safety

Training-Time Batch Normalization Reshapes Local Partition Geometry in Piecewise-Affine Networks

arXiv:2605.04946v1 Announce Type: new Abstract: Batch normalization (BN) is central to modern deep networks, but its effect on the realized function during training remains less understood than its op

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Two years ago. I stand by both predictions.

DGX agent

Gary Marcus reflects on predictions he made two years prior and reaffirms his confidence in their accuracy. The post suggests Marcus is reviewing his track record on forecasts, likely related to artif

safetygary-marcus--x

7 May 2026

UAV-VL-R1: Generalizing Vision-Language Models via Supervised Fine-Tuning and Multi-Stage GRPO for UAV Visual Reasoning

arXiv:2508.11196v2 Announce Type: replace Abstract: Recent advances in vision-language models (VLMs) have demonstrated strong generalization in natural image tasks. However, their performance often de

UI2Code^N: UI-to-Code Generation as Interactive Visual Optimization

arXiv:2511.08195v3 Announce Type: replace Abstract: UI-to-code aims to translate UI screenshots into executable front-end code. Despite progress with vision-language models (VLMs), most existing metho

ULF-Loc: Unbiased Landmark Feature for Robust Visual Localization with 3D Gaussian Splatting

arXiv:2605.04730v1 Announce Type: new Abstract: Visual localization is a core technology for augmented reality and autonomous navigation. Recent methods combine the efficient rendering of 3D Gaussian

Uncertainty-Aware Exploratory Direct Preference Optimization for Multimodal Large Language Models

arXiv:2605.04874v1 Announce Type: cross Abstract: Direct Preference Optimization (DPO) has proven to be an effective solution for mitigating hallucination in Multimodal Large Language Models (MLLMs) b

safetyarxiv-cs-cl

Unifying Dynamical Systems and Graph Theory to Mechanistically Understand Computation in Neural Networks

arXiv:2605.03598v2 Announce Type: cross Abstract: Understanding how biological and artificial neural networks implement computation from connectivity is a central problem in neuroscience and machine l

UniMoCo: Unified Modality Completion for Robust Multi-Modal Embeddings

arXiv:2505.11815v2 Announce Type: replace Abstract: Current vision-language models have been explored for multi-modal embedding tasks like information retrieval. However, they face significant challen

Using Common Random Numbers for Simulation-based Planning with Rollouts

arXiv:2605.04732v1 Announce Type: new Abstract: Simulation-based planning with rollouts is a widely-deployed technique for decision making in stochastic environments. The primary instrument of simulat

Variance Matters: Improving Domain Adaptation via Stratified Sampling

arXiv:2512.05226v2 Announce Type: replace Abstract: Domain shift remains a key challenge in deploying machine learning models to the real world. Unsupervised domain adaptation (UDA) aims to address th

What happens when AIs become smarter than us? Why would they keep humans around if given the choice? Our new paper argues that only trying t…

What happens when AIs become smarter than us? Why would they keep humans around if given the choice? Our new paper argues that only trying to control AIs is a limited strategy, and that a stable, mutu

safetydan-hendrycks--x

When Life Gives You BC, Make Q-functions: Extracting Q-values from Behavior Cloning for On-Robot Reinforcement Learning

arXiv:2605.05172v1 Announce Type: new Abstract: Behavior Cloning (BC) has emerged as a highly effective paradigm for robot learning. However, BC lacks a self-guided mechanism for online improvement af

safetyarxiv-cs-ro

Why Expert Alignment Is Hard: Evidence from Subjective Evaluation

arXiv:2605.04972v1 Announce Type: new Abstract: Aligning large language models with expert judgment is especially difficult in subjective evaluation tasks, where experts may disagree, rely on tacit cr

safetyarxiv-cs-cl

Worth remembering when @janleike quit over OpenAI safety concerns.

Jan Leike, OpenAI's safety and alignment lead, resigned in May 2024, citing concerns about the company's commitment to safety practices and the deprioritization of safety work relative to product deve

Wow this paper has been “just published” many times for nearly a year. I have called out at least two other “influencers” for the same thing…

Wow this paper has been “just published” many times for nearly a year. I have called out at least two other “influencers” for the same thing on the same paper, including another one earlier this week.

yep, really embarassing

yep, really embarassing One of the richest VCs, Marc Andreessen, has shown again that he thinks generative AI works perfectly well if you use the right prompts, thus ignoring the statistical nature of

100% stood the test of time: “What Ilya saw” was Sam’s bad behavior, not AGI.

100% stood the test of time: “What Ilya saw” was Sam’s bad behavior, not AGI. When Altman got fired “What did Ilya see?” became a wildly popular conspiracy meme. I think we can safely say now that wha

2nd episode of The Roman Forum is an interview with AI Safety/Governance expert Connor Leahy @NPCollapse. Connor is a great speaker and is l…

2nd episode of The Roman Forum is an interview with AI Safety/Governance expert Connor Leahy @NPCollapse. Connor is a great speaker and is lobbying to get government to ban Superintelligence. My first

safetyconnor-leahy--x

A Knowledge-Driven LLM-Based Decision-Support System for Explainable Defect Analysis and Mitigation Guidance in Laser Powder Bed Fusion

arXiv:2605.01100v1 Announce Type: new Abstract: This work presents a knowledge-driven decision-support system that integrates structured defect knowledge with LLM-based reasoning to provide explainabl

A Robust Unsupervised Domain Adaptation Framework for Medical Image Classification Using RKHS-MMD

arXiv:2605.03787v1 Announce Type: new Abstract: Labeling medical images is a major bottleneck in the field of medical imaging, as it requires domain-specific expertise, and it gets further complicated

A Three-Stage Offline SDRE-Based Control Framework for Human Motion Reproduction on a Suspended Bipedal Robot

arXiv:2506.04680v2 Announce Type: replace Abstract: During the development of wearable exoskeletons, evaluations involving human subjects pose inherent safety risks. Therefore, systematic testing is o

safetyarxiv-cs-ro

A Universal Reproducing Kernel Hilbert Space from Polynomial Alignment and IMQ Distance

arXiv:2605.03262v1 Announce Type: new Abstract: We introduce the Yat kernel $k_{b,arepsilon}(mathbf{w},mathbf{x})=frac{(mathbf{w}^opmathbf{x}+b)^2}{|mathbf{x}-mathbf{w}|^2+arepsilon},qquad bge 0, arep

A US appeals court strikes down a 2023 FCC rule banning broadband access discrimination based on income, race, and more; Chair Brendan Carr welcomes the ruling (Jon Brodkin/Ars Technica)

Jon Brodkin / Ars Technica: A US appeals court strikes down a 2023 FCC rule banning broadband access discrimination based on income, race, and more; Chair Brendan Carr welcomes the ruling — An appeals

safetytechmeme

A Vision-Based Shared-Control Teleoperation Scheme for Controlling the Robotic Arm of a Four-Legged Robot

arXiv:2508.14994v3 Announce Type: replace-cross Abstract: In hazardous and remote environments, robotic systems perform critical tasks demanding improved safety and efficiency. Among these, quadruped

Adaptive 3D-RoPE: Physics-Aligned Rotary Positional Encoding for Wireless Foundation Models

arXiv:2605.00968v1 Announce Type: cross Abstract: Positional encoding plays a pivotal role in determin?ing the extrapolation and generalization performance of wireless foundation models for channel st

ADAPTS: Agentic Decomposition for Automated Protocol-agnostic Tracking of Symptoms

arXiv:2605.03212v1 Announce Type: cross Abstract: Modeling latent clinical constructs from unconstrained clinical interactions is a unique challenge in affective computing. We present ADAPTS (Agentic

safetyarxiv-cs-cl

Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models

arXiv:2605.02911v1 Announce Type: new Abstract: Future sixth-generation (6G) mobile networks are envisioned to be equipped with a diverse set of powerful, yet highly specialized, optimization experts.

Algebraic Semantics of Governed Execution: Monoidal Categories, Effect Algebras, and Coterminous Boundaries

arXiv:2605.01032v2 Announce Type: new Abstract: We present an algebraic semantics for governed execution in which governance is axiomatized, compositional, and coterminous with expressibility. The fra

Aligning Inductive Bias for Data-Efficient Generalization in State Space Models

arXiv:2509.20789v4 Announce Type: replace Abstract: The remarkable success of modern AI has been closely tied to scaling laws, yet the finite supply of high-quality data makes data efficiency--learnin

🚨Alphabet $GOOGL trades at 133x free cash flow. For context: its pre-COVID multiple was ~20x. And free cash flow hasn't grown since 2021. G…

🚨Alphabet GOOGL trades at 133x free cash flow. For context: its pre-COVID multiple was ~20x. And free cash flow hasn't grown since 2021. GQG Partners — one of the world's top institutional investors —

Am I right that hyperscaling compute is the biggest bet in history? Any counter examples? It’s way more expensive than the Manhattan Project…

Am I right that hyperscaling compute is the biggest bet in history? Any counter examples? It’s way more expensive than the Manhattan Project, the Apollo project, and railways across the US. If it does

Amazing the shit X gave me in November 2023 for saying Sam wasn’t always candid. He wasn’t. Period.

Amazing the shit X gave me in November 2023 for saying Sam wasn’t always candid. He wasn’t. Period. Murati testified that, by fall 2023, Sam Altman was “not always” candid with her, and undermined her

And how many times have I told you that Sam is no longer the right CEO for OpenAI?

And how many times have I told you that Sam is no longer the right CEO for OpenAI? Murati: My issues with Sam were very much around management and providing direction to the organization and decision

Anthropic researchers detail 'model spec midtraining', which adds a stage between pretraining and fine-tuning to improve generalization from alignment training (Anthropic)

Anthropic: Anthropic researchers detail “model spec midtraining”, which adds a stage between pretraining and fine-tuning to improve generalization from alignment training — Sara Price2, Samuel Marks2,

safetytechmeme

Architectural Obsolescence of Unhardened Agentic-AI Runtimes

arXiv:2605.01740v1 Announce Type: cross Abstract: An agentic-AI runtime issues tool calls, sends messages, and actuates devices on behalf of an LLM. Catching the four ways an action can diverge from i

Audio-Visual Intelligence in Large Foundation Models

arXiv:2605.04045v1 Announce Type: new Abstract: Audio-Visual Intelligence (AVI) has emerged as a central frontier in artificial intelligence, bridging auditory and visual modalities to enable machines

Beyond Activation Alignment: The Geometry of Neural Sensitivity

arXiv:2605.03222v1 Announce Type: new Abstract: Activation-alignment measures such as Representational Similarity Analysis (RSA), Canonical Correlation Analysis (CCA), and Centered Kernel Alignment (C

BifrostUMI: Bridging Robot-Free Demonstrations and Humanoid Whole-Body Manipulation

arXiv:2605.03452v1 Announce Type: new Abstract: High-quality data collection is a fundamental cornerstone for training humanoid whole-body visuomotor policies. Current data acquisition paradigms predo

safetyarxiv-cs-ro

Bootstrapped Mixed Rewards for RL Post-Training: Injecting Canonical Action Order

arXiv:2512.04277v3 Announce Type: replace Abstract: Post-training with reinforcement learning (RL) typically optimizes a single scalar objective and ignores structure in how solutions are produced. We

Can Semantic Methods Enhance Team Sports Tactics? A Methodology for Football with Broader Applications

arXiv:2601.00421v2 Announce Type: replace Abstract: This paper explores how semantic-space reasoning, traditionally used in computational linguistics, can be extended to tactical decision-making in te

Catching the Infection Before It Spreads: Foresight-Guided Defense in Multi-Agent Systems

arXiv:2605.01758v1 Announce Type: new Abstract: Large multimodal model-based Multi-Agent Systems (MASs) enable collaborative complex problem solving through specialized agents. However, MASs are vulne

Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents

arXiv:2604.06132v2 Announce Type: replace Abstract: Large language models are increasingly deployed as autonomous agents for multi-step workflows in real-world software environments. However, existing

C’mon BBC. Zilis was sharp as a tack on the stand, on her role in the OpenAI nonprofit board, and how she managed conflicts as they began …

C’mon BBC. Zilis was sharp as a tack on the stand, on her role in the OpenAI *nonprofit* board, and how she managed conflicts as they began to develop, and this (which is not really even news since it

Coherent Hierarchical Multi-Label Learning to Defer for Medical Imaging

arXiv:2605.02734v1 Announce Type: new Abstract: Learning to Defer (L2D) enables a model to predict autonomously or defer to an expert, but prior work largely assumes flat label spaces. We study the fi

Correction: the jury is advisory only. It’s the judge who decides; if she sees it as I do, OpenAI loses.

Gary Marcus clarifies that in the legal proceeding he's discussing, the jury serves an advisory role while the judge retains decision-making authority on the case outcome. Marcus expresses confidence

Deciphering Shortcut Learning from an Evolutionary Game Theory Perspective

arXiv:2605.02658v2 Announce Type: new Abstract: Shortcut learning causes deep learning models to rely on non-essential features within the data. However, its formation in deep neural network training

Descent-Guided Policy Gradient for Scalable Cooperative Multi-Agent Learning

arXiv:2602.20078v3 Announce Type: replace-cross Abstract: Scaling cooperative multi-agent reinforcement learning (MARL) is fundamentally limited by cross-agent noise. When agents share a common reward

DGPO: Distribution Guided Policy Optimization for Fine Grained Credit Assignment

arXiv:2605.03327v1 Announce Type: new Abstract: Reinforcement learning is crucial for aligning large language models to perform complex reasoning tasks. However, current algorithms such as Group Relat