RESEARCH
Every Research story from RXed AI News over the last 30 days — 20 items.
August European tech funding falls 63% to €3.2B as dealmaking slows.
tech.eu/2026/09/04/august-european-tech-funding-fal…
OpenAI achieved its goal of fielding an automated research intern by September 2026. @OpenAI
unite.ai/openai-hits-goal-of-building-an-automated-…
SMELT re-evaluates Looped Transformers by fixing compute, not model size, to isolate architectural gains.
huggingface.co/papers/2609.01343
Embodied AI hits a computational wall, requiring new mathematics to advance physical robot capabilities.
therobotreport.com/edge-ai-wall-why-embodied-ai-req…
CritICL makes small LLMs reason like big ones at inference time, no retraining needed.
huggingface.co/papers/2608.27455
WikiSkill auto-discovers reusable skills from agent experience for persistent knowledge evolution.
arxiv.org/abs/2608.27454v1
LAION-BVD: 10M hours of open video for multimodal AI training.
huggingface.co/papers/2608.24845
New AI-accelerated Newton method reduces computation to one linear solve per iteration for convex functions.
arxiv.org/abs/2608.21359v1
AI4AI-Bench tests LLM agents on recursive self-improvement, aiming to make AI systems improve their own training algorithms.
arxiv.org/abs/2608.20318v1
Semantic Task Completion Video Benchmarks require AI to achieve intended outcomes, not just generate frames.
huggingface.co/papers/2608.17426
SPADE uses adaptive executable environments for AI agents to generate diverse, self-improving goals.
arxiv.org/abs/2608.19197v1
Capability-centric data design generalizes image generation beyond isolated task-specific datasets.
arxiv.org/abs/2608.18076v1
AlphaEvolve improves matrix multiplication exponent bounds using modern optimization and combination loss analysis.
arxiv.org/abs/2608.16884v1
Test-time compute inefficiency bounds large reasoning models, shifting focus to Thought-Level Beam Search.
huggingface.co/papers/2608.08020
@google may retire Gems in October, replacing them with Skills requiring a paid subscription.
testingcatalog.com/google-may-retire-gems-in-octobe…
Google Deepmind's WeatherNext forecasts cyclones a day further than leading models, matching a decade of traditional progress.
the-decoder.com/google-deepminds-weathernext-predic…
Language models can be tricked by a single bad signal. New training builds resistance. arxiv.org/abs/2608.06377v1
GaussianSelector lets you pick 3D objects with minimal clicks using graph optimization.
huggingface.co/papers/2608.01492
SmartMage uses dynamic modality orchestration for 3D scene understanding in physical AI. huggingface.co/papers/2608.05137
World-to-Wrist models treat main and wrist views as parallel, ignoring their distinct roles in robot manipulation. huggingface.co/papers/2608.05369