LLMS

Apple Launches Multilingual Foundation Models for On-Device AI

By

–

27 July 2025 16h45

10. Apple Intelligence Foundation Language Models Apple introduces two multilingual, multimodal foundation models: a 3B-parameter on-device model optimized for Apple silicon and a scalable server model using a novel PT-MoE transformer architecture.

Deep Researcher: Test-Time Diffusion for Report Generation

By

–

27 July 2025 16h45

8. Deep Researcher with Test-Time Diffusion Rather than relying on static inference strategies like CoT or best-of-n sampling, this work frames the report generation process as a diffusion process.

MCPEval: Open-Source LLM Agent Evaluation Framework

By

–

27 July 2025 16h45

9. MCPEval MCPEval is an open-source framework that automates end-to-end evaluation of LLM agents using a standardized Model Context Protocol, eliminating manual benchmarking.

Inverse Scaling in Large Reasoning Models Test-Time Compute

By

–

27 July 2025 16h45

6. Inverse Scaling in Test-Time Compute Presents a systematic study of inverse scaling in large reasoning models, where increasing the test-time compute (i.e., reasoning length) harms rather than helps model performance.

Compute-Optimal Strategies for Many-Shot In-Context Learning

By

–

27 July 2025 16h45

7. Towards Compute-Optimal Many-Shot In-Context Learning Proposes practical strategies for reducing the cost of many-shot in-context learning while preserving or improving performance. https://
arxiv.org/abs/2507.16217

In-Context Learning Without Weight Updates in LLMs

By

–

27 July 2025 16h45

5. Learning without Training This paper provides a theoretical and empirical explanation for how LLMs exhibit in-context learning, the ability to learn from examples in a prompt without weight updates.

Routine: Structured Planning Format for LLM Agent Tool-Calling

By

–

27 July 2025 16h45

4. Structural Planning for LLM Agents This paper introduces Routine, a structured planning format designed to improve the stability and accuracy of LLM agents executing multi-step tool-calling tasks in enterprise settings.

→ View original post on X — @karpathy

Model Convergence Through Distillation and Subliminal Learning Transfer

By

@karpathy

–

27 July 2025 16h12

This model convergence is quite perplexing. Possibly related to recent results on subliminal learning? Basically deeper knowledge correlations transfer when training via distillation. As the amount of data online from LLMs increases, it’s possible this makes them converge to some