safety

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

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

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

DGX agent

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

safetyarxiv-cs-ai

11 May 2026

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

How Value Induction Reshapes LLM Behaviour

arXiv:2605.07925v1 Announce Type: new Abstract: Conversational Large Language Models are post-trained on language that expresses specific behavioural traits, such as curiosity, open-mindedness, and em

safetyarxiv-cs-cl

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

Intention assimilation control for accurate tracking with variable impedance in teleoperation

arXiv:2605.07037v1 Announce Type: new Abstract: Robot systems for teleoperation commonly use a spring-like force pulling the follower robot towards the leader's position to track their movements. With

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

InvThink: Premortem Reasoning for Safer Language Models

arXiv:2510.01569v3 Announce Type: replace Abstract: We present InvThink, a training and prompting framework that requires the model to enumerate, analyze, and constrain potential failures before gener

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

Learned Lyapunov Shielding for Adaptive Control

arXiv:2605.06934v1 Announce Type: new Abstract: We augment the Slotine--Li adaptive controller for Euler--Lagrange systems with three learned components: a structured-quadratic Lyapunov function (V_ps

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

Many-to-Many Multi-Agent Pickup and Delivery

arXiv:2605.07835v1 Announce Type: new Abstract: Multi-robot systems in automated warehouses must manage continuous streams of pickup-and-delivery tasks while ensuring efficiency and safety. Prior work

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

MORPH-U: Multi-Objective Resilient Motion Planning for V2X-Enabled Autonomous Driving in High-Uncertainty Environments via Simulation

arXiv:2605.07370v1 Announce Type: cross Abstract: V2X can warn an autonomous vehicle about hazards beyond line-of-sight, but it also brings uncertainty: messages may be delayed, dropped, or even forge

MPD^2-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis

arXiv:2605.08024v1 Announce Type: new Abstract: Learning-to-defer (L2D) can make glaucoma screening safer by routing difficult/uncertain cases to humans, yet standard formulations overlook expert avai

Multi-environment Invariance Learning with Missing Data

arXiv:2601.07247v2 Announce Type: replace-cross Abstract: Learning models that can handle distribution shifts is a key challenge in domain generalization. Invariance learning, an approach that focuses

Multi-Environment POMDPs with Finite-Horizon Objectives

arXiv:2605.07537v1 Announce Type: new Abstract: Partially Observable Markov Decision Processes (POMDPs) are systems in which one agent interacts with a stochastic environment, and receives only partia

Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation

arXiv:2603.16876v2 Announce Type: replace-cross Abstract: We propose MARL-Rad, a multi-modal multi-agent reinforcement learning framework for radiology report generation that trains the entire agentic

Multi-Objective Multi-Agent Bandits: From Learning Efficiency to Fairness Optimization

arXiv:2605.06864v1 Announce Type: new Abstract: We study multi-objective multi-agent multi-armed bandits (MO-MA-MAB) under stochastic rewards, where agents observe heterogeneous reward vectors and com

Mythos found a single vulnerability in cURL (along with three false positives, and one issue they classified as a bug). The founder/lead dev…

Mythos identified one genuine vulnerability in cURL while also reporting three false positives and one issue classified as a bug during their security analysis. The post references cURL's founder or l

@NameInteger @GaryMarcus @geoffreyhinton Hinton's argument was about encoding: LLMs don't encode text as text, but as weighted matrices. It …

@NameInteger @GaryMarcus @geoffreyhinton Hinton's argument was about encoding: LLMs don't encode text as text, but as weighted matrices. It makes no difference. Memorised data is memorised data no mat

No Forgetting Learning: Buffer-free Continual Learning Classification

arXiv:2503.04638v3 Announce Type: replace Abstract: Most Continual Learning (CL) methods maintain performance on earlier tasks by storing exemplars in a replay buffer, introducing memory overhead that

NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in World Action Models

arXiv:2605.07794v1 Announce Type: new Abstract: World Action Models (WAMs) are an emerging family of policies that tie robot action generation to future-observation modeling. In this work, we focus on

Not even surprised by horrific stories like these anymore. The mission of getting LLMs aligned with human values has largely been a failure.

Not even surprised by horrific stories like these anymore. The mission of getting LLMs aligned with human values has largely been a failure. NEW: ChatGPT advised the FSU shooter that a mass shooting w

OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search

arXiv:2604.03675v2 Announce Type: replace Abstract: Agentic search enables language models to solve knowledge-intensive tasks by adaptively acquiring external evidence over multiple steps. Reinforceme

Object Hallucination-Free Reinforcement Unlearning for Vision-Language Models

arXiv:2605.08031v1 Announce Type: new Abstract: Vision-language models (VLMs) raise growing concerns about privacy, copyright, and bias, motivating machine unlearning to remove sensitive knowledge. Ho

Offline Policy Optimization with Posterior Sampling

arXiv:2605.07393v1 Announce Type: new Abstract: A fundamental challenge in model-based offline reinforcement learning (RL) lies in the trade-off between generalization and robustness against exploitat

oh. my. god. 😱

oh. my. god. 😱 FT Exclusive: NHS England has granted external staff from companies including Palantir “unlimited access” to identifiable patient data while working on a part of its flagship data platf

On the Meta-Design of Allocation Problems

arXiv:2602.08786v4 Announce Type: replace-cross Abstract: There is an extensive literature that studies how to find optimal policies in resource allocation problems, taking the underlying design param