All entries

Since Hinton has actually replied let me clarify some things - LLMS don’t always regurgitate - LLMs don’t literally store full texts - but…

Since Hinton has actually replied let me clarify some things - LLMS don’t *always* regurgitate - LLMs don’t literally store full texts - but given the mechanisms that they use they do sometimes regurg

safetygary-marcus--x

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Safety

Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

DGX agent

arXiv:2605.06130v2 Announce Type: replace Abstract: A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupl

safetyarxiv-cs-ai

11 May 2026

Skip-It? Theoretical Conditions for Layer Skipping in Vision-Language Models

arXiv:2509.25584v2 Announce Type: replace Abstract: Vision-language models achieve incredible performance across a wide range of tasks, but their large size makes inference costly. Recent work has sho

researcharxiv-cs-ai

🧵 Slime: The Most Elegant & Comfortable RL Training Framework Ever A deep dive into why Slime redefines LLM RL training with clean architec…

🧵 Slime: The Most Elegant & Comfortable RL Training Framework Ever A deep dive into why Slime redefines LLM RL training with clean architecture & production-grade engineering ✨ Insights from Zhihu con

hardwarezhipu-ai--x

SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks

model-releasesarxiv-cs-ai

arXiv:2603.24755v2 Announce Type: replace-cross Abstract: Software development is iterative, yet agentic coding benchmarks hide design issues through their single-shot setup. Recent iterative benchmar

Tutorials

SLOPE: Optimistic Potential Landscape Shaping for Model-based Reinforcement Learning

arXiv:2602.03201v3 Announce Type: replace Abstract: Model-based reinforcement learning (MBRL) is sample-efficient but struggles in sparse reward settings. A critical bottleneck arises from the lack of

tutorialsarxiv-cs-lg

Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation

arXiv:2605.07950v1 Announce Type: new Abstract: We study Slowly Annealed Langevin Dynamics (SALD), a sampler for tracking a path of moving target distributions and approximating the terminal target th

safetyarxiv-cs-lg

SmellBench: Evaluating LLM Agents on Architectural Code Smell Repair

model-releasesarxiv-cs-cl

arXiv:2605.07001v1 Announce Type: cross Abstract: Architectural code smells erode software maintainability and are costly to repair manually, yet unlike localized bugs, they require cross-module reaso

SMT-Based Active Learning of Weighted Automata

arXiv:2605.07758v1 Announce Type: cross Abstract: We present an SMT-based active learning algorithm for nondeterministic weighted automata (WFAs) as a practical and robust alternative to Hankel/L*-sty

SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests

researchmicrosoft-research

Using SocialReasoning Bench, we observed a stable pattern across models—agents execute competently, but fail to consistently improve the user’s position, even with explicit instructions to optimize fo

SOCKET: SOft Collision Kernel EsTimator for Sparse Attention

arXiv:2602.06283v2 Announce Type: replace Abstract: Exploiting sparsity during long-context inference is key to scaling large language models, as attention dominates the cost of autoregressive decodin

hardwarearxiv-cs-lg

SOD: Step-wise On-policy Distillation for Small Language Model Agents

arXiv:2605.07725v1 Announce Type: cross Abstract: Tool-integrated reasoning (TIR) is difficult to scale to small language models due to instability in long-horizon tool interactions and limited model

Applications

SoftBank partners with South Korea's Cosmos Lab and DeltaX to mass produce next-gen data center batteries in Japan at the scale of one GWh per year in FY 2027 (Min-Jeong Lee/Bloomberg)

Min-Jeong Lee / Bloomberg: SoftBank partners with South Korea's Cosmos Lab and DeltaX to mass produce next-gen data center batteries in Japan at the scale of one GWh per year in FY 2027 — SoftBank Gro

applicationstechmeme

Local Ai

SoftSAE: Dynamic Top-K Selection for Adaptive Sparse Autoencoders

arXiv:2605.06610v2 Announce Type: replace-cross Abstract: Sparse Autoencoders (SAEs) have become an important tool in mechanistic interpretability, helping to analyze internal representations in both

local-aiarxiv-cs-cv

SoLAR: Error-Resilient Streamable Long-Horizon Free-Viewpoint Video Reconstruction with Anchor Activation and Latent Recalibration

arXiv:2605.07346v1 Announce Type: new Abstract: Free-Viewpoint Video (FVV) has emerged as a cornerstone of next-generation immersive media systems and attracted widespread attention. Previous methods

Solving Max-Cut to Global Optimality via Feasibility-Preserving Graph Neural Networks

arXiv:2605.07113v1 Announce Type: new Abstract: Exact solution of hard combinatorial optimization problems often relies on strong convex relaxations, but solving these relaxations repeatedly inside a

Agents

SOM: Structured Opponent Modeling for LLM-based Agents via Structural Causal Model

arXiv:2605.07301v1 Announce Type: new Abstract: Accurately predicting opponents' behavior from interactions is a fundamental capability for large language model (LLM)-based agents in multi-agent and g

agentsarxiv-cs-ai

Sony's failed war against Internet piracy may doom other copyright lawsuits

In a 9-0 Supreme Court decision, Cox Communications prevailed against Sony Music Entertainment in a landmark copyright case decided March 25, 2026, with the Court finding that internet service provide

industryars-technica

Source: new revenue sharing terms cap OpenAI's payments to Microsoft at 38B; they previously could have totaled 135B through 2030 if long-term goals were met (The Information)

The Information: Source: new revenue sharing terms cap OpenAI's payments to Microsoft at 38B; they previously could have totaled 135B through 2030 if long-term goals were met — OpenAI Chief Financial

Source: Robinhood has filed confidentially for its second publicly traded venture fund, Robinhood Venture Fund II, focusing on early-stage startups (Lucinda Shen/Axios)

Lucinda Shen / Axios: Source: Robinhood has filed confidentially for its second publicly traded venture fund, Robinhood Venture Fund II, focusing on early-stage startups — Robinhood has filed confiden

Sources: German defense tech startup Helsing is set to raise 1.2B led by Dragoneer and Lightspeed at a valuation of about 18B, up from $14B in June 2025 (Financial Times)

Financial Times: Sources: German defense tech startup Helsing is set to raise 1.2B led by Dragoneer and Lightspeed at a valuation of about 18B, up from 14B in June 2025 — German company backed by Spot

Sources: Kuaishou plans to spin off its Kling AI video unit for an IPO in 2027 and is seeking a $20B valuation in pre-IPO funding talks with potential investors (The Information)

The Information: Sources: Kuaishou plans to spin off its Kling AI video unit for an IPO in 2027 and is seeking a $20B valuation in pre-IPO funding talks with potential investors — Chinese social media

Sources: Masayoshi Son has held talks with French President Emmanuel Macron about unveiling a multibillion-dollar AI data center project in the coming weeks (Bloomberg)

Bloomberg: Sources: Masayoshi Son has held talks with French President Emmanuel Macron about unveiling a multibillion-dollar AI data center project in the coming weeks — SoftBank Group Corp. founder M

Sources: OpenAI facilitated a 6.6B secondary share sale in October 2025 for 600+ current and ex-employees; over 75 people cashed out the maximum cap of 30M (Berber Jin/Wall Street Journal)

Berber Jin / Wall Street Journal: Sources: OpenAI facilitated a 6.6B secondary share sale in October 2025 for 600+ current and ex-employees; over 75 people cashed out the maximum cap of 30M — Employee

Sources: the White House's Office of the National Cyber Director and Commerce Department's CAISI are fighting over which agency should lead AI model evaluations (Washington Post)

Washington Post: Sources: the White House's Office of the National Cyber Director and Commerce Department's CAISI are fighting over which agency should lead AI model evaluations — As the White House g

safetytechmeme

SPACEX: Marmac 31 'You'll Thank Me Later' now has a cover installed. This barge will transport the Super Heavy Boosters from Starbase to the…

SpaceX has installed a cover on the Marmac 31 barge, a vessel designated to transport Super Heavy Boosters from Starbase to unspecified destinations. The barge, nicknamed 'You'll Thank Me Later,' repr

industryelon-musk--x

Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache

arXiv:2605.06763v1 Announce Type: new Abstract: Sparse attention improves LLM inference efficiency by selecting a subset of key-value entries, but at the cost of potential accuracy degradation. In par

hardwarearxiv-cs-lg

Sparse Attention as Compact Kernel Regression

arXiv:2601.22766v3 Announce Type: replace Abstract: Recent work has revealed a link between self-attention mechanisms in transformers and test-time kernel regression via the Nadaraya-Watson estimator,

Sparse Autoencoders as Plug-and-Play Firewalls for Adversarial Attack Detection in VLMs

arXiv:2605.07447v1 Announce Type: cross Abstract: Vision-language models (VLMs) have advanced rapidly and are increasingly deployed in real-world applications, especially with the rise of agent-based

Sparse Random-Feature Neural Networks with Krylov-Based SVD for Singularly Perturbed ODE

model-releasesarxiv-cs-lg

arXiv:2605.07286v1 Announce Type: cross Abstract: Random-feature neural networks (RFNNs), including architectures with fixed hidden layers and analytically determined output weights, offer fast traini

Sparser, Faster, Lighter Transformer Language Models

arXiv:2603.23198v2 Announce Type: replace-cross Abstract: Scaling autoregressive large language models (LLMs) has driven unprecedented progress but comes with vast computational costs. In this work, w

hardwarearxiv-cs-cl

SparseRL-Sync: Lossless Weight Synchronization with ~100x Less Communication

arXiv:2605.07330v1 Announce Type: cross Abstract: In large-scale reinforcement learning (RL) systems with decoupled Trainer-Rollout execution, the Trainer must regularly synchronize policy weights to

speaking of things that have gotten over a threshold for me, the combo of the new ChatGPT model, personality, and personalization feels like…

Sam Altman comments on OpenAI's new ChatGPT model, noting that the combination of improved capabilities, personality features, and personalization options has crossed an important threshold of functio

industrysam-altman--x

SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting

arXiv:2605.07243v1 Announce Type: new Abstract: Speculative decoding accelerates LLM inference by drafting a tree of candidate continuations and verifying it in one target forward. Existing drafters f

researcharxiv-cs-cl

Spectral Dynamics in Deep Networks: Feature Learning, Outlier Escape, and Learning Rate Transfer

arXiv:2605.07870v1 Announce Type: cross Abstract: We study the evolution of hidden-weight spectra in wide neural networks trained by (stochastic) gradient descent. We develop a two-level dynamical mea

researcharxiv-cs-ai

Spectral Filtering for Complex Linear Dynamical Systems

arXiv:2601.22400v2 Announce Type: replace-cross Abstract: We study the problem of learning complex-valued linear dynamical systems (CLDS) with sector-bounded spectrum. This class captures oscillatory

researcharxiv-cs-ai

Spectral Surgery: Class-Targeted Post-Hoc Rebalancing via Hessian Spike Perturbation

arXiv:2605.07790v1 Announce Type: cross Abstract: The Hessian spectrum of trained deep networks exhibits a characteristic structure: a continuous bulk of near-zero eigenvalues and a small number of la

Spectrum-Adaptive Generalization Bounds for Trained Deep Transformers

arXiv:2605.07297v1 Announce Type: cross Abstract: Understanding why trained Transformers generalize well is a fundamental problem in modern machine learning theory, and complexity-based generalization

SphereVAD: Training-Free Video Anomaly Detection via Geodesic Inference on the Unit Hypersphere

arXiv:2605.08003v1 Announce Type: new Abstract: Video anomaly detection (VAD) aims to automatically identify events that deviate from normal patterns in untrimmed surveillance videos. Existing methods

SpikingBrain: Spiking Brain-inspired Large Models

arXiv:2509.05276v4 Announce Type: replace-cross Abstract: Mainstream Transformer-based large language models face major efficiency bottlenecks: training computation scales quadratically with sequence

hardwarearxiv-cs-ai

Applications

SplatWeaver: Learning to Allocate Gaussian Primitives for Generalizable Novel View Synthesis

arXiv:2605.07287v1 Announce Type: new Abstract: Generalizable novel view synthesis aims to render unseen views from uncalibrated input images without requiring per-scene optimization. Recent feed-forw

applicationsarxiv-cs-cv

SR^2-LoRA: Self-Rectifying Inter-layer Relations in Low-Rank Adaptation for Class-Incremental Learning

model-releasesarxiv-cs-cv

arXiv:2605.07420v1 Announce Type: cross Abstract: Pre-trained models with parameter-efficient fine-tuning (PEFT) have demonstrated promising potential for class-incremental learning (CIL), yet catastr

SREGym: A Live Benchmark for AI SRE Agents with High-Fidelity Failure Scenarios

model-releasesarxiv-cs-ai

arXiv:2605.07161v1 Announce Type: new Abstract: AI agents are increasingly used to diagnose and mitigate failures in production systems, known as agentic Site Reliability Engineering (SRE). Current SR

SSP-based construction of evaluation-annotated data for fine-grained aspect-based sentiment analysis

arXiv:2605.07446v1 Announce Type: new Abstract: We report the construction of a Korean evaluation-annotated corpus, hereafter called 'Evaluation Annotated Dataset (EVAD)', and its use in Aspect-Based

researcharxiv-cs-cl

Local Ai

ST-Gen4D: Embedding 4D Spatiotemporal Cognition into World Model for 4D Generation

arXiv:2605.07390v1 Announce Type: new Abstract: Generative models have achieved success in producing apparently coherent 2D videos, but remain challenging in the physical world due to lack of 4D spati

local-aiarxiv-cs-cv

Stabilized neural Hamilton--Jacobi--Bellman solvers: Error analysis and applications in model-based reinforcement learning

arXiv:2605.07116v1 Announce Type: cross Abstract: Physics-informed neural solvers offer a promising route to model-based reinforcement learning in continuous time, where optimal feedback synthesis is

Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control

model-releasesarxiv-cs-lg

arXiv:2605.07182v1 Announce Type: new Abstract: Training a family of large language models (LLMs), either from scratch or via iterative compression, is prohibitively expensive and inefficient, requiri

STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation

arXiv:2605.08029v1 Announce Type: new Abstract: Deep generative models have advanced rapidly across text and vision, motivating unified multimodal systems that can understand, reason over, and generat