• Verification is emerging as a fourth scaling axis. After pretraining, post-training, and test-time compute, LLM-as-a-Verifier squeezes another layer out of your strong model — no training, no binary judgments, just an external verifier giving agentic tasks fine-grained feedback.

• PixWorld rips the VAE middleware out of 3D generation. Reconstruction and generation both move back to pixel space under one supervision signal. One fewer pretrained component, one fewer layer of encoding loss, and the objective shifts from aligning a latent to aligning 3D fidelity.

• Models keep getting stronger after they ship. EdgeBench analyzed 38,000 hours of real agent interactions and found post-deployment environment learning follows a log-sigmoid curve fitting at R²=0.998 — with learning speed doubling every three months.

• There's finally a ruler for whether deformable objects should be learned in pixels or geometry. Deform360 releases 198 objects and 215 hours of visuotactile data, putting the 2D-pixel and 3D-geometry camps on the same table.

• Add audio to a big model and text intelligence takes a hit — or does it. Audex builds on a strong text MoE backbone with a single unified-token decoder, claiming audio SOTA with "almost no regression" on text. That last claim is exactly what needs checking.

Also Notable

• Vision Foundation Models Favor Semantic Invariance and Drop Spatial Detail — this one recenters pretraining on boundary/shape discontinuities to recover dense spatial perception, 35 upvotes with code. Vision Pretraining for Dense Spatial Perception
• KV Cache Eviction Stops Relying on Static Heuristics — KVpop directly supervises the keep-or-drop decision, learning a fixed-budget eviction policy — a practical piece for saving memory in long-context decoding. KVpop: Key-Value Cache Compression with Predictive Online Pruning
• Long Video Is Either Single-View Temporal Autoregression or Multi-View Clips — MV-Forcing packs both kinds of autoregression into one diffusion model via a 4D geometry bridge, generating long, multi-view-consistent dynamic scenes (ECCV). MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing
• Multi-Vector Image-Text Retrieval Keeps Fine-Grained Evidence via Late Interaction, but Storage and Compute Are Costly — SaMer does object-aware token merging, compressing into K representative centroids without losing region-level evidence. Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval
• The First World Model for Highly Dynamic Multi-Agent Environments — conditioned on multiple agents' action streams, it attributes scene changes to the correct player, tested on Rocket League. Multiplayer Interactive World Models with Representation Autoencoders
• End-to-End Driving VLA Tries to Skip the Usual 2D-to-3D Trajectory Prediction — PixelPilot takes the scalable route (ECCV). PixelPilot: Scalable Vision-Language-Action Models for End-to-End Autonomous Driving
• T2I Painting Rare Attribute-Object Combinations Keeps Dropping Concepts or Getting Overrun by Dominant Entities — RADIANCE uses relative adaptive denoising plus an IP-Adapter to reinforce novel concepts. RADIANCE: Relative Adaptive Denoising with IP-Adapter for Novel Concept Enhancement
• How to Govern Failures That Only Surface After Deployment and That Pre-Evaluation Missed — a survey of open problems in AI incident governance, from definition and taxonomy to monitoring, reporting, and incident response (ICML). Open Problems in AI Incident Governance
• Existing Multi-Bit Watermarks Either Expose All Metadata or None — selective-disclosure watermarking lets you reveal only part of the embedded information on demand, adding a control granularity for LLM text provenance (ICML). Selective Disclosure Watermarking for Large Language Models
• Monocular-to-Stereo Video Stalls on Occlusion Inpainting and Lacks Stereo Training Data — Geometric Reciprocity unlocks self-supervision through geometric reciprocity, sidestepping scarce stereo annotations (ICML). Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation

Read the full edition →