, naive RAG, and a multi-agent “Agentic RAG” system with targeted retrieval, third-party checks, and explicit multi-step rule evaluation. The agentic setup delivers the biggest improvements when information is missing or requires multi-step reasoning, aiming to make automated decisions more auditable and human-in-the-loop friendly.
Context Graphs for Proactive Enterprise Agents
Instead of waiting for users to ask, this work introduces “Context Graphs” – a live relational structure that tracks enterprise entities, relationships, and state changes over time. A Delta Detection Engine and Proactivity Scorer continuously find urgent, relevant updates, and an LLM-based surfacing layer delivers grounded notifications, reporting Precision@5 of 0.83 and cutting time-to-surface from 47 minutes to under 30 seconds.
Infinity-Parser2: A Multimodal Document Parsing Leap
Infinity-Parser2 couples controllable data synthesis with multi-task reinforcement learning to tackle the bottleneck of scarce, faithful document parsing annotations. The team builds Infinity-Doc2-5M (5M bilingual samples) and reports strong results including 87.6% on olmOCR-Bench and 74.3% on ParseBench, with variants tuned for low-latency or precision-critical deployments.
Alignment Plausibility for Safer AI in Healthcare
This technical proposal argues that today’s mental-health LLM safety work is too reactive to acute harms and insufficient for longer-term risks like dependency or boundary erosion. It introduces “alignment plausibility” as a regulatory-style construct – requiring consistent value specification, value-embedded training, and deployment oversight that detects drift and longer-term harm – to support trust that systems are aligned to positive health outcomes.
Meta Turns Off Instagram Feature That Enabled AI Deepfakes From Public Accounts
After backlash, Meta disabled a feature tied to its Muse Image AI model that allowed users to generate images by @-mentioning public Instagram accounts, a workflow critics said enabled creation without the account owner’s permission. The Verge reports Meta’s reversal, highlighting how quickly AI image tools can collide with consent, rights, and deepfake fears.
Apple Sues OpenAI Over Alleged Trade-Secret Theft Tied to Hardware Plans
Apple has filed suit alleging that OpenAI engineers stole Apple trade secrets after moving from Apple, naming OpenAI and other parties including Jony Ive’s IO Products along with specific employees. The complaint frames the issue as misappropriation connected to OpenAI’s hardware strategy, escalating the competitive – and legal – race around AI devices.
AI-Integrated Modeling for Agricultural Resilience
This work blends economic and biophysical simulation by integrating GTAP (economic models) with APSIM (biophysical crop/agronomy models) to analyze agricultural supply-chain shocks. The result is an AI-powered query-and-response interface that helps policymakers and market participants estimate cross-disciplinary impacts from disruptions.
Real-Time Hand Gesture Recognition With Graph Neural Networks From sEMG
Researchers propose a graph-network representation of forearm muscle activation patterns using sEMG, then apply a graph neural network for real-time gesture recognition. Using an 8-electrode myoband setup across eight subjects, the method reports 99% average classification accuracy with fast construction and prediction latency (48ms on an M1 Pro CPU), targeting prosthetics and AR control.
Adversarial Social Epistemology for Human-Agent Communication Networks
This paper argues that “misinformation” framing is too narrow for environments where assertions are scaffolded by chains of testimony, inference, certification, and tacit trust. It outlines adversarial social epistemology – focusing on how agents can subvert the auditability of inferential chains and exploit the mechanisms that normally make scaffolded information trustworthy.
Idiobionics: Unifying Privacy and Intelligent Robotic Prostheses
Introducing “idiobionics,” the authors examine how privacy risks emerge as bionic limbs become more perceptive and AI-controlled through tight coupling of biological and digital systems. The paper defines the field and discusses preliminary attack ideas and a set of open research questions aimed at making wearable autonomous devices safer and more adoption-ready.
VectorizationLLM: A RAG-Grounded Math Assistant for MATLAB Analysis
VectorizationLLM is an open-weight-model-based tutor designed to teach vectorization topics like Fourier analysis and differential equations, with a RAG knowledge base and a system-prompt structure that emphasizes explanations and examples rather than handing over direct answers. The course-oriented setup targets learning outcomes for MATLAB-centric computational analysis.
Aligning Clinical Needs and AI Capabilities: A Survey on LLM Medical Reasoning
This survey maps LLM medical reasoning onto a dual framework: clinical competencies aligned to Miller’s Pyramid and computational reasoning types (deductive, inductive, abductive) tied to medical tasks. It benchmarks 18 models across a five-level reasoning dataset, reporting a split where specialist models tend to lead in diagnosis-centric work while general models shine in decision support and dialogue, alongside persistent issues like hallucination and grounding.