Search: “safety”

EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle

arXiv:2510.16079v2 Announce Type: replace-cross Abstract: Current Large Language Model (LLM) agents show strong performance in tool use, but lack the crucial capability to systematically learn from th

Safety

Ex-OpenAI co-founder Ilya Sutskever claims to have collected proof of Sam Altman’s alleged dishonesty https://www.theglobeandmail.com/busine…

DGX agent

Ex-OpenAI co-founder Ilya Sutskever claims to have collected proof of Sam Altman’s alleged dishonesty https://www.theglobeandmail.com/business/article-openai-trial-sutskever-sam-altman-musk/?utm_sourc

safetygary-marcus--x

11 May 2026

Excellent explainer video by @FryRsquared on the risks of AI agents. She also raises a crucial point: we shouldn’t make the mistake of think…

Excellent explainer video by @FryRsquared on the risks of AI agents. She also raises a crucial point: we shouldn’t make the mistake of thinking current limitations will necessarily persist. As we’ve s

safetyyoshua-bengio--x

Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models

arXiv:2605.07244v1 Announce Type: cross Abstract: We introduce Mutual Reinforcement Learning, a framework for concurrent RL post-training in which heterogeneous LLM policies exchange typed experience

f-Divergence Regularized RLHF: Two Tales of Sampling and Unified Analyses

arXiv:2605.06977v1 Announce Type: cross Abstract: Reinforcement Learning from Human Feedback (RLHF) has become a cornerstone technique for post-training large language models. While most existing appr

Flatness and Gradient Alignment Are Both Necessary: Spectral-Aware Gradient-Aligned Exploration for Multi-Distribution Learning

arXiv:2605.07914v1 Announce Type: cross Abstract: Sharpness-aware and gradient-alignment methods have been shown to improve generalization, however each family of methods targets a single geometric pr

Flexible Entropy Control in RLVR with a Gradient-Preserving Perspective

arXiv:2602.09782v2 Announce Type: replace-cross Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a critical method for enhancing the reasoning capabilities of Large Langu

Flow-OPD: On-Policy Distillation for Flow Matching Models

arXiv:2605.08063v1 Announce Type: cross Abstract: Existing Flow Matching (FM) text-to-image models suffer from two critical bottlenecks under multi-task alignment: the reward sparsity induced by scala

From 0-Order Selection to 2-Order Judgment: Combinatorial Hardening Exposes Compositional Failures in Frontier LLMs

arXiv:2605.07268v1 Announce Type: new Abstract: Multiple-choice reasoning benchmarks face dual challenges: rapid saturation from advancing models and data contamination that undermines static evaluati

safetyarxiv-cs-cl

From Model to Data (M2D): Shifting Complexity from GNNs to Graphs for Transparent Graph Learning

arXiv:2605.06814v1 Announce Type: new Abstract: Graph Neural Networks (GNNs) achieve high performance but can be opaque to humans, making it difficult to understand and compare the many proposed archi

From Specification to Deployment: Empirical Evidence from a W3C VC + DID Trust Infrastructure for Autonomous Agents

arXiv:2605.06738v1 Announce Type: cross Abstract: Autonomous AI agents now transact at production scale -- 69,000 bots executing 165 million transactions across 50 million USDC in cumulative volume on

From Time Series Analysis to Question Answering: A Survey in the LLM Era

arXiv:2506.11512v2 Announce Type: replace-cross Abstract: Recently, Large Language Models (LLMs) have introduced a novel paradigm in Time Series Analysis (TSA), leveraging strong language capabilities

GASim: A Graph-Accelerated Hybrid Framework for Social Simulation

arXiv:2605.07692v1 Announce Type: new Abstract: Large-scale social simulators are essential for studying complex social patterns. Prior work explores hybrid methods to scale up simulations, combining

Generalised Linear Models in Deep Bayesian RL with Learnable Basis Functions

arXiv:2512.20974v3 Announce Type: replace-cross Abstract: Bayesian Reinforcement Learning (BRL), a subclass of Meta-Reinforcement Learning (Meta-RL), provides a principled framework for generalisation

🤔Going to be in Vancouver tomorrow? 🧐Curious about whether or not scaling AI is a bad bet? 😎Catch TPN founder @zacharykarabell's chat wit…

🤔Going to be in Vancouver tomorrow? 🧐Curious about whether or not scaling AI is a bad bet? 😎Catch TPN founder @zacharykarabell's chat with @GaryMarcus at @WebSummit at 10:20am. 🎟️Get your tickets here

Gonna keep retweeting, because this flawed move never dies:

Gonna keep retweeting, because this flawed move never dies: AI hype has become a giant game of bait and switch. the bait: we are going to make an AI that can solve any problem a expert human could sol

GRAPE: Let GRPO Supervise Query Rewriting by Ranking for Retrieval

arXiv:2509.23370v2 Announce Type: replace Abstract: The CLIP model has established itself as a cornerstone of large-scale retrieval systems. However, its performance often degrades under distributiona

Guidance Is Not a Hyperparameter: Learning Dynamic Control in Diffusion Language Models

arXiv:2605.07701v1 Announce Type: new Abstract: Classifier-Free Guidance (CFG) is a widely used mechanism for controlling diffusion-based generative models, yet its guidance scale is typically treated

safetyarxiv-cs-cl

GustPilot: A Hierarchical DRL-INDI Framework for Wind-Resilient Quadrotor Navigation

arXiv:2603.19966v2 Announce Type: replace Abstract: Wind disturbances remain a key barrier to reliable autonomous navigation for lightweight quadrotors, where the rapidly varying airflow can destabili

safetyarxiv-cs-ro

HAIC: Humanoid Agile Object Interaction Control via Dynamics-Aware World Model

arXiv:2602.11758v2 Announce Type: replace Abstract: Humanoid robots show promise for complex whole-body tasks in unstructured environments. Although Human-Object Interaction (HOI) has advanced, most m

safetyarxiv-cs-ro

HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models

arXiv:2605.07973v1 Announce Type: new Abstract: Text-to-image diffusion models can generate visually stunning images, yet, controlling what appears and how it appears, remains surprisingly difficult,

How Log-Barrier Helps Exploration in Policy Optimization

arXiv:2603.15001v2 Announce Type: replace-cross Abstract: Recently, it has been shown that the Stochastic Gradient Bandit (SGB) algorithm converges to a globally optimal policy with a constant learnin

How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment

arXiv:2605.06850v1 Announce Type: cross Abstract: Reinforcement Learning (RL) has emerged as a crucial paradigm for unlocking the advanced reasoning capabilities of Large Language Models (LLMs), encom

How to utilize failure demo data?: Effective data selection for imitation learning using distribution differences in attention mechanism

arXiv:2605.07560v1 Announce Type: new Abstract: Imitation learning for robotic tasks has relied primarily on policies trained only on successful demonstrations, although failures are unavoidable durin

safetyarxiv-cs-ro

If you believe this nonsense, and a lot of people do, I beg you to read my newsletter called “Misplaced Panic Over AI progress” 🙏

If you believe this nonsense, and a lot of people do, I beg you to read my newsletter called “Misplaced Panic Over AI progress” 🙏 we are exactly 4.5 steps away from achieving AGI!! People are not read

Implicit Compression Regularization: Concise Reasoning via Internal Shorter Distributions in RL Post-Training

arXiv:2605.07316v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards improves LLM reasoning but often induces overthinking, where models generate unnecessarily long reasoning

Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer

This newsletter covers three main topics: the relationship between AI capabilities relative to human intelligence (RSI) and its potential economic impacts, regulatory approaches that emphasize flexibi

safetyimport-ai

Important Nature Neuroscience paper shows how humans differ from LLMs. Many people currently believe that humans are just next-word predicto…

Important Nature Neuroscience paper shows how humans differ from LLMs. Many people currently believe that humans are just next-word predictors, like LLMs. But this new paper by Zou, Poeppel and Ding s

🚨 In his testimony just now, at the Musk-OpenAI trial, Satya Nadella came off as shrewd, calm, and (mostly) honest, an impressive leader – …

🚨 In his testimony just now, at the Musk-OpenAI trial, Satya Nadella came off as shrewd, calm, and (mostly) honest, an impressive leader – but also selectively blind. How? He seemed unable to believe

Inference-Time Attribute Distribution Alignment for Unconditional Diffusion

arXiv:2605.07456v1 Announce Type: new Abstract: Inference-time controllable generation is essential for real-world applications of unconditional diffusion models. However, most existing techniques foc

InfoGeo: Information-Theoretic Object-Centric Learning for Cross-View Generalizable UAV Geo-Localization

arXiv:2605.07099v1 Announce Type: new Abstract: Cross-view geo-localization (CVGL) is fundamental for precise localization and navigation in GPS-denied environments, aiming to match ground or UAV imag

InterCoG: Towards Spatially Precise Image Editing with Interleaved Chain-of-Grounding Reasoning

arXiv:2603.01586v3 Announce Type: replace Abstract: Emerging unified editing models have demonstrated strong capabilities in general object editing tasks. However, it remains a significant challenge t

Inverse Reinforcement Learning with Just Classification and a Few Regressions

arXiv:2509.21172v2 Announce Type: replace Abstract: Inverse reinforcement learning (IRL) aims to infer rewards from observed behavior, but rewards are not identified from the policy alone: many reward

Is the Future Compatible? Diagnosing Dynamic Consistency in World Action Models

arXiv:2605.07514v1 Announce Type: cross Abstract: World Action Models (WAMs) enable decision-making through imagined rollouts by predicting future observations and actions. However, the reliability of

It is insane that we allow big tech to race toward superintelligence without any oversight. This is a grave national security risk. Industry…

It is insane that we allow big tech to race toward superintelligence without any oversight. This is a grave national security risk. Industry will not regulate itself. Get trained with Torchbearer to a

safetyconnor-leahy--x

KL for a KL: On-Policy Distillation with Control Variate Baseline

arXiv:2605.07865v1 Announce Type: cross Abstract: On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for large language models, especially for reasoning domains. However, OP

Learning Cross-Atlas Consistent Brain Disorder Representations via Disentangled Multi-Atlas Functional Connectivity Learning

arXiv:2605.07026v1 Announce Type: cross Abstract: Functional connectivity (FC) derived from resting-state fMRI is widely used to characterize large-scale brain network alterations in neurological and

Learning to Track Instance from Single Nature Language Description

arXiv:2605.07064v1 Announce Type: new Abstract: How to achieve vision-language (VL) tracking using natural language descriptions from a video sequence extbf{without relying on any bounding-box ground

Learning Visual Feature-Based World Models via Residual Latent Action

arXiv:2605.07079v1 Announce Type: cross Abstract: World models predict future transitions from observations and actions. Existing works predominantly focus on image generation only. Visual feature-bas

Lightweight Unpaired Smartphone ISP Transfer with Semantic Pseudo-Pairing

arXiv:2605.07495v1 Announce Type: new Abstract: Unpaired smartphone ISP is a challenging problem due to the lack of scene and color alignment between RAW and target RGB images. Many existing methods e

MAGIQ: A Post-Quantum Multi-Agentic AI Governance System with Provable Security

arXiv:2605.06933v1 Announce Type: new Abstract: Our computing ecosystem is being transformed by two emerging paradigms: the increased deployment of agentic AI systems and advancements in quantum compu

MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge

arXiv:2507.21183v5 Announce Type: replace-cross Abstract: As the era of large language models (LLMs) unfolds, Preference Optimization (PO) methods have become a central approach to aligning LLMs with

Masks Can Talk: Extracting Structured Text Information from Single-Modal Images for Remote Sensing Change Detection

arXiv:2605.07178v1 Announce Type: new Abstract: Remote sensing change detection is pivotal for urban monitoring, disaster assessment, and environmental resource management. Yet, unimodal deep learning

Michael Burry urged investors to scale back exposure to surging technology stocks, saying the current market environment has reached histori…

Michael Burry urged investors to scale back exposure to surging technology stocks, saying the current market environment has reached historically dangerous extremes reminiscent of prior speculative bu

Mind the Gap: Geometrically Accurate Generative Reconstruction from Disjoint Views

arXiv:2605.07550v1 Announce Type: new Abstract: 3D vision systems are fundamentally constrained by their reliance on visual overlap: reconstruction methods require it for geometric alignment, while ge

Miner:Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Models

arXiv:2601.04731v2 Announce Type: replace Abstract: Current critic-free RL methods for large reasoning models suffer from severe inefficiency when training on positive homogeneous prompts (where all r

MoCoTalk: Multi-Conditional Diffusion with Adaptive Router for Controllable Talking Head Generation

arXiv:2605.08050v1 Announce Type: new Abstract: Talking-head generation requires joint modeling of identity, head pose, facial expression, and mouth dynamics. Existing methods typically address only a

Modality Gap-Driven Subspace Alignment Training Paradigm For Multimodal Large Language Models

arXiv:2602.07026v2 Announce Type: replace-cross Abstract: Despite the success of multimodal contrastive learning in aligning visual and linguistic representations, a persistent geometric anomaly, the