• Native Structural Reasoning Puts Prediction And Explanation On One Chain. SciReasoner discretizes structure into citable tokens instead of squashing it into an embedding. It takes 67 SOTAs across 86 benchmarks, and double-blind experts rate its reasoning traces on par with a frontier model 98% of the time.

• Picking A Coding Agent Means Picking Who Won't Steer You Into A Ditch. AgentLens moves evaluation from a single pass/fail bit to the whole trajectory. Where formal verification can't decide, an LLM writes a readable review — and it already runs in a nightly pipeline to catch product regressions.

• A Sparse Memory Bank Fixes Linear Attention's Recall Gap. Sparse Delta Memory scales state capacity by orders of magnitude while touching only a small slice each step. The bet: grow state without growing compute.

• The Hard Part Of Async RL Isn't Throughput — It's Staying Stable After The Speedup. SAO uses single-rollout sampling to cut off-policy bias, trains stably for a thousand steps, beats GRPO, and already ships in the GLM-5.2 (750B-A40B) training pipeline.

Also Notable

• City-Scale Spatial Intelligence Is Starved For Real-Scale Data. WildCity uses an autonomous-vehicle fleet to capture a tens-of-square-kilometer multimodal testbed, putting rendering, simulation, and spatial representation to the test at real city scale.
• Loosen Exploration Without Losing Stability, And Get Both. UP uses "unbounded positive asymmetric optimization" to loosen exploration while keeping training stable, chewing on the same RL post-training stability problem as today's async RL.

Read the full edition →