safety

LLM Judges Inconsistently Disagree Across Safety Criteria and Harm Categories

arXiv:2605.31381v1 Announce Type: new Abstract: We evaluate the consistency of automated judges in conducting a multi-dimensional safety evaluation in a reference-free setup. Our results indicate that

Human

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

LVSA: Training-Free Sparse Attention for Long Video Diffusion

DGX agent

arXiv:2605.31057v1 Announce Type: new Abstract: Dense self-attention is the compute and quality bottleneck of long-video diffusion inference: cost grows quadratically with the sequence length, and bey

safetyarxiv-cs-cv

1 Jun 2026

Mental Damage: Caption Poisoning Attacks on Retrieval-Augmented Text-to-Music Generation

arXiv:2605.30365v1 Announce Type: cross Abstract: Retrieval-augmented text-to-music (TTM) systems augment underspecified user prompts using captions retrieved from a music caption dataset. This design

MergeTok: Unified Continuous and Discrete Visual Tokenization via Token Merging

arXiv:2605.30904v1 Announce Type: new Abstract: Most visual tokenizers for image generation are bifurcated into two families with complementary limitations: continuous VAEs offer high-fidelity reconst

Modeling a digital twin of a food supply chain using BigQuery Graph

The example of a growing restaurant Imagine you are running a restaurant chain. You just can't physically feel and touch things to know how your business operates. You need tools and a digital replica

safetygoogle-cloud-ai

Multi-Agent Teams Hold Experts Back

arXiv:2602.01011v4 Announce Type: replace-cross Abstract: Multi-agent LLM systems are increasingly deployed as autonomous collaborators, where agents interact freely rather than execute fixed, pre-spe

Neuron-Level Interventions for Gendered and Gender-Neutral Generation in Language Models

arXiv:2605.30717v1 Announce Type: new Abstract: Language models (LMs) can produce gendered language and stereotypes even when given neutral prompts. Most prior work on gender bias in LMs primarily exa

Neurosymbolic rising!

Neurosymbolic rising! I've wanted to work on deep neurosymbolic integration for a while: if you make an 800k transformer reason *like* a logical solver, you get 100% on sudoku-extreme w. 15m of train

safetygary-marcus--x

OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation

arXiv:2605.30519v1 Announce Type: new Abstract: Autoregressive (AR) video generation extends videos by producing latent chunks sequentially, but scaling to long videos requires repeated access to a gr

On the Illusion of Gender Bias in Face Recognition: Explaining the Fairness Issue Through Non-demographic Attributes

arXiv:2501.12020v2 Announce Type: replace Abstract: Face recognition systems (FRS) exhibit significant accuracy differences based on the user's gender. Since such a gender gap reduces the trustworthin

On the 'Induction Bias' in Sequence Models

arXiv:2602.18333v2 Announce Type: replace-cross Abstract: Despite the remarkable practical success of transformer-based language models, recent work has raised concerns about their ability to perform

On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders

arXiv:2605.31518v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) decompose neural network activations into interpretable features, but many learned features never activate, a problem called

Optimizing Rank for High-Fidelity Implicit Neural Representations

arXiv:2512.14366v2 Announce Type: replace Abstract: Implicit Neural Representations (INRs) based on vanilla Multi-Layer Perceptrons (MLPs) are widely believed to be incapable of representing high-freq

Organizational Adaptation to Generative AI in Cybersecurity

arXiv:2506.12060v2 Announce Type: replace-cross Abstract: Cybersecurity organizations are adapting to GenAI integration through modified frameworks and hybrid operational processes, with success influ

Our views on AI policy and political advocacy

OpenAI outlines its positions on AI policy priorities and explains its approach to political engagement and advocacy efforts. The document likely covers key regulatory areas OpenAI supports, such as s

safetyopenai

PAC-Bayesian Reinforcement Learning Trains Generalizable Policies

arXiv:2510.10544v3 Announce Type: replace-cross Abstract: We derive a novel PAC-Bayesian generalization bound for reinforcement learning that explicitly accounts for Markov dependencies in the data, t

Parallel Tempering Initial Sampling in Inference-Time Reward Alignment

arXiv:2605.30991v1 Announce Type: cross Abstract: Inference-time reward alignment steers pretrained diffusion and flow-based generative models to satisfy user-specified rewards without retraining. Rec

PASTA: A Scalable Framework for Multi-Policy AI Compliance Evaluation

arXiv:2601.11702v3 Announce Type: replace-cross Abstract: AI compliance is becoming increasingly critical as AI systems grow more powerful and pervasive. Yet the rapid expansion of AI policies creates

*-PLUIE: Personalisable metric with Llm Used for Improved Evaluation

arXiv:2602.15778v2 Announce Type: replace Abstract: Evaluating the quality of automatically generated text often relies on LLM-as-a-judge (LLM-judge) methods. While effective, these approaches are com

Population-Free Pareto Tracking for Sample-Efficient Multi-Policy MORL

arXiv:2508.02217v2 Announce Type: replace Abstract: Multi-objective reinforcement learning (MORL) is a fundamental framework for real-world decision-making problems involving multiple conflicting crit

PostCam: Camera-Controllable Novel-View Video Generation with Query-Shared Cross-Attention

arXiv:2511.17185v2 Announce Type: replace Abstract: We propose PostCam, a streamlined framework for novel-view video generation that achieves superior detail preservation and precise camera trajectory

Prediction: Nobody knows when this will all collapse, but 2026 will be remembered in hindsight as the year in which retail investors and ind…

Prediction: Nobody knows when this will all collapse, but 2026 will be remembered in hindsight as the year in which retail investors and index funds were left holding the bag. It's funny how people th

safetygary-marcus--x

Preference-Aware Rubric Learning for Personalized Evaluation

arXiv:2605.31545v1 Announce Type: new Abstract: As Large Language Models (LLMs) evolve from general-purpose assistants to user-centric agents, personalization has become central to aligning model beha

PReMISE: Policy Rubrics as Measurement Specifications for LLM Judges

arXiv:2605.30803v1 Announce Type: new Abstract: LLM judges are increasingly used to evaluate open-ended responses, but their scores depend strongly on the rubrics that condition them. A vague rubric a

Primitive Subspaces Mediate Few-Shot Transfer in VLAs

arXiv:2605.30695v1 Announce Type: new Abstract: Deploying vision-language-action (VLA) policies in industrial environments requires the ability to teach new tasks at low cost, a property current VLAs

safetyarxiv-cs-ro

Rationalize: Shared Semantic Reasoning for Human-AI Alignment

arXiv:2605.30632v1 Announce Type: cross Abstract: We introduce Rationalize, a role-pair framework for shared semantic reasoning between humans and AI models in data-driven sensemaking. Building on ide

RDGen: Demonstration Generation for High-Quality Robot Learning via Reinforcement Learning

arXiv:2605.30957v1 Announce Type: new Abstract: Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robot control. However, their performance remains fundament

safetyarxiv-cs-ro

Reading Between the Citations: A Typed Claim Network for Scientific Literature

arXiv:2605.30966v1 Announce Type: cross Abstract: Knowledge graphs over corpora of inter-referencing documents - scholarly papers, legal opinions, policy briefs - encode the topology of reference but

REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge

arXiv:2603.17145v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly deployed as automated evaluators that assign numeric scores to model outputs, a paradigm known a

Reassessing Extractive QA Datasets at Scale: LLM-as-a-Judge and In-Depth Analyses

arXiv:2504.11972v3 Announce Type: replace Abstract: Extractive QA tasks are commonly evaluated using Exact Match (EM) and F1-score, but these metrics often fail to reflect true model performance. Rece

Reinforced sequential Monte Carlo for amortised sampling

arXiv:2510.11711v2 Announce Type: replace Abstract: This paper proposes a synergy of amortised and particle-based methods for sampling from distributions defined by unnormalised density functions. We

Reinforcement Learning Amplifies Emergent Misalignment from Harmless Rewards

arXiv:2605.31328v1 Announce Type: new Abstract: Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples.

Reinterpreting Safety Thresholds as Neuron Spiking Thresholds

arXiv:2605.30368v1 Announce Type: cross Abstract: Surrogate Safety Measures (SSMs) are extensively utilised in the evaluation of traffic risk in automated driving contexts. However, the majority of SS

Representation Collapse in Sequential Post-Training of Large Language Models

arXiv:2605.30524v1 Announce Type: new Abstract: Large language models are now adapted through chains of post-training stages rather than through a single instruction-tuning pass. This paper studies wh

Rethinking Multimodal Few-Shot 3D Point Cloud Segmentation: From Fused Refinement to Decoupled Arbitration

arXiv:2601.01456v2 Announce Type: replace-cross Abstract: In this paper, we revisit multimodal few-shot 3D point cloud semantic segmentation (FS-PCS), identifying a conflict in 'Fuse-then-Refine' para

// Reusable Context Engineering // Context bloat quietly kills long-horizon runs, but you can fix it from the outside without fine-tuning th…

// Reusable Context Engineering // Context bloat quietly kills long-horizon runs, but you can fix it from the outside without fine-tuning the underlying agent. (bookmark this) Context management is us

safetydair-ai--x

Revisiting Zeroth-Order Hessian Approximation: A Single-Step Policy Optimization Lens

arXiv:2605.30960v1 Announce Type: new Abstract: Accurate Zeroth-Order (ZO) Hessian estimation is a cornerstone of derivative-free methods, essential for tasks such as bilevel optimization, Bayesian in

Routing on the Stiefel Manifold: When Does Adaptive Subspace Selection Help for Cross-Domain EEG Decoding?

arXiv:2605.31043v1 Announce Type: cross Abstract: Cross-domain EEG decoding remains challenging despite advances in Riemannian deep learning: covariance matrices from different subjects occupy systema

Safeguarding Text-to-Image Generation via Inference-Time Prompt-Noise Optimization

arXiv:2412.03876v2 Announce Type: replace Abstract: Text-to-Image (T2I) diffusion models are widely recognized for their ability to generate high-quality and diverse images based on text prompts. Howe

Scalable Constrained Multi-Agent Reinforcement Learning via State Augmentation and Consensus for Separable Dynamics

arXiv:2605.30461v1 Announce Type: cross Abstract: We present a distributed approach for constrained Multi-Agent Reinforcement Learning (MARL) that combines state-augmented policy learning with distrib

Scaling Multi-Agent Environment Co-Design with Diffusion Models

arXiv:2511.03100v2 Announce Type: replace-cross Abstract: The agent-environment co-design paradigm jointly optimises agent policies and environment configurations in search of improved system performa

SCOPE: Selective Conformal Optimized Pairwise LLM Judging

arXiv:2602.13110v3 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly used as scalable judges in pairwise evaluation, but they remain prone to miscalibration and bias

SDM-Q: Cost-Aware Staged Decision-Making for Multi-Omics Classification with Deep Q-Learning

arXiv:2605.31014v1 Announce Type: new Abstract: Multi-omics data provide complementary molecular characterizations of disease phenotypes and play an important role in disease diagnosis and subtype cla

Secure AI agents with Policy and Lambda interceptors in Amazon Bedrock AgentCore gateway

In this post, we use a lakehouse data agent to demonstrate how you can use Policy for deterministic access control and Lambda interceptors for dynamic validation. We then show how to combine Lambda in

safetyaws-ml-blog

Seeing Before Agreeing: Aligning Multi-Agent Consensus with Visual Evidence

arXiv:2605.30698v1 Announce Type: cross Abstract: Vision-language models (VLMs) have achieved strong performance on visual question answering (VQA). To mitigate individual hallucinations and blind spo

Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures

arXiv:2605.30608v1 Announce Type: new Abstract: Learning a shared representation between spoken text and gesture is central to co-speech gesture retrieval, synthesis, and understanding, but remains ch

SemStruct: Contextualizing Semantic Embeddings with Structural Information for Schema Matching

arXiv:2605.30729v1 Announce Type: new Abstract: Schema matching is a fundamental step in integrating heterogeneous data sources. While Pre-trained Language Models (PLMs) have revolutionized this task

Simulation of collision avoidance behavior in crowd movement by data-driven approach

arXiv:2605.31210v1 Announce Type: cross Abstract: Crowd movement simulation is essential for pedestrian safety management and facility layout optimization. Data-driven models enhance trajectory predic