News

A near-frontier mobile agent is published with fully open training environments and recipes: MobileGym++ simulates 35 commercial apps with cross-app workflows, with controllable reset and verification. Relevant for anyone building on-device agents.

A new Benchmark measuring how well video generation models keep a Coherent memory of the world they generate: previously observed content preserved and coherent evolution over time, alongside visual fidelity and duration.

ClawLoop (LongCat) trains sandbox LLM agents through a self-sustaining three-phase cycle — task synthesis, SFT/RL training, rollout diagnosis — where the diagnosis feeds the next task synthesis, removing the external model that normally synthesizes tasks and trajectories. With Qwen3.6-27B it improves on SkillsBench, Claw-Eval-Live and ClawGym-Bench; the refined tasks transfer with gains even to Qwen3.5-4B and 9B.

New paper on Hugging Face: aligning the latent representations of image generators with different architectures (diffusion and flow matching) is still an open problem. InstantFusion proposes a shared latent interface to make generators interoperable.

A new paper shows a single Transformer block, applied recurrently, matches a full-depth vision encoder's accuracy at comparable Inference FLOPs. The key is representing the FFN at each recurrent depth as a convex combination of a small shared expert bank, with no intermediate feature distillation.

A new paper shows how a multimodal model can internalize the spatial reasoning ability of a tool-using agent, cutting Inference cost: the model learns from verified traces of its own executions, with no reliance on external tools.

A Transformer trained to predict the conditional marginals of missing variables, the key ingredient for computing Bayes risk and the value of one more observation before deciding under incomplete information.

The PaLoRA paper shows that the step-size restriction in LoRA-based continual learning should grow with the effective rank of accumulated knowledge, and proposes an adaptive pacing scheme gaining 4% accuracy on 50-task benchmarks.

Kandinsky Lab publishes the Kandinsky 6.0 Video paper: foundation models for synchronized video and audio generation, the day's #1 paper on Hugging Face with 8 models and 6 Spaces already linked.

Lightricks releases LTX-2.5, open weights: video generation with synchronized audio from text, image, or video, with native multishot that holds identity and style across scenes, a new diffusion video decoder replacing the VAE stage, and a Gemma 4 12B text encoder. The weights are built for running and Fine-tuning on your own infrastructure.

University of Central Florida researchers propose TACO, a ternary column-wise one-sparse optimizer for LLM Fine-tuning: it keeps first-order gradients while shrinking persistent optimizer state from 27.7 GB to 0.16 GB on OPT-13B.

A new method, WUSH-KV, cuts the memory and bandwidth the KV cache takes up in long-context Inference by quantizing it at low precision with data-calibrated transforms. At 2 bits, the approach matches or beats the OSCAR transform across all tested models and tasks. It's a sign that memory optimization is no longer only about the weights, but also the context memory.

The PISA paper proposes a pyramid Top-K selection strategy that brings block sparse attention to O(N log N) complexity, cutting the quadratic bottleneck of block selection; Triton kernels for training and Inference.