safety

The Uncomfortable Truth About AI “Reasoning” | World Science Festival https://youtu.be/iFYF_e1GSGI?si=pFNjvhp48TMKp-PE via @YouTube @GaryMar…

This video from the World Science Festival, shared by cognitive scientist Gary Marcus, examines the gap between AI systems' claimed reasoning capabilities and their actual mechanisms, likely arguing t

safetygary-marcus--x

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Safety

think you know @garymarcus because read a few of his tweets? try watching this, to see the real deal, and why, for example, the US Senate in…

DGX agent

think you know @garymarcus because read a few of his tweets? try watching this, to see the real deal, and why, for example, the US Senate invited him to testify. Superb conversation between @bgreene a

safetygary-marcus--x

20 May 2026

ThoughtTrace: Understanding User Thoughts in Real-World LLM Interactions

arXiv:2605.20087v1 Announce Type: cross Abstract: Conversational AI has now reached billions of users, yet existing datasets capture only what people say, not what they think. We introduce ThoughtTrac

Toward an AI-Powered Computational Testbed for Workforce Policy

arXiv:2605.19064v1 Announce Type: cross Abstract: Workforce transformations are difficult to forecast and costly to mismanage. In particular, the integration of artificial intelligence into knowledge

Towards Distillation Guarantees under Algorithmic Alignment for Combinatorial Optimization

arXiv:2605.20074v1 Announce Type: new Abstract: Distillation transfers knowledge from a large model trained on broad data to a smaller, more efficient model suitable for deployment. In structured pred

Trustworthy Agent Network: Trust in Agent Networks Must Be Baked In, Not Bolted On

arXiv:2605.19035v1 Announce Type: new Abstract: The rapid advancement of Large Language Models has given rise to autonomous LLM-based agents capable of complex reasoning and execution. As these agents

TSR: Trajectory-Search Rollouts for Multi-Turn RL of LLM Agents

arXiv:2602.11767v3 Announce Type: replace Abstract: Advances in large language models (LLMs) are driving a shift toward using reinforcement learning (RL) to train agents from iterative, multi-turn int

two trillion dollars to build “pathologically dishonest” AI

two trillion dollars to build “pathologically dishonest” AI I think @RyanPGreenblatt's recent post summarizes the vibe of this behavior well. Not the sort of thing that would be acceptable for a human

safetygary-marcus--x

Universal Skeleton Understanding via Differentiable Rendering and MLLMs

arXiv:2603.18003v4 Announce Type: replace Abstract: Multimodal large language models (MLLMs) exhibit strong visual-language reasoning, yet cannot process structured, non-visual data such as human skel

When Critics Disagree: Adaptive Reward Poisoning Attacks in RIS-Aided Wireless Control System

arXiv:2605.20037v1 Announce Type: cross Abstract: Reward-poisoning attacks present a significant risk to learning-based wireless control systems. Given this, we propose a Disagreement-Guided Reward Po

When Preference Labels Fall Short: Aligning Diffusion Models from Real Data

arXiv:2605.19839v1 Announce Type: new Abstract: Preference alignment aims to guide generative models by learning from comparisons between preferred and non-preferred samples. In practice, most existin

When Tabular Foundation Models Meet Strategic Tabular Data: A Prior Alignment Approach

arXiv:2605.19662v1 Announce Type: new Abstract: Tabular foundation models based on pretrained prior-data fitted networks~(PFNs) have shown strong generalization on diverse tabular tasks, but they are

When to Stop Reusing: Dynamic Gradient Gating for Sample-Efficient RLVR

arXiv:2605.19425v1 Announce Type: cross Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for advanced reasoning in Large Language Models (LLMs), but rol

Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making

arXiv:2602.07008v2 Announce Type: replace Abstract: Reliable models should not only predict correctly, but also justify decisions with acceptable evidence. Yet conventional supervised learning typical

Worst-Group Equalized Odds Regularization for Multi-Attribute Fair Medical Image Classification

arXiv:2605.19214v1 Announce Type: cross Abstract: Diagnostic performance in medical AI varies systematically across demographic groups, yet subgroup AUC can mask clinically important disparities. At a

A Fourier perspective on the learning dynamics of neural networks: from sample complexities to mechanistic insights

arXiv:2605.16913v1 Announce Type: cross Abstract: Neural networks trained with gradient-based methods exhibit a strong simplicity bias: they learn simpler statistical features of their data before mov

A Simplex Witness Certificate for Constant Collapse in Variational Autoencoders

arXiv:2605.18224v1 Announce Type: cross Abstract: This note studies exact constant collapse in variational autoencoders, where the encoder mean becomes independent of the input. The goal is to make th

A study done on 361,645 job applications in almost 30 countries over the last 40 years discovered the hiring bias in society is actually aga…

A large-scale meta-analysis examining over 361,000 job applications across nearly 30 countries spanning 40 years found that hiring discrimination based on protected characteristics (such as race, gend

safetyelon-musk--x

A Visual Reinforcement Learning-Based Separate Primitive Policy for Peg-in-Hole Tasks

arXiv:2504.14820v2 Announce Type: replace Abstract: For peg-in-hole tasks, humans rely on binocular visual perception to locate the peg above the hole surface and then proceed with insertion. This pap

safetyarxiv-cs-ro

Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications

arXiv:2605.17413v1 Announce Type: cross Abstract: Safety-aligned language models often refuse cybersecurity requests whose wording resembles misuse, even when the task is authorized and defensive. Thi

Actionable World Representation

arXiv:2605.18743v1 Announce Type: new Abstract: Inspired by the emergent behaviors in large language models that generalized human intelligence, the research community is pursuing similar emergent cap

Activation Steering with a Feedback Controller

arXiv:2510.04309v3 Announce Type: replace Abstract: Controlling the behaviors of large language models (LLM) is fundamental to their safety alignment and reliable deployment. However, existing steerin

Adaptive Control in Autonomous Driving via Real-Time Recurrent RL

arXiv:2602.02236v4 Announce Type: replace-cross Abstract: We study online fine-tuning of pretrained control policies for autonomous driving using Real-Time Recurrent Reinforcement Learning (RTRRL), a

Adaptive Experimentation for Censored Survival Outcomes

arXiv:2605.18459v1 Announce Type: new Abstract: Adaptive experimentation enables efficient estimation of causal effects, but existing methods are not designed for survival data with censoring, where e

Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification

arXiv:2605.17609v1 Announce Type: new Abstract: Many inference-time language-model pipelines combine a cheap reward signal with an expensive verifier, such as exact answer checking in mathematical rea

Adversarial Fragility and Language Vulnerability in Clinical AI: A Systematic Audit of Diagnostic Collapse Under Imperceptible Perturbations and Cross-Lingual Drift in Low-Resource Healthcare Settings

arXiv:2605.16993v1 Announce Type: cross Abstract: Current clinical artificial intelligence (AI) systems are evaluated almost exclusively on clean, standardised, English-language inputs, conditions tha

AffordVLA: Injecting Affordance Representations into Vision-Language-Action Models via Implicit Feature Alignment

arXiv:2605.17517v1 Announce Type: new Abstract: Recent advances in Vision-Language-Action (VLA) models have shown strong potential for general-purpose robotic manipulation. However, the visual represe

safetyarxiv-cs-ro

Agent Bazaar: Enabling Economic Alignment in Multi-Agent Marketplaces

arXiv:2605.17698v1 Announce Type: new Abstract: The deployment of Large Language Models (LLMs) as autonomous economic agents introduces systemic risks that extend beyond individual capability failures

AI Agents May Always Fall for Prompt Injections

arXiv:2605.17634v1 Announce Type: cross Abstract: Prompt injection is the most critical vulnerability in deployed AI agents. Despite recent progress, we show that the prevailing defense paradigm (data

AI Alignment Breaks at the Edge

arXiv:2602.20042v2 Announce Type: replace Abstract: General Alignment has improved average-case helpfulness and safety, but current alignment practice still rewards confident, single-turn responses. T

AI of the People, by the People, for the People: A Social Choice Approach to Collective Control of Artificial Intelligence

arXiv:2605.16291v1 Announce Type: cross Abstract: With the growing adoption of AI systems, reasoning about how society can exert control over AI becomes an increasingly urgent problem. Existing work o

AIM: Adversarial Information Masking for Faithfulness Evaluation of Saliency Maps

arXiv:2605.16905v1 Announce Type: cross Abstract: Post-hoc saliency methods are widely used to interpret deep neural networks, but their faithfulness is difficult to evaluate reliably. Existing evalua

Algorithmic Cultivation: How Social Media Feeds Shape User Language

arXiv:2605.17010v1 Announce Type: cross Abstract: Algorithmic feeds have become primary environments for encountering information online, yet while they shape what people see, less is known about how

Alignment and Safety of Diffusion Models via Reinforcement Learning and Reward Modeling: A Survey

arXiv:2505.17352v2 Announce Type: replace Abstract: Diffusion models have become a central paradigm for image and multimodal generation, yet their deployment raises persistent questions about alignmen

Alignment Drift in Long-Term Human-LLM Interaction: A Mechanism-Oriented Framework

arXiv:2605.16516v1 Announce Type: cross Abstract: Long-term interaction with LLM-based systems may produce alignment drift: a gradual process in which system outputs become less constrained by the use

AMATA: Adaptive Multi-Agent Trajectory Alignment for Knowledge-Intensive Question Answering

arXiv:2605.17352v1 Announce Type: new Abstract: Despite substantial advances in large language models (LLMs), generating factually consistent responses for knowledge-intensive question answering remai

AMR-SD: Asymmetric Meta-Reflective Self-Distillation for Token-Level Credit Assignment

arXiv:2605.18529v1 Announce Type: new Abstract: The alignment of Large Language Models (LLMs) for complex reasoning heavily relies on Reinforcement Learning with Verifiable Rewards (RLVR). However, st

An Assessment of Human vs. Model Uncertainty in Soft-Label Learning and Calibration

arXiv:2605.18648v1 Announce Type: cross Abstract: Central to human-aligned AI is understanding the benefits of human-elicited labels over synthetic alternatives. While human soft-labels improve calibr

An Efficient Streaming Video Understanding Framework with Agentic Control

arXiv:2605.17921v1 Announce Type: new Abstract: Streaming video requires handling dynamic information density under strict latency budgets. Yet, existing methods typically employ static strategies, su

An Empirical Study of Privacy Leakage Chains via Prompt Injection in Black-Box Chatbot Environments

arXiv:2605.18133v1 Announce Type: cross Abstract: LLM-based chatbot agents increasingly process user requests by combining natural-language reasoning with external tools such as web browsing. These ca

AnchorDiff: Topology-Aware Masked Diffusion with Confidence-based Rewriting for Radiology Report Generation

arXiv:2605.17071v1 Announce Type: new Abstract: Radiology report generation (RRG) aims to automatically produce clinically accurate textual reports from medical images. Existing methods predominantly

Anytime and Difficulty-Adaptive PAC-Bayes for Constrained Density-Ratio Network with Continual Learning Guarantees

arXiv:2605.17212v1 Announce Type: new Abstract: A unified framework for learning under covariate shift is presented, in which a constrained density-ratio network approximates the Radon-Nikodym derivat

Are Multimodal LLMs Ready for Surveillance? A Reality Check on Zero-Shot Anomaly Detection in the Wild

arXiv:2603.04727v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) have demonstrated impressive general competence in video understanding, yet their reliability for rea

ARROW: Augmented Replay for RObust World models

arXiv:2603.11395v2 Announce Type: replace-cross Abstract: Continual reinforcement learning challenges agents to acquire new skills while retaining previously learned ones with the goal of improving pe

Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making

arXiv:2605.17228v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support and medical documentation. However, the

As a Soviet historian who has spent years writing about the extreme, repressive control Soviet Communism exercised over its unfortunate citi…

As a Soviet historian who has spent years writing about the extreme, repressive control Soviet Communism exercised over its unfortunate citizens, I find it really hard to bring a similar accusation ag

safetyelon-musk--x

Assessing Localization Technologies for Pedestrian Collision Avoidance

arXiv:2605.18295v1 Announce Type: new Abstract: Robust pedestrian safety is crucial to the next-generation of intelligent transportation systems. Such systems rely on active pedestrian localization an

safetyarxiv-cs-ro

Assured autonomy: How operations research powers and orchestrates generative AI systems

arXiv:2512.23978v2 Announce Type: replace Abstract: Generative artificial intelligence (GenAI) is shifting from conversational assistants toward agentic systems -- autonomous decision-making systems t