Mistral unveiled a new model, and OpenAI shared mathematical results.

Today’s roundup: a fresh release from Mistral, mathematical research from OpenAI, and a couple of studies on making AI agents more reliable.

🔥 Hot:

🔹 Mistral introduced the Le Chonk model: According to the company, the new model competes with the best open models from China. 🔹 OpenAI published mathematical research and the code for its model: The roundup covers work on open mathematical problems.

➡️ Useful materials:

🔹 Researchers proposed checking VLM responses from different perspectives: This approach can detect plausible but incorrect answers without relying on an external evaluator. 🔹 EPOCH proposes managing AI-agent search through evidence verification: The paper examines the risk of mistaking a fragile result for a genuine discovery. 🔹 FluidPD adapts LLM server resources to changing workloads: The system reallocates resources between request-processing stages depending on speed requirements.

➡️ Discussions and case studies:

🔹 A Habr author showed how he uses one neural network to generate code with the help of another: In the article, he publishes the generated text and the prompt.

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