A compact pass over the week, with an emphasis on what each item establishes rather than what it was announced as.

Evaluation kept being the story

Meta entered its models in five STEM olympiads, reporting a perfect theory score at the Asian Physics Olympiad. Olympiads are a better instrument than most benchmarks — fresh problems, human grading, no contamination — while remaining a test of well-posed problems with guaranteed solutions. Real result, narrower scope than the headline implies.

Skill entropy was proposed as a way to characterise long-horizon reasoning tasks by the variety of skills they demand rather than by their length. The underlying observation is correct: binary success on a two-hundred-step episode throws away nearly all the information the episode contains.

MerchantBench targets long-term coherence for agents in e-commerce operations — consistency across a long sequence of coupled decisions, rather than correctness within one task. This is the failure mode that most often kills pilots and least often appears in evaluation.

Human preference rankings came in for scrutiny again. They measure a real thing — which model people prefer interacting with — and drift from capability in predictable ways: length, confident phrasing, immediate legibility.

Science applications continue to look strongest

DeepMind announced WeatherNext, reporting a breakthrough in cyclone forecasting. Weather remains the cleanest large-scale test the field has, because the ground truth arrives on schedule and no one negotiates the rubric.

Round-trip consistency for bidirectional diffusion models offers rollout error prediction without ground truth. A model that knows when to distrust itself is worth more operationally than one with slightly better average error.

Systems and practice

A vLLM internals walkthrough drew a large audience, which it should. Serving throughput is decided by continuous batching, KV cache paging, and scheduling — not by model choice, which is where most teams spend their attention.

Turning recurring LLM traces into typed deterministic pipelines was raised as a question and deserves to be a practice. The most reliable part of any agent system is the part that stopped being an agent.

Governance

Third-party cyber evaluations appeared from several directions. Independent evaluation has gone from rare to routine; reporting standards have not followed, and without a shared severity scale or configuration disclosure, findings are hard to compare.

OpenAI announced work with the American Psychological Association on youth mental health guidance. The structural argument is sound — labs can measure behaviour but are poorly placed to set outcome standards. Whether it is substantive depends on access and publication independence, neither yet visible.

Worth reading if you missed them

• Inside vLLM: Anatomy of a High-Throughput LLM Inference System
• Toward Skill-Native LLMs: Skill Entropy
• Towards Physics of Multimodal Pretraining

Full sources for everything above are in the news feed.