Chatbot Spots Romance Scams in Real-Time on Line
A new chatbot uses stage classification to detect romance scams as they happen in Line conversations.
Research Digest
Preprints and papers, rewritten in plain English.
RICE-Alpha: Event Graphs Improve LLM Stock Forecasts
A new framework models temporal dependencies in corporate events to make LLM-based stock forecasting more reliable.
Data-Informed Hydra-Nets for Heterogeneous Federated LLM Training
A new arXiv paper proposes a model-heterogeneous federated learning approach for LLMs using Hydra-Nets shaped by client data.
ReasonUni Adds Reasoning to Reconstruction for Unified Multimodal Models
A new training approach couples image understanding and generation by injecting reasoning into reconstructive post-training.
Interpreting Language Model Parameters, Not Just Activations
A new arXiv paper focuses on how neural networks use their parameters directly, moving beyond the usual study of intermediate representations.
Thermal Interfaces in AVs Are Not Security Boundaries, Survey Finds
A new survey highlights how thermal management and night-vision interfaces in autonomous vehicles create overlooked side-channel and perception attack surfaces.
JevAgentHarness aims to cut costly LLM calls in agent harnesses
A new framework proposes training smaller decision models to share control with LLMs, potentially reducing cost and latency in long agent tasks.
Latent World Models Learn Intrinsic Dynamics Despite Visual Noise
New work proposes learning action-consequence dynamics in latent space to avoid distraction from irrelevant visual variation.
ARGUS Framework Aims to Make LLM User Simulators More Reliable
A new agenda-driven framework targets reliability failures in LLM-based user simulators used for agent training and evaluation.
ClaimScope Benchmark Audits Mathematical Claims Beyond Proofs
A new 72-task benchmark, ClaimScope, tests mathematical agents on issuing verdicts, finding counterexamples, and proposing repairs for claims.
LLM Agents Need to Forget: MemCut Targets Long-Term Memory
A new arXiv paper argues that dependable long-term memory in LLM agents requires the ability to forget, not just retrieve.
LLMs and Embeddings for Ontology Refinement: CASTON Case Study
A new arXiv study explores using embeddings and large language models to refine class hierarchies in biomedical ontologies.
Localized Context Distillation Helps LLMs Learn from Interaction
A new arXiv paper proposes a localized form of context distillation to let large language models improve from real-world interaction experience.
Separating Decisions from Generation in Multi-Agent LLM Workflows
A new framework proposes using lightweight decision models for coordination tasks in multi-agent systems, reserving generative LLMs for expert execution.
Agent Status Protocol: Lightweight State Reporting for AI Agents
A proposed protocol lets heterogeneous AI agents report operational state to a central server, enabling real-time dashboards and monitoring.
Graph Convolutional Network Flags Low-Quality Documents by Consistency
A new arXiv paper proposes a hierarchical graph convolutional network that assesses document quality by modelling internal consistency.
QREF Aims to Fix Misaligned Search Queries in DeepSearch Agents
A proposed query-refinement and evidence-filtering method targets the common failure mode of intent-misaligned and redundant searches in iterative agents.
MetaExp: Teaching Language Agents to Turn Feedback into Reusable Experience
A new arXiv paper proposes MetaExp, a method for language-model agents to self-improve by converting interaction feedback into reusable experience through verifiable cross-task environment scaling.