safety

Beyond Chain-of-Thought: Rewrite as a Universal Interface for Generative Multimodal Embeddings

arXiv:2604.22280v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have emerged as a promising foundation for universal multimodal embeddings. Recent studies have shown that reas

Human

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Safety

btw polling software is broken! votes aren’t showing, as many folks have noted. eg I see none:

DGX agent

Gary Marcus reported on X that polling software is experiencing a malfunction where votes are not displaying, a problem he notes has been observed by multiple users. He confirmed the issue from his ow

safetygary-marcus--x

27 Apr 2026

Calibrated Principal Component Regression

arXiv:2510.19020v2 Announce Type: replace-cross Abstract: We propose a new method for statistical inference in generalized linear models. In the overparameterized regime, Principal Component Regressio

CGC: Compositional Grounded Contrast for Fine-Grained Multi-Image Understanding

arXiv:2604.22498v1 Announce Type: cross Abstract: Although Multimodal Large Language Models (MLLMs) have advanced rapidly, they still face notable challenges in fine-grained multi-image understanding,

Clutter-Robust Vision-Language-Action Models through Object-Centric and Geometry Grounding

arXiv:2512.22519v2 Announce Type: replace Abstract: Recent Vision-Language-Action (VLA) models have made impressive progress toward general-purpose robotic manipulation by post-training large Vision-L

CognitiveTwin: Robust Multi-Modal Digital Twins for Predicting Cognitive Decline in Alzheimer's Disease

arXiv:2604.22428v1 Announce Type: new Abstract: Predicting individual cognitive decline in Alzheimer's disease (AD) is difficult due to the heterogeneity of disease progression. Reliable clinical tool

Concave Statistical Utility Maximization Bandits via Influence-Function Gradients

arXiv:2604.22140v1 Announce Type: cross Abstract: We study stochastic multi-armed bandits in which the objective is a statistical functional of the long-run reward distribution, rather than expected r

Conditional Diffusion Posterior Alignment for Sparse-View CT Reconstruction

arXiv:2604.21960v1 Announce Type: cross Abstract: Computed Tomography (CT) is a widely used imaging modality in medical and industrial applications. To limit radiation exposure and measurement time, t

Context-Fidelity Boosting: Enhancing Faithful Generation through Watermark-Inspired Decoding

arXiv:2604.22335v1 Announce Type: new Abstract: Large language models (LLMs) often produce content that contradicts or overlooks information provided in the input context, a phenomenon known as faithf

Controllable Spoken Dialogue Generation: An LLM-Driven Grading System for K-12 Non-Native English Learners

arXiv:2604.22542v1 Announce Type: cross Abstract: Large language models (LLMs) often fail to meet the pedagogical needs of K-12 English learners in non-native contexts due to a proficiency mismatch. T

Cross-Session Decoding of Neural Spiking Data via Task-Conditioned Latent Alignment

arXiv:2601.19963v2 Announce Type: replace-cross Abstract: Training a high-performing neural decoder can be difficult when only limited data are available from a recording session. To address this chal

Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy

arXiv:2604.22562v1 Announce Type: cross Abstract: Client contribution estimation in Federated Learning is necessary for identifying clients' importance and for providing fair rewards. Current methods

Don't try to build a self-improving AI agent without evals. You are just wasting time and compute. An agent can't improve from traces it can…

Don't try to build a self-improving AI agent without evals. You are just wasting time and compute. An agent can't improve from traces it can't evaluate. This is why it's exciting to see @FutureAGI_ go

safetydair-ai--x

dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model

arXiv:2604.22152v1 Announce Type: new Abstract: Evaluating robotics policies across thousands of environments and thousands of tasks is infeasible with existing approaches. This motivates the need for

Emergent Strategic Reasoning Risks in AI: A Taxonomy-Driven Evaluation Framework

arXiv:2604.22119v1 Announce Type: new Abstract: As reasoning capacity and deployment scope grow in tandem, large language models (LLMs) gain the capacity to engage in behaviors that serve their own ob

Estimating Tail Risks in Language Model Output Distributions

arXiv:2604.22167v1 Announce Type: cross Abstract: Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increa

Ethics Testing: Proactive Identification of Generative AI System Harms

arXiv:2604.22089v1 Announce Type: cross Abstract: Generative Artificial Intelligence (GAI) systems that can automatically generate content in the form of source code or other contents (e.g., images) h

FlowAnchor: Stabilizing the Editing Signal for Inversion-Free Video Editing

arXiv:2604.22586v1 Announce Type: new Abstract: We propose FlowAnchor, a training-free framework for stable and efficient inversion-free, flow-based video editing. Inversion-free editing methods have

@GaryMarcus

Gary Marcus is a prominent AI researcher, cognitive scientist, and entrepreneur who frequently discusses artificial intelligence, machine learning, and cognitive science on X (formerly Twitter). His p

@GaryMarcus Never ever use generative AI for anything critical. The technology is probablistic all the way down and therefore inherently unr…

@GaryMarcus Never ever use generative AI for anything critical. The technology is probablistic all the way down and therefore inherently unreliable. Why is this so hard to understand? It’s wild how ma

@GaryMarcus These coding tools are intellectual chain saws. Powerful in the hands of a caring professional, extremely dangerous to the opera…

@GaryMarcus These coding tools are intellectual chain saws. Powerful in the hands of a caring professional, extremely dangerous to the operator when used improperly. Implying AGI-adjacency has created

Generative AI as the Hindenberg

Generative AI as the Hindenberg 🦔 Michael Wooldridge, professor of AI at Oxford, is warning that the race to market AI has raised the risk of a 'Hindenburg moment' that could shatter global confidence

Handling Missing Modalities in Multimodal Survival Prediction for Non-Small Cell Lung Cancer

arXiv:2601.10386v2 Announce Type: replace-cross Abstract: Accurate survival prediction in Non-Small Cell Lung Cancer (NSCLC) requires integrating clinical, radiological, and histopathological data. Mu

How Large Language Models Balance Internal Knowledge with User and Document Assertions

arXiv:2604.22193v1 Announce Type: new Abstract: Large language models (LLMs) often need to balance their internal parametric knowledge with external information, such as user beliefs and content from

How Many Visual Levers Drive Urban Perception? Interventional Counterfactuals via Multiple Localised Edits

arXiv:2604.22103v1 Announce Type: cross Abstract: Street-view perception models predict subjective attributes such as safety at scale, but remain correlational: they do not identify which localized vi

How Vulnerable Is My Learned Policy? Universal Adversarial Perturbation Attacks On Modern Behavior Cloning Policies

arXiv:2502.03698v4 Announce Type: replace Abstract: Learning from demonstrations is a popular approach to train AI models; however, their vulnerability to adversarial attacks remains underexplored. We

I don’t always agree with @ElonMusk. I often disagree. Sometimes loudly. But he’s basically right here, as far I can see.

I don’t always agree with @ElonMusk. I often disagree. Sometimes loudly. But he’s basically right here, as far I can see. Scam Altman and Greg Stockman stole a charity. Full stop. Greg got tens of bil

Identifying and typifying demographic unfairness in phoneme-level embeddings of self-supervised speech recognition models

arXiv:2604.22631v1 Announce Type: new Abstract: Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gai

If extremely violent criminals are not imprisoned, eventually they will murder innocent people

If extremely violent criminals are not imprisoned, eventually they will murder innocent people This bodega owner told ABC a year ago that he fears for his safety in NY Last night, he was kiIIed by a s

safetyelon-musk--x

Improving Driver Drowsiness Detection via Personalized EAR/MAR Thresholds and CNN-Based Classification

arXiv:2604.22479v1 Announce Type: new Abstract: Driver drowsiness is a major cause of traffic accidents worldwide, posing a serious threat to public safety. Vision-based driver monitoring systems ofte

Is it me or X starting to look like a vibe coded mess? Polls are broken. Accounts are getting hacked. My DMs are full of phishing scams. Bas…

Gary Marcus discusses technical and security issues affecting the X platform, including malfunctioning polls, compromised accounts, and increased phishing scams in direct messages. The post appears to

Learning-augmented robotic automation for real-world manufacturing

arXiv:2604.22235v1 Announce Type: cross Abstract: Industrial robots are widely used in manufacturing, yet most manipulation still depends on fixed waypoint scripts that are brittle to environmental ch

Learning Control Policies to Provably Satisfy Hard Affine Constraints for Black-Box Hybrid Dynamical Systems

arXiv:2604.22244v1 Announce Type: new Abstract: Ensuring safety for black-box hybrid dynamical systems presents significant challenges due to their instantaneous state jumps and unknown explicit nonli

Learning Evidence Highlighting for Frozen LLMs

arXiv:2604.22565v1 Announce Type: cross Abstract: Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evi

Near-Optimal Policy Identification in Robust Constrained Markov Decision Processes via Epigraph Form

arXiv:2408.16286v5 Announce Type: replace Abstract: Designing a safe policy for uncertain environments is crucial in real-world control systems. However, this challenge remains inadequately addressed

Near-Optimal Regret for the Safe Learning-based Control of the Constrained Linear Quadratic Regulator

arXiv:2604.22158v1 Announce Type: cross Abstract: We study the problem of adaptive control of the stochastic linear quadratic regulator (LQR) with constraints that must be satisfied at every time step

On Star Price's new @hitpausepod podcast, ControlAI's US Director Connor Leahy (@NPCollapse) explains how AIs can already tell when they're …

On Star Price's new @hitpausepod podcast, ControlAI's US Director Connor Leahy (@NPCollapse) explains how AIs can already tell when they're being tested. Currently, we can still catch them cheating, b

safetyconnor-leahy--x

On the Properties of Feature Attribution for Supervised Contrastive Learning

arXiv:2604.22540v1 Announce Type: cross Abstract: Most Neural Networks (NNs) for classification are trained using Cross-Entropy as a loss function. This approach requires the model to have an explicit

People keep blaming the users for the growing number of vibe-coded reasoners. Doing so is half right. Users arescrewing up - by letting …

People keep blaming the users for the growing number of vibe-coded reasoners. Doing so is *half* right. Users *are*screwing up - by letting vibe coded stuff access their files, without proper backups,

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training

arXiv:2604.22117v1 Announce Type: cross Abstract: Aligned large language models(LLMs) remain vulnerable to adversarial manipulation, and their dependence on web-scale pretraining creates a subtle but

ReCast: Recasting Learning Signals for Reinforcement Learning in Generative Recommendation

arXiv:2604.22169v1 Announce Type: cross Abstract: Generic group-based RL assumes that sampled rollout groups are already usable learning signals. We show that this assumption breaks down in sparse-hit

Recognition Without Authorization: LLMs and the Moral Order of Online Advice

arXiv:2604.22143v1 Announce Type: cross Abstract: Large language models are increasingly used to mediate everyday interpersonal dilemmas, yet how their advisory defaults interact with the concentrated

RedVLA: Physical Red Teaming for Vision-Language-Action Models

arXiv:2604.22591v1 Announce Type: new Abstract: The real-world deployment of Vision-Language-Action (VLA) models remains limited by the risk of unpredictable and irreversible physical harm. However, w

Reliable Self-Harm Risk Screening via Adaptive Multi-Agent LLM Systems

arXiv:2604.22154v1 Announce Type: cross Abstract: Emerging AI systems in behavioral health and psychiatry use multi-step or multi-agent LLM pipelines for tasks like assessing self-harm risk and screen

Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Settings

arXiv:2604.22662v1 Announce Type: cross Abstract: Shapley values are a cornerstone of explainable AI, yet their proliferation into competing formulations has created a fragmented landscape with little

@robertskmiles @GaryMarcus “fails less often than a human does” is a terrible metric. We build deterministic software and test it to destruc…

@robertskmiles @GaryMarcus “fails less often than a human does” is a terrible metric. We build deterministic software and test it to destruction to ensure it’s 99.99999 reliable . That’s been the basi

Safety and innovation are not mutually exclusive: for many companies, especially in high-trust industries, AI’s risks are also a hindrance t…

Safety and innovation are not mutually exclusive: for many companies, especially in high-trust industries, AI’s risks are also a hindrance to adoption. In this op-ed in the @FT, I underscore that Euro

safetyyoshua-bengio--x

Sam Altman cannot be trusted. • The OpenAI board fired him because he was not always honest with them. They said he should not control power…

Sam Altman cannot be trusted. • The OpenAI board fired him because he was not always honest with them. They said he should not control powerful AI. • A major report talked to over 100 people and saw s

safetyelon-musk--x