• Fine-tuning alone teaches LLMs to output multiple tokens per step. MARS needs no architecture changes and no extra parameters. Qwen2.5-7B hits 1.71x wall-clock speedup with near-zero migration cost.

• Image autoencoder collapse isn't a channel problem. TC-AE shows the real bottleneck is token utilization. A two-stage compression path fixes it without adding complexity.

• World models no longer trade spatial consistency for real-time speed. INSPATIO-WORLD splits the two concerns into separate modules, generating navigable 4D scenes from a single video input.

• RL alignment for diffusion models doesn't need full precision everywhere. FP4 for exploration, BF16 for training. Convergence speeds up by up to 4.64x with no quality loss.

Also Notable

• Text, Layout, and Editing Instructions All Become Visual Prompts — FlowInOne unifies multimodal generation as image-in image-out flow matching, removing text as a required control interface.
• Motion Control and Camera Angle Finally Decoupled — NVIDIA's MoRight lets users specify object motion without inadvertently affecting camera movement, with physically plausible chain reactions.
• Reward Model Benchmarks' Blind Spot: Personal Preferences — Personalized RewardBench reveals that existing evaluations test general quality but not whether models distinguish individual user preferences.
• Not All Regions Need Full Resolution — Q-Zoom lets MLLMs adaptively select which visual regions need fine-grained perception based on the query, preventing attention saturation from irrelevant tokens.
• Catastrophic Forgetting in Test-Time Training Has a Fix — Elastic weight consolidation stabilizes inference-time updates in long-sequence 3D reconstruction, preventing new observations from overwriting old memories.
• Which KV Cache Entries Matter at a Million Tokens? — StructKV retains structural skeleton tokens rather than high-attention-score ones, rethinking compression strategy for long-context inference.
• MoE Expert Weights Compressed to 1-Bit — MoBiE achieves extreme binarization while handling inter-expert redundancy, opening new compression territory for MoE deployment.
• Where Does the Reasoning Chain Break? — Step Saliency pinpoints fracture points in long reasoning chains, finding errors often occur in intermediate steps rather than final outputs.
• Users Correct RAG Errors Post-Deployment, but Benchmarks Don't Care — Existing RAG benchmarks are fully static, ignoring whether systems can learn from deployed user feedback.
• Pretraining Synthetic Data Should Fuse Across Documents — WRAP++ upgrades from single-document rewriting to cross-document fusion, exposing models to cross-source reasoning patterns during pretraining.

Read the full edition →