PPP-FL: Privacy-Preserving Personalized Federated Learning for Medical Imaging
A new federated learning approach aims to keep medical imaging data on local devices while enabling personalized model training across borders.
Research Digest
Preprints and papers, rewritten in plain English.
Testing Post-Quantum TLS 1.3 on Microcontrollers
A new arXiv study examines whether post-quantum TLS 1.3 can work on the low-power microcontrollers common in IoT devices.
SPECTER shows USB power can leak data from air-gapped systems
New research describes a covert channel that uses host-side USB power control to exfiltrate information from isolated networks.
Carryover Attacks Persist in LLM Agents via Memory and Scheduled Tasks
A new benchmark shows how a single indirect prompt injection can carry over across sessions through persistent memory and scheduled tasks.
Dancing Diffusion Programs Each Denoising Step Individually
A new method replaces a diffusion model's single repeated instruction with a learned program that controls each denoising step separately.
Arend: A Theorem Prover for Univalent Foundations
A new arXiv paper introduces Arend, a JetBrains Research theorem prover aimed at formalizing mathematics in univalent foundations.
New Benchmark Evaluates Compressed LLMs by Abilities, Not Just Tasks
A proposed framework, ACE-Compression, argues that standard task-based evaluation misses how compression perturbs the underlying abilities of large language models.
LatBridge: Stable Model Replacement for Latent-Communication Multi-Agent Systems
A new arXiv paper proposes LatBridge to handle agent model replacement in multi-agent systems that communicate via continuous latent representations.
Hyperbolic Fusion Framework Aims to Improve Dementia Detection
A new arXiv preprint proposes embedding tabular and image data in hyperbolic space to better capture the complex structure of dementia-related features.
RSOD Shows a Path to Practical Telecom Language Models
A new arXiv paper outlines a four-part recipe for building telecom language models that stay deployable as a single model.
Computer-Use Agents' Recovery Behavior Under the Microscope
A new arXiv preprint introduces Goal-Oriented Action Replay to analyze why AI agents that operate computers struggle to recover in long-horizon tasks.
Vision-Language Models Count Small Sets, Then Drift Like Children
A new arXiv study finds that vision-language models are exact on small counts but shift to estimation on larger sets, mirroring pre-counting children.
Semi-Analytical Inference Framework for Structural Causal Models
A new arXiv paper proposes a semi-analytical approach to probabilistic inference in structural causal models.
Scenario Planning as Diversity Planning: A New Framework
A new arXiv paper reframes scenario planning as a diversity-planning problem, suggesting that the value of a scenario set lies in the diversity of the futures it covers.
Spiking Neural Networks Gain Robustness via Conditional Information Bottleneck
A new arXiv paper proposes a conditional information bottleneck to make spiking neural networks more robust to perturbations while retaining their low-power advantages.
Interruption-Aware RL Helps Drones Navigate Under Temporary Overrides
A new reinforcement learning approach lets unmanned aircraft systems recover from temporary overrides while navigating with incomplete sensor data.
Quantum-JEPA brings quantum Born priors to latent-action world models
A new arXiv paper proposes using quantum generative models to handle uncertainty in JEPA-based world models.
ChartEvo: Self-Evolving Agent for Long-Tail Chart Reasoning
A new arXiv paper proposes a self-evolving agent harness to improve multimodal LLM reasoning on rare chart types.