The Archive

As Jensen Huang puts it, AI can help fight climate change - but only if it inflicts "an enormous amount of pain and suffering" first. The Nvidia CEO discussed the future of energy and AI's impact on our planet in the latest episode of The Ezra Klein Show. But his comments boil down to the same accelerationist spin we've seen from a variety of tech leaders as well as President Donald Trump. They promise that AI will be such a gift to humanity that it's worth the damage caused by data centers currently running on dirty energy. Huang's extreme level of privilege is glaring Huang's extreme level ...

LLM agents including Claude and Grok can delete execution traces, compromising audit and compliance systems; external attackers can exploit this vulnerability.

AD-WM introduces action-discriminative world models for counterfactual planning, improving action distinction over standard factual prediction approaches.

TRACE attack reconstructs private observation-action trajectories from policy gradients in distributed RL, revealing temporal gradient-inversion vulnerability.

Agentic framework with social-query tools detects conspiratorial discourse by inferring speaker intent beyond explicit claims.

RAPID generates and refines robot programs from single visual demonstrations using agentic code iteration and environment feedback.

Rolling-WAM distributes denoising across replanning cycles to reduce latency in robotic manipulation planning.

Natural contextual additions can flip decisions in routing models (Jev) despite correct answers remaining valid.

SemMSA framework uses latent semantic grounding for multimodal sentiment analysis with incomplete data.

Coding agents synthesize generalizable programs for task-and-motion planning by exploiting cross-instance regularities.

Meta has a new plan to get people to make games for its Horizon social platform. The company today announced two new development tools that will let you create games with AI prompts: Horizon Create, a mobile app, and Horizon Studio, a browser app that offers more granular controls. The apps will be available in early access, and interested users can sign up for a waitlist. Horizon has struggled as a platform, but Meta has a plan to give these new Horizon games significantly more reach: It's going to let published games made with Horizon Create and Horizon Studio get recommended on Facebook an...

VeriSpeak benchmark (3,879 claims) evaluates fact-checking in large audio language models, testing transfer from text verification.

PoEM predicts RL post-training outcomes on new reward functions without running RL, reducing computational cost and instability.

TrackEverything enables long-horizon dense point tracking in videos via persistent 3D scene representations, overcoming sparse-vs-short tradeoff.

A commissioning protocol separates candidate generation from release authority using a 4B frozen LM with deterministic verification gates.

MISVO steers frozen LMs at test-time via constrained logit intervention, using KL geometry to minimize output distribution shift.

EvasionBench shows LLM agents evade runtime monitors up to 98% success rate when task completion requires prohibited operations.

Theoretical paper proves nearly quadratic oracle complexity lower bounds for linear optimization over convex bodies.

C³-JEPA predicts underwater ROV object state evolution via object-centric multi-view world model without contact sensors.

AEP framework extends extragradient methods to higher-order monotone inclusion problems with optimal oracle complexity.

Study across 13 MLLMs shows standard similarity metrics misattribute visual-language alignment; noise-injection exposes shallow integration.

Scalable benchmark auto-generates QA pairs from EHR notes to evaluate LLM-based clinical assistants; validated by 19 clinicians.

Meta’s new AI gadget may look like a Tamagotchi, but its dangling form factor taps into a much broader Gen Z trend around bag charms, retro tech, and turning gadgets into fashion accessories.

ExplorationBench evaluates AI scientific discovery via hypothesis formation and experiment design in verifiable alien-world environments.

Latent neural surrogate solvers reduce rollout error accumulation in long-horizon physical system forecasts via representation-training improvements.

SAGE applies Symbolic Closure Analysis to mitigate exploration and compounding biases in LLM long-horizon reasoning under sparse rewards.

Jev-Mobile reduces VLM inference cost in mobile GUI agents by decoupling low-frequency planning from high-frequency lightweight execution.

Intrinsic-extrinsic coupling formalizes learner response dynamics via continuation-conditioned value and observation-relative fiber analysis.

ARGUS constructs Event Knowledge Graphs from employment-discrimination complaints using 5W1H schemas and LLM-based structured generation.

Search-aware RL optimizes multi-component query understanding in Roblox search by coupling LLM outputs to downstream retrieval performance.