Microsoft released MAI-Thinking-1, an MoE model with 35B active parameters and 1T total parameters, pretrained from scratch on 30T tokens without third-party model distillation.
Why it matters: HKR-H/K/R all pass: Microsoft released MAI-Thinking-1 with concrete MoE scale and training-token figures. Benchmarks, access, and pricing are not disclosed, so it stays in the 78–84 band rather than P1.
NTU MARS Lab and collaborators released a 43-page survey on robot world models, covering definitions, architectures, applications, benchmarks, and challenges around action-conditioned consistency, inference efficiency, and physical grounding.
Why it matters: HKR-H and HKR-K pass: the hook is robot world models, and the post cites a 43-page survey with benchmarks and action-consistency framing. HKR-R is weak, so this stays at the featured threshold.
Zhejiang University and Microsoft introduced World-R1, training Wan 2.1 with about 3,000 text-only prompts, Flow-GRPO, and a four-part reward; the 1.3B version improves PSNR over the baseline by 10.23 dB.
Why it matters: HKR-H/K/R all pass: the hook is unusual, and the post gives 3,000 text samples, Flow-GRPO, and a +10.23 dB PSNR gain. Strong multimodal research, but not a foundation-model launch, so 78.
Emory University, Microsoft, and partners introduced STReasoner for spatio-temporal time-series reasoning, with ST-Bench covering four task types. It uses Network SDE plus Multi-Agent data generation, then Align, SFT+CoT, and S-GRPO training. The article claims inference cost is 0.004× closed models, with code on GitHub.
Why it matters: HKR-H and HKR-K pass: the story has a “first framework” hook plus ST-Bench, S-GRPO, 0.004× cost, and code release. HKR-R is weak because spatiotemporal reasoning is a narrower research lane.