一篇论文改写AI科研评价规则！中国公司拿出实践数据，双榜第一

TL;DR - Moderna and Merck report that their personalized mRNA cancer therapy, intismeran autogene plus Keytruda, met recurrence-free and distant-metastasis-free survival endpoints in a Phase III trial for high-risk melanoma. The result could mark a major clinical milestone for individualized neoantigen therapies, though the article’s claim that AI “cured cancer” overstates the reported findings.

• The global randomized, double-blind trial enrolled 1,137 patients with surgically resected stage IIB–IV melanoma and compared the combination therapy with Keytruda alone.
• The workflow sequences each patient’s tumor and healthy tissue, selects up to 34 tumor-specific neoantigens computationally, and encodes them in a personalized mRNA vaccine.
• The vaccine trains T cells to recognize tumor mutations, while Keytruda removes PD-1-mediated immune inhibition; no new unexpected safety signals were reported.
• Moderna and Merck plan regulatory discussions and are testing the platform in additional solid tumors, including lung, pancreatic, breast, bladder, and kidney cancers.

TL;DR - Alibaba Cloud reported 45% year-over-year revenue growth and a 133% increase in adjusted EBITA, suggesting that AI’s clearest near-term monetization is accruing to cloud infrastructure and chip providers. Continued compute shortages, rising MaaS usage, and Alibaba’s vertically integrated stack could sustain growth and improve margins.

• AI training and inference demand drove Alibaba Cloud quarterly revenue to RMB 48.737 billion; Alibaba expects growth to accelerate over the next several quarters.
• Adjusted EBITA margin reached 11.6%, supported by expanding MaaS revenue, infrastructure optimization, and greater use of in-house chips.
• Alibaba says its chip, networking, storage, and scheduling optimizations improved model-output throughput by 2.5%–5.6% without adding hardware.
• Its new T-Head Zhenwu M890-based supernode instances were serving more than 650 customers, strengthening Alibaba’s full-stack position across chips, cloud, models, and applications.

TL;DR - A new “discovery episode” framework evaluates AI scientists on complete research cycles rather than closed-book answers, scoring hypothesis formation, experiment execution, and result interpretation. Deep Principle’s MIRA platform illustrates the approach with an agentic system connected to computation and automated laboratories.

• The framework records and assesses each decision across the scientific workflow, including failures, rigor, reproducibility, and whether findings are genuinely novel.
• MIRA uses collaborating agents, high-performance computing and automated wet-lab execution, plus reusable research memory to support closed-loop experimentation.
• The article reports that MIRA ranked first on Research Claw Benchmark and Science Agent Arena, scoring 81.1% on the latter while also achieving the lowest average task cost.
• Deep Principle says it is applying the platform to internal R&D pipelines in batteries, industrial coolants, and new-energy materials.

TL;DR - Chinese cloud AI chipmaker Enflame has begun its Shanghai STAR Market IPO process, seeking RMB 6 billion to fund fifth- and sixth-generation chips and hardware-software co-design. Its expanding product stack and reported commercial deployments signal growing demand for domestic AI training and inference infrastructure.

• Enflame has developed four architecture generations and five cloud AI chips, spanning accelerators, modules, computing systems, clusters, and programming software.
• Its proprietary stack includes GCU-CARE compute units, GCU-LARE chip interconnects, and the TopsRider software platform for model development and migration.
• The company reports revenue-generating deployments of thousand- and ten-thousand-accelerator computing centers, alongside work on supernode and high-speed interconnect designs.
• IPO proceeds are designated for next-generation chip commercialization and advanced AI hardware-software co-innovation.