Search: “safety”

Do agents listen to you… or themselves? While evaling subagent behavior in deep agent systems, we noticed an interesting quirk in our agents…

Do agents listen to you… or themselves? While evaling subagent behavior in deep agent systems, we noticed an interesting quirk in our agents' alignment with hand-written system prompts vs. the instruc

safetyharrison-chase--x

Safety

Enhancing Speech Large Language Models through Reinforced Behavior Alignment

DGX agent

arXiv:2509.03526v2 Announce Type: replace Abstract: The recent advancements of Large Language Models (LLMs) have spurred considerable research interest in extending their linguistic capabilities beyon

safetyarxiv-cs-cl

21 May 2026

EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models

arXiv:2510.05942v3 Announce Type: replace Abstract: We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a mo

safetyarxiv-cs-cl

extit{Stochastic} MeanFlow Policies: One-Step Generative Control with Entropic Mirror Descent

arXiv:2605.21282v1 Announce Type: new Abstract: Online off-policy reinforcement learning (RL) is shaped by two coupled choices: the policy class and the update rule. Gaussian policies are fast and hav

FEAT: A Linear-Complexity Foundation Model for Extremely Large Structured Data

arXiv:2603.16513v3 Announce Type: replace Abstract: Structured data is widely used in domains such as healthcare, finance, and scientific data management. Recent studies on structured data foundation

FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching

arXiv:2605.20910v1 Announce Type: new Abstract: Extending the generation horizon of video diffusion models to long sequences remains a long-standing and important challenge. Existing training-free app

Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification

arXiv:2605.20248v1 Announce Type: new Abstract: In the transductive setting, where the full graph is observed but node labels are only partially available, progress in semi-supervised node classificat

GROW: Aligning GRPO with State-Action Modeling for Open-World VLM Agents

arXiv:2605.20246v1 Announce Type: new Abstract: Recently, vision-language model (VLM) agents have shown promising progress in open-world tasks, where successful task completion often requires multiple

Here's how the corruption works: Thursday: RJ Reynolds donates $5M to Trump Saturday: Trump invites RJR execs to Mar a Lago; execs ask to lo…

Here's how the corruption works: Thursday: RJ Reynolds donates $5M to Trump Saturday: Trump invites RJR execs to Mar a Lago; execs ask to loosen regs on flavored vapes; Trump calls up RFK Jr. and tell

safetyyann-lecun--x

HORST: Composing Optimizer Geometries for Sparse Transformer Training

arXiv:2605.21104v1 Announce Type: new Abstract: Sparsifying transformers remains a fundamental challenge, as standard optimizers fail to simultaneously encourage sparsity and maintain training stabili

I don't think anyone has a good intuitive sense about what this means, and that failure of imagination is a generally bad thing for planning…

I don't think anyone has a good intuitive sense about what this means, and that failure of imagination is a generally bad thing for planning, investment, and policy. I also don't have an easy solution

safetyethan-mollick--x

I was happy to join the @ScienceBoard_UN podcast to discuss deception among frontier AI models and the need to manage its potential global i…

I was happy to join the @ScienceBoard_UN podcast to discuss deception among frontier AI models and the need to manage its potential global impacts and risks. 🎙️What happens when AI learns to lie? Scie

safetyyoshua-bengio--x

if you own a tesla, you are the product

if you own a tesla, you are the product Astonishing potential, especially if one takes all the galaxies into account and factors them into the calculation. 😃 And this comes from a man who removed Auto

Improved convergence rate of kNN graph Laplacians: differentiable self-tuned affinity

arXiv:2410.23212v2 Announce Type: replace-cross Abstract: In graph-based data analysis, k-nearest neighbor (kNN) graphs are widely used due to their adaptivity to local data densities. Allowing weight

Inference Time Policy Optimization for Offline RL with Differentiable World Models

arXiv:2603.22430v2 Announce Type: replace Abstract: Offline Reinforcement Learning (RL) learns optimal policies from fixed datasets, training a policy once and deploying it at inference time without f

It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs

arXiv:2605.20258v1 Announce Type: new Abstract: Contextual Integrity (CI) defines privacy not merely as keeping information hidden, but as governing information flows according to the norms of a given

It’s crazy that OpenAI and xAI were built partly as ways to mitigate @demishassabis’s power and how both are more reckless and less focused …

It’s crazy that OpenAI and xAI were built partly as ways to mitigate @demishassabis’s power and how both are more reckless and less focused on the good part, viz AI for science. And then Anthropic was

Latent Geometry as a Structural Monitor: Eigenspace Alignment for Anomaly Detection in Anonymity Networks

arXiv:2605.20391v1 Announce Type: cross Abstract: Traditional anomaly detection marks events when measured signals cross predefined thresholds. This captures the moment of transition but not the struc

Learning Robust Dexterous In-Hand Manipulation from Joint Sensors with Proprioceptive Transformer

arXiv:2605.21330v1 Announce Type: new Abstract: In-hand object manipulation is a fundamental yet challenging capability for dexterous robots. Despite significant progress in dexterous manipulation, ex

safetyarxiv-cs-ro

Learning to Think in Physics: Breaking Shortcut Learning in Scientific Diffusion via Representation Alignment

arXiv:2605.20780v1 Announce Type: cross Abstract: Physics-informed diffusion models typically enforce PDE constraints only on final outputs, leaving intermediate representations unconstrained and pron

Letting Trajectories Spread: Quality-Preserving Control for Diverse Flow Matching

arXiv:2510.09060v2 Announce Type: replace-cross Abstract: Flow-based text-to-image models follow deterministic trajectories, making it costly to explore diverse modes under limited sampling budgets. E

Linear-DPO: Linear Direct Preference Optimization for Diffusion and Flow-Matching Generative Models

arXiv:2605.21123v1 Announce Type: new Abstract: Direct Preference Optimization (DPO) is successful for alignment in LLMs but still faces challenges in text-to-image generation. Existing studies are co

Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex

arXiv:2605.06139v2 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large language models (LLMs) post-training to incentivize r

LLM Pretraining Shapes a Generalizable Manifold: Insights into Cross-Modal Transfer to Time Series

arXiv:2605.20449v1 Announce Type: new Abstract: Can language-pretrained transformers become effective time-series forecasters, and why? In this paper, we show that cross-modal transfer arises because

longer comment on yesterday’s stories

longer comment on yesterday’s stories the nuance between yesterday’s stories that X influencers won’t tell you: https://open.substack.com/pub/garymarcus/p/checking-the-math-behind-openai-and?r=8tdk6&u

Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX

arXiv:2605.20577v1 Announce Type: cross Abstract: Riichi Mahjong is a multi-player, imperfect-information game characterized by stochasticity and high-dimensional state spaces. These attributes presen

Mind the Sim-to-Real Gap & Think Like a Scientist

arXiv:2605.21458v1 Announce Type: cross Abstract: Suppose a planner has a pre-trained simulator of a sequential decision problem and the option to run real experiments in the field. The simulator is c

Mitigating Label Bias with Interpretable Rubric Embeddings

arXiv:2605.21455v1 Announce Type: new Abstract: Statistical decision algorithms are increasingly deployed in domains where ground-truth labels are hard to obtain, such as hiring, university admissions

Mobile UMI: Cross-View Diffusion Policy with Decoupled Kinematics for Mobile Manipulation

arXiv:2605.20894v1 Announce Type: new Abstract: Mobile imitation learning on portable demonstration interfaces faces two coupled bottlenecks: locomotion-contaminated action labels and inference-induce

safetyarxiv-cs-ro

Multi-Head Attention as Ensemble Nadaraya-Watson Estimation: Variance Reduction, Decorrelation, and Optimal Head Diversity

arXiv:2605.20271v1 Announce Type: cross Abstract: We develop a rigorous statistical theory of multi-head attention (MHA) as an ensemble of Nadaraya-Watson (NW) kernel regression estimators. Building o

Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards

arXiv:2605.20865v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) plays a pivotal role in improving the reasoning ability of large language models. However, widely

Multimodal LLMs under Pairwise Modalities

arXiv:2605.21059v1 Announce Type: new Abstract: Despite the impressive results achieved by multimodal large language models (MLLMs), their training typically relies on jointly curated multimodal data,

NaP-Control: Navigating Diffusion Prior for Versatile and Fast Character Control

arXiv:2605.20209v1 Announce Type: cross Abstract: Achieving precise, versatile whole-body character control in physics-based animation remains challenging. Recent diffusion-based policies generate ric

Neural Collapse by Design: Learning Class Prototypes on the Hypersphere

arXiv:2605.20302v1 Announce Type: cross Abstract: Supervised classification has a theoretical optimum, Neural Collapse (NC), yet neither of its two dominant paradigms reaches it in practice. Cross ent

nice measure of how insane SpaceX’s claims are

nice measure of how insane SpaceX’s claims are I don't offer investment advice, but when I see that SpaceX claims that their total addressable market is the size of the *complete* US economy (with eve

nor do we know how the (new) model works nor how it does on anything else nor how it was trained. scientists wait for facts; cheerleaders (o…

nor do we know how the (new) model works nor how it does on anything else nor how it was trained. scientists wait for facts; cheerleaders (over and over) rush to judgments that have often been wrong.

note that @jeffBezos is not betting on chatbots, but rather focusing on domain-specific tools for engineering, which is also something I wou…

note that @jeffBezos is not betting on chatbots, but rather focusing on domain-specific tools for engineering, which is also something I would encourage. the financial story for what he is doing may b

OcclusionFormer: Arranging Z-Order for Layout-Grounded Image Generation

arXiv:2605.21343v1 Announce Type: new Abstract: Recent layout-to-image models have achieved remarkable progress in spatial controllability. However, they still struggle with inter-object occlusion. Wh

👀OpenAI Q1 operating margin was -122% even when excluding SBC and other items.

OpenAI reported a negative 122% operating margin in Q1, indicating the company spent significantly more on operations than it generated in revenue, even after excluding stock-based compensation and ot

our company, which lost money last quarter, is going to take a big slice of $28.5 trillion market that we project will exist, primarily in a…

our company, which lost money last quarter, is going to take a big slice of 28.5 trillion market that we project will exist, primarily in a field in which we have almost no current presence, which is

Parallel LLM Reasoning for Bias-Resilient, Robust Conceptual Abstraction

arXiv:2605.20194v1 Announce Type: new Abstract: Large language models (LLMs) have been increasingly used to analyze text. However, they are often plagued with contextual reasoning limitations when ana

safetyarxiv-cs-cl

Pareto-Enhanced Portrait Generation: Vision-Aligned Text Supervision for Alignment, Realism, and Aesthetics

arXiv:2605.20640v1 Announce Type: new Abstract: Text-to-image diffusion models often face a severe trilemma in human portrait generation: text-image alignment, photorealism, and human-perceived aesthe

PointACT: Vision-Language-Action Models with Multi-Scale Point-Action Interaction

arXiv:2605.21414v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models have shown strong potential for general-purpose robotic manipulation by leveraging large pretrained vision-languag

Pretty profound for an ipo filing: “For the entirety of its existence, human civilization has lived on a single celestial body: Earth. The c…

Pretty profound for an ipo filing: “For the entirety of its existence, human civilization has lived on a single celestial body: Earth. The current paradigm, in which human civilization is confined to

safetyelon-musk--x

Principled RL for Flow Matching Emerges from the Chunk-level Policy Optimization

arXiv:2510.21583v2 Announce Type: replace Abstract: Recent Progress in post-training flow matching for text-to-image (T2I) generation with Group Relative Policy Optimization (GRPO) has demonstrated st

Q-SpiRL: Quantum Spiking Reinforcement Learning for Adaptive Robot Navigation

arXiv:2605.20801v1 Announce Type: new Abstract: Adaptive robot navigation in dynamic environments requires policies that can reach the target reliably while producing efficient and stable trajectories

safetyarxiv-cs-ro

Quantum End-to-End Learning for Contextual Combinatorial Optimization

arXiv:2605.20222v1 Announce Type: cross Abstract: Contextual combinatorial optimization (CCO) plays a critical role in decision-making under uncertainty, yet remains a significant challenge. We presen

Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis

arXiv:2605.20277v1 Announce Type: new Abstract: Medical vision-language models (VLMs) have rapidly advanced as general-purpose multimodal assistants, yet their deployment in 3D Computed Tomography (CT