• The interface design sets the ceiling on spatial reasoning. SpatialClaw changes nothing about the model and skips fine-tuning. It rewrites the agent's action interface into a stateful code kernel and hits 59.9% average accuracy across 20 benchmarks — 11.2 points above recent spatial agents, with consistent gains on six VLM backbones.

• Science automation's last mile. LabVLA plugs a vision-language-action model into a real lab bench, using a simulation data engine to fill the gap left by scarce lab data, and tops every baseline on LabUtopia both in and out of distribution.

• MaxProof trains math proving as three separate skills — generation, verification, repair — then organizes them with population-level test-time scaling. It scores 35/42 on IMO 2025 and 36/42 on USAMO 2026, both past the human gold-medal line.

• Image models finally get interleaved generation. InterleaveThinker bolts a multi-agent pipeline plus single-step RL onto any generator, letting it think and emit text and images in turns, matching Nano Banana and GPT-5.

• 2D supervision is enough to learn 3D motion. VideoMDM uses no 3D ground truth at all. Precise 2D poses from monocular video train a coherent 3D human motion prior that nearly matches MDM trained on full 3D data.

Also Notable

• The Bottleneck in Autonomous Research Is Environment Design, Not Agent Workflow — EurekAgent treats optimizable metrics and the execution environment as the main battleground, echoing recent environment-engineering work.
• The Lever for World-Action Models Is a Semantic Vision-Action Tokenizer, Not Reconstruction Fidelity — RepWAM wants that representation to connect future prediction with robot control.
• Turn the Agent Harness From Hand Engineering Into a Trainable Plugin — HarnessBridge lets the layer that connects to the environment optimize alongside the task.
• Build a Walkable Surround World in Real Time From One Narrow Image — MoVerse separates "world building" from "observation rendering."
• Make Hidden-State Recurrent Latent Reasoning Switchable and Trainable With On-Policy RL — easing the old problem that latent CoT is hard to optimize and interpret.
• Post-Training Quantization Compresses LLMs to Ternary Weights and Low-Bit Activations — TWLA pushes deployment-grade compression to the limit.
• Step-Level Caching for Diffusion Models Drops the Threshold Heuristic — it makes budget-constrained caching decisions against final output quality directly.
• Fix the "Bag of Words" Flaw in CLIP-Style Models — cross-modal masked composition recovers object relations and attribute binding.
• DoorDash in Production: Multi-Agent RL Learns Three-Sided Dispatch Weights From Delayed Market Feedback — a rare production-system case study.
• Detect Hallucinations Under Zero-Information-Source Constraints — no model internals, no external reference, just human-like-criterion probes.

Read the full edition →