Articles & Guides

REBASE introduces a training-free framework for one-shot in-context segmentation that suppresses misleading background correspondences. By projecting features away from a reference-background subspace, it produces cleaner matches for promptable segmenters such as SAM.

Xiaomi-Robotics-1 explores whether robotics can benefit from the same scaling dynamics that transformed language and vision models. By pre-training on more than 100,000 hours of real manipulation trajectories, the model shows consistent gains as data and model size increase.

A KAIST AI paper introduces robot-centric pointmaps to reduce the mismatch between camera observations and robot-frame actions in VLA models. The method encodes 3D scene coordinates in the robot frame while keeping the image-like grid structure used by existing 2D vision backbones.

RAGU is an open-source GraphRAG engine that separates extraction from graph consolidation, aiming to make graph-based retrieval less noisy and less expensive.

A new paper argues that evaluating security agents by peak success rate alone misses a key operational constraint: cost. It compares offensive and defensive agents under fixed budgets and finds that red-team and blue-team tasks scale differently.

On-Policy Delta Distillation, or OPD², replaces direct imitation of a teacher’s full output distribution with a signal based on the difference between a reasoning-tuned teacher and its base model. The paper reports consistent gains across math, science, and code-reasoning benchmarks.

Qwen’s technical report introduces Qwen-Music, a music generation system designed for full songs with vocal singing. Its key idea is to plan melodies with melody tokens before rendering high-fidelity stereo audio.

S1-Omni aims to reduce fragmentation in AI for Science by bringing diverse scientific data types into one reasoning model. It supports more than 200 scientific tasks and releases model weights, inference code, and a 10K-sample corpus subset.

A new study tests Muon in sparse-reward agentic RL and finds that its gains depend heavily on the advantage estimator and learning rate, rather than on the optimizer alone.

A new paper argues that encrypted negotiation values are not enough: concession timing, offer trajectories, and convergence patterns can still reveal private constraints. The authors propose an adaptive randomized policy to reduce inference attacks while preserving negotiation performance.

NVIDIA has released Cosmos 3 Edge on Hugging Face, a 4B-parameter open world model for robotics and vision AI. The model is designed to combine scene understanding, prediction, simulation, and action generation on edge hardware.

Grab has built Palana, a platform designed to run autonomous AI workloads with Kubernetes-native isolation, centralized egress control, secret mediation, and operational auditability.

xHC revisits Hyper-Connections as a practical scaling axis for large language model pre-training. By expanding residual streams to N=16 while sparsely updating only part of them, it reports stronger downstream performance with modest added training cost.

A new paper featured on Hugging Face Daily Papers introduces Loopie, a Mixture-of-Experts looped Transformer family. The authors claim it outperforms standard Transformer baselines under the same compute budget and gains strong reasoning ability after post-training.

The paper uses chess as a controlled environment to study how pretraining choices shape reinforcement learning gains, and what RL changes inside a reasoning model.

Cura 1T treats healthcare as a connected set of skills, not a single medical QA problem. Its training loop aims to improve consultation, reasoning, diagnosis, and tool use together.

AV-Flamingo targets the hard problem of reasoning over long real-world audio-visual videos, not just short clips. It combines large-scale training data, staged curriculum learning, and timestamp-grounded reasoning to improve alignment and interpretability.

Recursive Harness Self-Improvement reframes the agent harness as something that can be optimized, not merely hand-written once. The paper reports that a few lightweight iterations can improve low-reasoning-effort agents while cutting inference cost.

DSWorld applies the world model idea to autonomous data science agents, predicting the outcome of workflow operations before costly execution. The paper reports around 14× faster RL-based training and 3–6× faster search-based inference.

Kimi has suspended new consumer subscriptions after a rapid increase in user requests pushed its compute cluster close to capacity. The company also plans to separate Kimi’s main product benefits from Kimi Code benefits for future users.

A new study suggests that simply having AI advice available can push people to abandon uncertainty and trust wrong answers. Confidence rises, but judgment gets worse.