• Picking an optimizer finally has a map you can consult. OmniOpt breaks 100-plus methods into a five-stage meta-pipeline and lines them up side by side, finding that most "new optimizers" touch only one or two stages — the mechanistic differences are systematically overstated.

• Gemma 4 wants the default slot in your self-hosted stack. From 2.3B to 31B, dense and MoE both, natively multimodal across the line. Evaluate it by deployment position: one tier for on-device, one for multimodal, one for activation cost.

• Game worlds are now generated frame by frame by a model. AlayaWorld synthesizes playable worlds autoregressively with a video world model, targeting the content-production pipeline itself. The whole stack is open source.

• RL training reaches the image-generation step for the first time. BRAID models interleaved text-image reasoning as an MDP, so policy gradient no longer stops at text steps. Generation becomes something reward can shape.

Also Notable

• Concept Erasure Gets a Manifold Constraint — Assume representations concentrate on a low-dimensional manifold, then keep the intervention inside it to avoid damaging other concepts tied to the target. MANCE
• Move Flow-Matching Supervision Out of VAE Latent Space — Do velocity regression in pretrained perceptual feature space instead, cutting sampling steps from 35-50 down to 4-8. Perceptual Flow Matching
• Don't Just Compress Context for Long-Horizon Action Prediction — PraMem replaces hard-compression memory management with "experience memory accumulated in practice" to infer the next action. PraMem
• Dense Video Captioning Is Too Slow Token by Token — Parallelized autoregressive decoding lifts inference efficiency for long videos with high event density. Parallelized Autoregressive Decoding
• GRPO Stalls Outside the Student's Own Sampling Support on Hard Problems — TREK uses distillation to expand exploration rather than imitate, then refines with RL. TREK
• Bidirectional Diffusion Is Consistent but Slow, Autoregressive Fast but Local — Flex-Forcing tries to unify both video-diffusion paradigms in one framework. Flex-Forcing
• 3D Chest CT Synthesis Gets RL Post-Training — A rectified-flow transformer generates in latent space, then RL aligns it to structured radiology conditions. CONFLUX
• A Heterogeneous-Graph Foundation Model for Metabolic Engineering — Aims to make strain design learn from experimental data instead of relying only on stoichiometric constraint models. Canopy
• With a DINOv2 Teacher, Semi-Supervised Segmentation Changes Its Bottleneck — The problem shifts from "which pseudo-label to trust" to how the embedding space is structured by class. PixCon
• Generate Randomized Counterfactual Backgrounds — Paired with cross-variant self-supervision, it breaks the model's spurious correlation between background and label. Breaking Spurious Correlations

Read the full edition →