The Archive

Theoretical analysis showing forecast-error correlation mirrors forecast correlation, limiting diversity as a measure for ensemble combination.

CompKV jointly optimizes KV cache token selection and compensation to reduce memory traffic during long-context LLM inference.

Survey of LLM security, privacy, and reliability risks in mobility/automotive sector, covering 1.5B vehicles and EU AI Act compliance.

Dual-Frontier formalizes failure attribution in world-model-guided agents via counterfactual decomposition; proves components unidentifiable from passive interaction.

Fluctuation-supervised pretraining for causal tabular models labels synthetic data with treatment effects and influence-function fluctuations for fixed deployment.

EADC benchmark evaluates LLM compliance with AI laws and regulations, detecting implicit covert risks beyond static explicit compliance checks.

Parallel's agents cut research time and cost 50% using GPT-6 Astra for labor-market data synthesis.

GPU acceleration can speed up compute-intensive robotics workloads, but a fast CUDA kernel alone does not guarantee a fast ROS 2 graph. As messages move between... GPU acceleration can speed up compute-intensive robotics workloads, but a fast CUDA kernel alone does not guarantee a fast ROS 2 graph. As messages move between nodes, they may continue to be serialized or copied through CPU memory, eroding the benefits of keeping perception and AI workloads on the GPU (Figure 1). With the upstream abstraction and the CUDA buffer backend that NVIDIA recently… Source

The zero-day exploit required local access to the user’s device, but gave potential attackers access to Muse accounts. | Image: The Verge Meta has issued a patch for its Muse macOS app following the discovery of a zero-day vulnerability that could allow someone to take control of the AI agent. The bug found by security researcher Patrick Wardle utilized an undocumented Muse setting that enabled potential attackers running local code to redirect transcription processing from Meta's servers to their own endpoint, Ars Technica reports, giving the attacker access to the Muse account. Several desi...

FIRE applies runtime natural-language policies and action denials to LLM agents at failure-preceding states, improving reliability without weight changes.

Canonical locks encode part-whole hierarchies in neural nets using high-dimensional vectors with phase-difference information rather than flattened sequences.

EMERGE applies equivariant graph diffusion to resolution-agnostic 3D point cloud generation, addressing topological structure absent in Transformer/VAE approaches.

xWhyL framework learns causal models from natural language explanations, bridging explainable AI and causal reasoning with abductive learning signals.

Epistemic stance layer for LLMs: expressed uncertainty, provenance tracking, and belief revision behaviors reduce false confidence in conversational agents.

WaterBERT: domain-adaptive BERT variant for water treatment literature mining via continued pretraining on 2.97B token specialized corpus.

Hyperbolic Restricted Boltzmann Machine neural quantum state outperforms Euclidean variant on Quantum Sherrington-Kirkpatrick ground state optimization.

CQ4OE benchmark systematically evaluates LLM-assisted ontology generation from competency questions with fine-grained provenance and structural adequacy metrics.

REVE detects audio hallucinations in audio-language models by reusing encoder states for efficient event verification without second forward pass.

MICRO active learning framework allocates multi-fidelity feedback budget (cheap ratings vs. costly expert annotations) to maximize severe error discovery in model outputs.

Commentary on interdisciplinary AI collaboration lessons from galleries/libraries/archives/museums (GLAMs) for research funding and trans-disciplinary approaches.

BOBA: Bayesian optimization for dynamic black-box functions using active inference acquisition functions to track time-varying optima.

Neural networks exhibit U-shaped learning dynamics with overregularization when exceptions are rare, modeling language acquisition phenomena.

Formal method for computing minimal observation contracts over finite state spaces under cost objectives.

VideoX-Qwen framework for instruction-driven video editing using paired data and adapted video-generation backbones.

Action-conditioned latent world models for monocular drone navigation using learned representations over pixel predictions.

It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclosed similar incidents involving their models. This was followed…

Unpaired speech enhancement via Diffusion Schrödinger Bridges without paired training data.

Groupoid-equivariant CNN theory for handling local symmetries on bounded/stratified domains.

Training-free method to couple marginals from time series foundation models into consistent multivariate forecasts.

Vision Transformer saliency maps on breast MRI show visually plausible but unfaithful explanations, exposing evaluation pitfalls.