Search: “safety”

What Kind of Language is Easy to Language-Model Under Curriculum Learning?

arXiv:2604.26844v1 Announce Type: new Abstract: Many of the thousands of attested languages share common configurations of features, creating a spectrum from typologically very rare (e.g., object-verb

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Safety

When Annotators Disagree, Topology Explains: Mapper, a Topological Tool for Exploring Text Embedding Geometry and Ambiguity

DGX agent

arXiv:2510.17548v2 Announce Type: replace Abstract: Language models are often evaluated with scalar metrics like accuracy, but such measures fail to capture how models internally represent ambiguity,

safetyarxiv-cs-cl

30 Apr 2026

// When to Retrieve During Reasoning // Pay attention to this one, AI devs. (bookmark it) Most RAG systems retrieve once, before the model s…

// When to Retrieve During Reasoning // Pay attention to this one, AI devs. (bookmark it) Most RAG systems retrieve once, before the model starts reasoning. Large reasoning models like o1 and R1 don't

safetydair-ai--x

Who is accountable? the latest from @marketoonist

This post likely discusses accountability in AI systems, featuring commentary or a cartoon from Marketoonist (a popular cartoonist who creates comics about business and technology). Gary Marcus, an AI

wow. incredible.

wow. incredible. A Chinese court has ruled it illegal to replace human workers with AI purely for the sake of cost-cutting. The court decided that companies hold a social responsibility to treat worke

Zuck is zucking up Meta

Zuck is zucking up Meta Mark Zuckerberg lacks the cloud-computing business to truly justify Meta's AI spending, @DaveLeeBBG says (via @opinion) https://www.bloomberg.com/opinion/articles/2026-04-30/me

A Blueprint for AI-Driven Software Quality: Integrating LLMs with Established Standards

arXiv:2505.13766v5 Announce Type: replace-cross Abstract: Software Quality Assurance (SQA) is critical for delivering reliable, secure, and efficient software products. The Software Quality Assurance

A Deep Reinforcement Learning Approach to Automated Stock Trading, using xLSTM Networks

arXiv:2503.09655v2 Announce Type: replace-cross Abstract: Traditional Long Short-Term Memory (LSTM) networks are effective for handling sequential data but have limitations such as gradient vanishing

A Systematic Post-Train Framework for Video Generation

arXiv:2604.25427v1 Announce Type: new Abstract: While large-scale video diffusion models have demonstrated impressive capabilities in generating high-resolution and semantically rich content, a signif

agents are going and are already changing every industry & vertical, awesome blog from the @MadrigalPharma team on using Deep Agents, Skills…

agents are going and are already changing every industry & vertical, awesome blog from the @MadrigalPharma team on using Deep Agents, Skills, & LangSmith at the frontier of biopharma it’s an awesome t

safetyharrison-chase--x

AI-Residual Economies (Data → Income) Users begin to expect compensation for their contribution to AI systems, not just free usage. These pl…

AI-Residual Economies (Data → Income) Users begin to expect compensation for their contribution to AI systems, not just free usage. These platforms track, attribute, and pay micro-royalties when user

safetyyohei-nakajima--x

ANCHOR: A Physically Grounded Closed-Loop Framework for Robust Home-Service Mobile Manipulation

arXiv:2604.25323v1 Announce Type: new Abstract: Recent advances in open-vocabulary mobile manipulation have brought robots into real domestic environments. In such settings, reliable long-horizon exec

safetyarxiv-cs-ro

Beyond Accuracy: Benchmarking Cross-Task Consistency in Unified Multimodal Models

arXiv:2604.25072v1 Announce Type: new Abstract: Unified Multimodal Models (uMMs) aim to support both visual understanding and visual generation within a shared representation. However, existing evalua

Model Releases

Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal

model-releasesarxiv-cs-cl

arXiv:2509.09708v3 Announce Type: replace Abstract: Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour re

Carbon-Taxed Transformers: A Green Compression Pipeline for Overgrown Language Models

arXiv:2604.25903v1 Announce Type: cross Abstract: The accelerating adoption of Large Language Models (LLMs) in software engineering (SE) has brought with it a silent crisis: unsustainable computationa

CHUCKLE -- When Humans Teach AI To Learn Emotions The Easy Way

arXiv:2510.09382v2 Announce Type: replace Abstract: Curriculum learning (CL) structures training from simple to complex samples, facilitating progressive learning. However, existing CL approaches for

- Comey indicted for tweeting a number. - Trump FCC threatens ABC's broadcast license. - Trump defacing more govt institutions with his name…

- Comey indicted for tweeting a number. - Trump FCC threatens ABC's broadcast license. - Trump defacing more govt institutions with his name and picture. - Trump's kids cashing in on huge govt contrac

safetyyann-lecun--x

Compute Aligned Training: Optimizing for Test Time Inference

arXiv:2604.24957v1 Announce Type: new Abstract: Scaling test-time compute has emerged as a powerful mechanism for enhancing Large Language Model (LLM) performance. However, standard post-training para

Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers

arXiv:2604.25891v1 Announce Type: new Abstract: Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavio

CORAL: Adaptive Retrieval Loop for Culturally-Aligned Multilingual RAG

arXiv:2604.25676v1 Announce Type: new Abstract: Multilingual retrieval-augmented generation (mRAG) is often implemented within a fixed retrieval space, typically via query or document translation or m

Crazy how many people expect some company or country to achieve a “insurmountable lead” in AI, when in reality no company seems to hold a le…

Crazy how many people expect some company or country to achieve a “insurmountable lead” in AI, when in reality no company seems to hold a lead of more that a few weeks or at most a few months before s

CroSearch-R1: Better Leveraging Cross-lingual Knowledge for Retrieval-Augmented Generation

arXiv:2604.25182v1 Announce Type: new Abstract: A multilingual collection may contain useful knowledge in other languages to supplement and correct the facts in the original language for Retrieval-Aug

Model Releases

Cross-Lingual Jailbreak Detection via Semantic Codebooks

model-releasesarxiv-cs-cl

arXiv:2604.25716v1 Announce Type: new Abstract: Safety mechanisms for large language models (LLMs) remain predominantly English-centric, creating systematic vulnerabilities in multilingual deployment.

Damning

Damning 🚨 Altman texts Musk: 'we offered you equity when we established the capped profit. you didn't want it at the time.' (Altman drafted this with Shivon Zilis, then sent it to Musk that night.) Op

DEGround: An Effective Baseline for Ego-centric 3D Visual Grounding with a Homogeneous Framework

arXiv:2506.05199v3 Announce Type: replace Abstract: A core task in embodied intelligence is ego-centric 3D visual grounding. Existing methods typically adopt two-stage, heterogeneous pipelines that pa

DGLight: DQN-Guided GRPO Fine-Tuning of Large Language Models for Traffic Signal Control

arXiv:2604.25259v1 Announce Type: new Abstract: Traffic signal control (TSC) plays a central role in reducing congestion and maintaining urban mobility. This dissertation introduces DGLight, a critic-

DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

arXiv:2604.25050v1 Announce Type: new Abstract: Unlike chatbots, physical AI must act while the world keeps evolving. Therefore, the inter-chunk pause of synchronous executors are fatal for dynamic ta

safetyarxiv-cs-ro

DSO: Direct Steering Optimization for Bias Mitigation

Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Y

safetyapple-ml-research

Elon’s lawsuit might have merit even his motivations are questionable, as I just explained on @cnni https://video.snapstream.net/Play/7vHDLC…

Gary Marcus discusses on CNN the legal merits of Elon Musk's lawsuit, arguing that despite questionable motivations behind the case, there may be legitimate legal grounds supporting it. The post refer

Enterprises turn to runtime security to close the agentic AI trust gap

As enterprises push agentic AI out of the proof-of-concept phase and into production, AI runtime security — the ability to enforce policy at the exact moment an agent acts — is proving to be the bedro

safetysiliconangle

Frictive Policy Optimization for LLMs: Epistemic Intervention, Risk-Sensitive Control, and Reflective Alignment

arXiv:2604.25136v1 Announce Type: new Abstract: We propose Frictive Policy Optimization (FPO), a framework for learning language model policies that regulate not only what to say, but when and how to

@GaryMarcus @TheParisSpleen Gary, that was a great article. Wonderful work as always sir. You have been proven accurate once again.

This is a complimentary response to an article by Gary Marcus, praising his accuracy and work quality. The tweet does not provide specific details about the article's content, only that it was well-re

How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum

arXiv:2604.25907v1 Announce Type: new Abstract: Adapting reasoning models to new tasks during post-training with only output-level supervision stalls under reinforcement learning from verifiable rewar

How RL Unlocks the Aha Moment in Geometric Interleaved Reasoning

arXiv:2603.01070v2 Announce Type: replace Abstract: Solving complex geometric problems inherently requires interleaved reasoning: a tight alternation between constructing diagrams and performing logic

I-INR: Iterative Implicit Neural Representations

arXiv:2504.17364v4 Announce Type: replace Abstract: Implicit Neural Representations (INRs) have revolutionized signal processing and computer vision by modeling signals as continuous, differentiable f

👇 I was there for this, sitting right next to Altman. Realizing a few months later he lied about it (by omission) was what turned me agains…

👇 I was there for this, sitting right next to Altman. Realizing a few months later he lied about it (by omission) was what turned me against him. We had sworn to tell the whole truth and nothing but t

Interactive Episodic Memory with User Feedback

arXiv:2604.24893v1 Announce Type: new Abstract: In episodic memory with natural language queries (EM-NLQ), a user may ask a question (e.g., 'Where did I place the mug?') that requires searching a long

La industria de las IA vive exclusivamente de nuestra creencias y supuesto de que podría hacer lo que no hace.

La industria de las IA vive exclusivamente de nuestra creencias y supuesto de que podría hacer lo que no hace. Sheer insanity. Amazon, Google, Microsoft, and Meta collectively are spending more money

Learning-Based Dynamics Modeling and Robust Control for Tendon-Driven Continuum Robots

arXiv:2604.25691v1 Announce Type: new Abstract: Tendon-Driven Continuum Robots (TDCRs) pose significant modeling and control challenges due to complex nonlinearities, such as frictional hysteresis and

safetyarxiv-cs-ro

Learning from Noisy Preferences: A Semi-Supervised Learning Approach to Direct Preference Optimization

arXiv:2604.24952v1 Announce Type: new Abstract: Human visual preferences are inherently multi-dimensional, encompassing aesthetics, detail fidelity, and semantic alignment. However, existing datasets

Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System

arXiv:2604.24921v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models are a promising paradigm for generalist robotic manipulation by grounding high-level semantic instructions into ex

LOL. Just drop the stupid supply chain risk designation, and admit your mistake.

LOL. Just drop the stupid supply chain risk designation, and admit your mistake. SCOOP: The White House is developing guidance that would allow agencies to get around Anthropic's supply chain risk des

MAIC-UI: Making Interactive Courseware with Generative UI

arXiv:2604.25806v1 Announce Type: new Abstract: Creating interactive STEM courseware traditionally requires HTML/CSS/JavaScript expertise, leaving barriers for educators. While generative AI can produ

Misleading indeed.

Misleading indeed. 🚨 OpenAI's lawyer William Savitt just attacked Musk on cross-examination: 'you pledged 1 billion, only delivered 38 million.' That's misleading. The $1B was a COLLECTIVE pledge from

MolReFlect: Towards In-Context Fine-grained Alignments between Molecules and Texts

arXiv:2411.14721v2 Announce Type: replace Abstract: Molecule discovery is a pivotal research field, impacting everything from medicine to materials. Recently, Large Language Models (LLMs) have been wi

Navigating Global AI Regulation: A Multi-Jurisdictional Retrieval-Augmented Generation System

arXiv:2604.25448v1 Announce Type: new Abstract: Navigating AI regulation across jurisdictions is increasingly difficult for policymakers, legal professionals, and researchers. To address this, we pres

NimbleReg: A light-weight deep-learning framework for diffeomorphic image registration

arXiv:2503.07768v2 Announce Type: replace Abstract: This paper presents NimbleReg, a light-weight deep-learning (DL) framework for diffeomorphic image registration leveraging surface representation of

One Refiner to Unlock Them All: Inference-Time Reasoning Elicitation via Reinforcement Query Refinement

arXiv:2604.25444v1 Announce Type: new Abstract: Large Language Models (LLMs) often fail to utilize their latent reasoning capabilities due to a distributional mismatch between ambiguous human inquirie