The Archive

GridSFM, a 15M-parameter GNN foundation model pretrained on 54 power grid topologies, solves AC Optimal Power Flow with 2.45% zero-shot error.

Audio language models represent phoneme distinctive features similarly across speech and text modalities when measured via mean offset cosine.

Token-level wildcard CTC arcs tolerate transcription ambiguity in speech recognition by preserving frame-level supervision on phonetically ambiguous segments.

RL applied to quantum network entanglement scheduling for simultaneous multipartite tasks.

A pair of developers say that with very little prompting, Meta's Muse will share its entire filesystem with you. Peter James and Jonny L. Saunders have said they both independently coaxed Muse into zipping up and sharing the entire contents of its root filesystem, Ubuntu system files, app templates, and internal documentation. Saunders posted on Mastodon that it was "extremely easy" to replicate James' results and that Muse had "Almost no prompt injection resistance." Meta denies that the incident represents a security breach. As noted in its announcement post, Meta's Muse runs in persistent ...

LLM explanations for rejections tested by inserting stated missing facts; reveals whether reasoning grounds in text.

Graph-based multi-agent RL for railway infrastructure maintenance scheduling at large scale.

GRASP: multi-stage LLM planning framework with isolated context modules for complex task decomposition and evaluation.

We seem to be entering into an AI agent renaissance. Meta's new consumer-facing AI agent, Muse, topped the App Store charts soon after its release and has 600,000 daily active users in the US, by an Apptopia estimate. And AI agent platform Instinct, whose eponymous creator is fundraising at a $2.5 billion valuation, has been making the rounds in the tech industry. Under the hood, though, there are some obvious similarities to the platform that started it all: OpenClaw. For the past week or so, some social media users have alleged that Muse is directly built on OpenClaw. The two platforms use ...

EnigmaForge: benchmark with SAT-generated logic puzzles hidden in document corpora; 25 models tested on 600 instances.

Simulation-based screening for LLM-powered customer experience agents pre-deployment; 140M-scale testing.

SVGLM: Vision-language model framework using SVG primitives to unify text and image in reasoning chains.

The AI-powered feature builds a virtual wardrobe from your photos, and is now broadly available after first rolling out to Android users in June.

The fireman outfit. Dead. | Photo by Sean Hollister / The Verge This is Optimizer, a weekly newsletter sent from Verge senior reviewer Victoria Song that dissects and discusses the latest gizmos and potions that swear they're going to change your life. Opt in for Optimizer here. Last night, I asked Blorbo - what I named my Muse AI agent - to help me set some health goals. I was underwhelmed. I'd asked for a workout regimen for a sub-30 minute 5K and regaining lost muscle. It generated an okay-ish but kind of mid weekly routine. But then I looked at its face. Based on what it learned about me ...

HEXIS: compiles agent skills into extended finite state machines separating knowledge from control flow.

Audio description generation formalized as constrained optimization over what, when, and how to narrate visual content.

Orbital Error Dynamics: speculative framework proposing procedural weight generation via quadratic maps; lacks empirical validation.

Proves sum-of-squares certifiability for log-concave distributions, improving moment bound algorithms.

RL-based compiler pass selection for hybrid HPC-quantum MLIR pipelines in NISQ era systems.

TechCrunch Founder Summit is a full-day gathering in Boston on November 4 where founders across all stages connect with top VCs and experienced entrepreneurs to gain tactical insights on building and scaling their companies.

R-DEIM Net: 76M-parameter dual-expert model for paraphrase detection with interpretable rationales.

TRACK: training-free trajectory routing reduces video diffusion inference cost via model-size switching.

PrivDrift: benchmark auditing LLM secret leakage persistence across topic drift in multi-turn conversations.

AT-SKM-Net: accelerated constraint-based optimization for dynamic graphs in infrastructure systems.

Learning policies to predict when answer revision improves retrieval-augmented QA via recoverability estimates.

ElevenLabs powers the AI voice on the other end of a lot of customer service calls, and its CEO told me this week that businesses should probably tell you that — at least until getting a machine is what everyone expects anyway.

Residual correlation diagnostic for multi-target GP coregionalisation uncertainty quantification benefits.

GraphStar sets extend neural network verification to GNNs with node/edge features for power systems.

Nuclear norm-regularized Bayesian matrix completion with uncertainty quantification for missing-data problems.

Self-supervised method for measuring L2 pronunciation deviation using native speech corpora and SVD without labeled data.