All Articles
16116 articles total
The Easy Trap: Why LLMs Underestimate Misconception-Driven Difficulty
A study reveals that LLMs underestimate the difficulty of educational tasks driven by student misconceptions, a phenomenon termed the 'Easy Trap'.
The Human Utility Factor: A Computable Welfare Metric That Reframes AI Governance as a Constrained Optimisation Problem
The Human Utility Factor (HUF) is proposed as a computable welfare metric to treat AI governance as a constrained optimization problem focused on socioeconomic stability.
AI Security Priorities: A Field-Wide Agenda
A multi-sector expert agenda is presented to prioritize AI security efforts, focusing on technical assurance and the governance of agentic AI.
SimpleWikiSearch: A Clean Offline Wikipedia Environment for Agentic Search
SimpleWikiSearch provides a controlled, offline environment for reproducible agentic-search evaluation using a cleaned Wikipedia dump.
GuidedRAG: Semantic Steering of Retrieval-Augmented Generation
GuidedRAG introduces semantic steering and a selection stage to improve retrieval relevance and reduce overhead in Retrieval-Augmented Generation.
IFCMemoryBench: Evaluating Long-Term Memory of LLM-Based Agents in BIM Information Retrieval
IFCMemoryBench is a new benchmark for evaluating long-term memory in LLM agents within the specialized domain of Building Information Modelling (BIM).
IDP AutoOpt: Agent-Driven Optimization of Document Processing Pipeline Configurations
IDP AutoOpt is an autonomous LLM agent that optimizes configurations for intelligent document processing pipelines, reducing tuning time from weeks to hours.
FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents
FinCacheServe is a cost-efficient RAG serving system that uses dependency-consistent answer reuse to reduce LLM calls for mutable enterprise documents.
DeepSeek-V4-Flash Update
DeepSeek-V4-Flash update discussed in Hacker News comments.
Simulating TCP loss and congestion in browser using Go/WASM
A project demonstrating how to simulate TCP loss and congestion within a web browser using Go and WebAssembly.
Predict before you train: Scaling Laws for particle physics foundation models
Researchers develop scaling laws for particle physics foundation models, allowing performance prediction of large models based on small-model training compute.
Forensic Reproducibility Audit of a Radiology Vision-Language Model Benchmark: From Intended Protocol to Released Artifact
A forensic audit reveals reproducibility issues in a radiology vision-language model benchmark, leading to the withdrawal of original performance claims.
Emergent Sparsity in Frozen Random CNN Feature Extractors for Deep Reinforcement Learning
Study finds that deep reinforcement learning agents using frozen random CNN feature extractors spontaneously develop sparse representations of task-relevant information.
Large-Scale ChatBot Validation Through Customer Digital Twin Simulations
A new framework for validating LLM chatbots in banking using synthetic customer digital twins to simulate diverse interaction styles.
Sim2Win: A Team-Agnostic, Event-Based Pre-Match Outcome Prediction and Tactical Profiling System for Football
Sim2Win is a team-agnostic football outcome prediction system using tactical profiles and event data instead of team identities.
Identifying Implicit Bias in LLM-based Chat AI Toward People with Intellectual Disabilities
Research identifies negative implicit biases toward people with intellectual disabilities in popular LLMs, including GPT-4o and Llama-3.
Archetypes or ability? Clustering for modelling student mathematical competence
A study on student mathematical competence suggests overall ability is the dominant factor for performance, challenging assumptions about discrete skill sets.
The Age of AI Agents Demands A New Scientific Paradigm To Sustain Trustworthy Science
An argument for a new scientific paradigm to maintain trust in science as AI agents begin to autonomously generate hypotheses and design experiments.
JEP 401: Value Objects (Preview) merged to OpenJDK master
JEP 401 introduces Value Objects to OpenJDK, providing a way to define classes that are treated as values rather than identities.
Human Honeypot
An exploration or implementation of a 'Human Honeypot', likely focused on cybersecurity or social engineering detection.