video games train ai for real-world robots
general intuition raised $320m to build ai agents that learn spatial reasoning from gameplay button presses, then control real robots.
topic
research
five open source omni ai models for multimodal tasks
a practical look at open source models that handle text, images, audio, and video for real-world ai applications.
on-device neural architecture search for sensor data
a lightweight neural architecture search runs directly on embedded devices to adapt tiny neural networks to new users and sensor data in real time.
hidden scale escapes per-loop loss in looped language models
dense per-loop cross-entropy does not control all recurrent state variables, leaving a readout blind spot that allows hidden-state scale to grow unchecked.
minimax pac bounds for learning in exogenous contextual mdps
new sample complexity bounds for reinforcement learning with i.i.d. contexts, showing how many oracle calls are needed before and during policy execution.
spectral phase diagram for few-shot learning
a new saturation index helps decide when to stop labeling data for binary classification.
safe multi-agent rl via constraint manifold control
a hierarchical framework enforces hard safety constraints in multi-agent reinforcement learning using a constraint manifold, providing theoretical guarantees and stable training.
federated causal discovery and inference survey
a systematic review of federated causal discovery and inference methods, covering design choices, privacy, and open challenges.
reasoning helps llms recall simple facts
google research shows that chain-of-thought reasoning improves factual recall in large language models even for simple questions, driven by a computational buffer effect and factual priming.
open far-field asr benchmark reveals real-world gap
the ffasr leaderboard is the first open benchmark evaluating asr models under realistic far-field acoustic conditions, showing a large performance gap compared to near-field tests.
what makes an ai agent truly autonomous
a new paper argues that real agency requires internalized goals, identity, decision-making, self-regulation, and learning, not just external prompts.
neuro-symbolic drive adds rule-grounded reasoning to driving vlas
a new framework extracts decision traces from rule-based planners to supervise driving vision-language-action models, making their reasoning more faithful and causally connected to motion.
gray-box workflow finds reaction networks from noisy data
a combined mcmc and gaussian process method extracts interpretable chemical equations and improves yield in experiments.
automated search finds airnet bias in mixture-of-experts design
a large-scale pipeline exploring 4-expert heterogeneous mixture-of-experts architectures reveals that alphabetical enumeration caused the search to be anchored to a single base family, limiting diversity.
offline reasoning training methods converge to similar weight updates
study finds sft, rft, and rift produce nearly identical weight changes in lora fine-tuning, while dft and offline grpo diverge.
math skills to master before coding for data science
a breakdown of essential math disciplines for data science and a practical learning roadmap.
cross-origin storage api cuts duplicate ai model downloads
a proposed browser api lets web apps share cached ai model and wasm files across origins using cryptographic hashes, reducing redundant downloads and storage.
porting moebius inpainting model to browser with claude code
a 0.2b image inpainting model was converted from pytorch to onnx and deployed as a browser demo using webgpu, with help from claude code.