FERPO: Entropy-Regularized Policy Optimization for Continuous Control
A new RL method targets the mismatch between critic training and policy improvement in continuous-control tasks.
Research Digest
Preprints and papers, rewritten in plain English.
Hybrid Meta-Learning and Autoencoder Approach for Ransomware Detection
A new framework pairs few-shot meta-learning with autoencoders to detect ransomware without large training sets.
UniBuc Uses Tailored Prompting for Clinical NLI
A team shows that a carefully prompted SOLAR Instruct model can tackle biomedical natural language inference without fine-tuning.
LLM-as-a-Judge for Romanian RAG: A New Evaluation Approach
A new paper explores using LLM-as-a-Judge to evaluate retrieval-augmented generation systems in Romanian, a low-resource language where standard metrics fall short.
Signed Lexical Confidence for Risk-Calibrated Intent Routing
A new method augments standard confidence scores with lexical signals to help AI assistants decide when to act and when to defer.
Federated Learning Approach Detects Cyberattacks on EV Charging Protocol
A new arXiv paper proposes a federated detection method for cyberattacks targeting the Open Charge Point Protocol 1.6 used by EV charging stations.
EurekaBench: Benchmarking AI’s Ability to Discover Scientific Insights
A new arXiv paper introduces EurekaBench, a benchmark designed to measure how well AI agents can perform the iterative, pattern-finding work behind scientific discovery.
IndoorBEV: Lightweight LiDAR Perception for Indoor Robots
A new bird's-eye-view perception system aims to give indoor mobile robots real-time LiDAR understanding under tight latency and memory limits.
New Attacks Target Imperfectly Complete QCCC Key Agreement
The first unconditional polynomial-query attacks against quantum-computation, classical-communication key agreement work even when the protocol has imperfect completeness.
Hybrid Quantum-Classical Models for Brain Deformation Prediction
A new arXiv study tests whether hybrid quantum-classical machine learning can make high-dimensional brain deformation prediction tractable.
Geometric Perturbation Keeps Collaborative Learning Private
A preprint describes a privacy-preserving method that lets participants share transformed data representations for collaborative learning.
Invariant-guided diffusion for tabular data under covariate shift
A new data-augmentation framework targets scarce, imbalanced tabular data when training and deployment distributions differ.
LLMs Show Query-Language Preference in Multilingual Search
A new arXiv paper examines how large language models favour the query language when synthesising multilingual evidence for knowledge-intensive tasks.
Bayesian fine-tuning makes language models better at hidden-variable reasoning
A new arXiv study shows that fine-tuning language models with Bayesian objectives improves their ability to infer hidden variables from limited observations.
Evaluating Context Strategies for Long-Horizon Clinical Reasoning in LLMs
A new arXiv study examines how well LLMs integrate evidence across extended patient histories.
Direct-Decision Models Show Ordinal-Scale Bias
A new study argues that accuracy alone is insufficient for evaluating direct-decision models, which must also respect ordinal scales.
Stale Binding: Why LLMs Stick to Old Values When Context Changes
A new paper identifies 'stale binding' as a failure mode where LLMs answer with outdated variable values even after context updates.
STRATA: A Tiered Multi-Agent Framework for Real-Time Strategy Games
A new arXiv paper proposes STRATA, a self-learning system that uses role-aligned tiered agents to handle the complex long-horizon coordination demands of real-time strategy games.