RISC-V in CPython, fair LLM benchmarks, and agents for databases

In today’s roundup: an important update for Python on RISC-V, reproducibility issues with LLM benchmarks, and practical ways to connect models to complex databases.

🔥 Hot:

🔹 CPython has received official support for the RISC-V architecture — The support is at tier three: builds and maintenance are available, but there are no guaranteed timelines for issue processing. 🔹 A study showed that LLM leaderboard results depend on test harness settings — The order of options, prompt wording, and response parsing method can significantly change the final score.

➡️ Useful materials:

🔹 SchemaRouter proposes selecting agent tools based on data fields — The approach is intended for heterogeneous RAG, where APIs, databases, vector stores, and graph databases are used simultaneously. 🔹 AIREP describes a signed protocol for auditing AI-runtime decisions — For each response, it records whether it was released, blocked, deferred, edited, or escalated. 🔹 A review examines model collapse and ways to reduce its risk — The focus is on the consequences of training subsequent models on synthetic data and available countermeasures. 🔹 CoreSec proposes that data center RCA systems abstain from guessing more often — Instead of unstable aggregation of assessments, the approach uses abstention algebra to work with noisy and incomplete telemetry.

➡️ Discussions and case studies:

🔹 An LLM was connected to a 253-table database and three approaches were compared — A benchmark based on 29 real-world questions evaluated accuracy, cost, and latency; the experiment showed that the errors were not caused by the models alone. 🔹 A LocalLLaMA user reported the first feature branch written entirely by a local model — The author ran Qwen 3.8 27B on a GeForce RTX 4060 Ti 16GB and used it as an assistant for development and knowledge base management.

📝 If you would like to add other news and materials to the list, write in the comments.