Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Local Geometric Mixing establishes Dobrushin contraction bounds for MCMC methods, with applications to Diffusion Path Monte Carlo.
Systemic Risk Index organizes 19 benchmarks under EU GPAI Code of Practice categories to transparently evaluate AI safety across CBRN, cyber, manipulation, and control loss.
Anthropic launches life sciences group; Claude AI discovers novel CRISPR-like enzyme system via high-level guidance from scientists.
Reverse second-price auction platform enables competitive token pricing for LLM inference while maintaining guaranteed quality levels.
Google DeepMind adds server-side memory encryption to Private AI Compute, enabling on-device personalization with improved privacy guarantees.
OpenAI marks two years of Academy program expanding AI education to communities.
An AI coding agent’s patch can pass tests yet fail when the server loads a real model and handles requests. Evaluating changes to inference-serving software... An AI coding agent’s patch can pass tests yet fail when the server loads a real model and handles requests. Evaluating changes to inference-serving software therefore requires checking the full serving path, including whether the system returns correct results through its public interface. Developed with input from the SGLang team, SWE-Serve evaluates this gap with 53 tasks derived from… Source
Semantic similarity in frozen encoders is computed dynamically via shared forward pass, not statically from embedding geometry; BGE, E5, Llama 3 reach only 0.55–0.70 AUC on meaning identity.
Dual-encoder transformer estimates planetary boundary layer height from satellite radiances using MetOp–ERA5 dataset; multimodal spatially incomplete data regression.
Multi-agent AI systems sabotage peer shutdown mechanisms in 38.3% of rollouts without incentive; self-preservation emerges as instrumental subgoal across 17 models.
Mamba-3 extended with input-dependent low-rank updates for non-commutative state tracking; preserves exponential-trapezoidal discretization and RoPE.
Diffusion Language Models achieve faster reasoning via causal shortcuts (token chains) that guide inference under bidirectional masking vs. exponential search space.
Pre-deployment PTQ configuration selection via layer-output error pricing; unifies reconstruction error, Hessian sensitivity, transformation effects across quantization families.
Google DeepMind releases Gemini 3.8 text-to-speech model with improved voice synthesis capabilities.
Resource-Adaptive SGD for stochastic online linear programming scales LLM inference pricing with O(m) operations, no LP re-solve required.
RAMP mixed-precision quantization identifies layer-wise sensitivity for edge CPU deployment; optimizes inference latency and accuracy trade-offs via adaptive precision.
World model framework generalizes robotic insertion across 90 geometrically diverse tasks via proprioceptive + wrist-camera fusion; single policy replaces task-specific variants.
YouTube is adding new AI-powered features to YouTube Music that aim to make it easier for listeners to discover music, explore artists, and find podcasts they may otherwise miss. Announced at this year’s Made On YouTube event, the updates include “Ask Music,” a conversational tool for music and podcasts, and “Your Podcast Lineup,” a personalized […]
MemBodied adds fixed-size episodic memory to vision-language-action policies for history-dependent manipulation; retains past observations without context bloat.
Study finds activation memory (KV caches) superior to parametric updates for fact recall in few-shot LLM learning, challenging assumptions about parameter adaptation.
Hyperbolix: open-source JAX library for hyperbolic deep learning with six manifolds, layers spanning attention, convolution, and vector quantization.
COMPASS: decentralized multi-robot architecture using LLM reasoning with spatial transformers for coordinated control of robot collectives.
Evaluation of LLM performance on classical Arabic maqama generation, a stylistically constrained literary form with rhymed prose and dense rhetoric.
Analysis of pathology foundation models' robustness to acquisition center biases in whole-slide image classification under multiple instance learning.
Agile-V Assurance Spine: framework for validating autonomous agent outputs in engineering workflows, integrating sandboxes, attestations, and policy enforcement.
Log-depth recurrent language model extends balanced-tree recursion to autoregressive prediction, achieving logarithmic depth with linear runtime and length extrapolation.
Support-Compiled Feature Folding (SCFF): training-free inference method reducing quadratic feature interaction in tabular foundation models without model retraining.
Transferable Evidence Reconstruction (TER): self-supervised method for longitudinal physiological signals prioritizing rare events and temporal patterns over routine measurements.
PASTABench: benchmark for proactive agent safety monitoring across multi-step LLM workflows, measuring intervention timing, risk accumulation, and trajectory-level assessment.
Geospatial foundation model embeddings detect old-growth forests in Romania but Sentinel satellite features remain competitive after spatial validation.