The AI Gazette

Aug 17 – Aug 23, 2026

Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference

A new design for small AI language models, built specifically to run fast on your personal computer. Instead of always re-reading the entire conversation history, this model uses a hybrid approach where most of its internal memory stays fixed, making it much quicker as conversations get longer.

Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference · Christos Koutsiaris · arXiv:2608.20210

Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation

This paper teaches AI models to be smarter about their "thinking" process, allowing them to decide whether to give a quick answer, a brief explanation, or a detailed one for each problem. This adaptive approach significantly cuts down on the computational power and time needed, making today's increasingly large AI systems much more efficient and practical.

Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation · Gijs Kassenaar, Zhao Yang, Vincent François-Lavet · arXiv:2608.20256

Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation

A look at how to efficiently assign tasks to the best AI model when figuring out which model is truly optimal for a given task costs money. It uses a classical mathematical framework, "Pandora's Box," to decide whether it's worth investing in a more thorough assessment of a model's capabilities, thereby saving significant computational resources.

Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation · Adam Fisch, Shubhendu Trivedi, Fantine Huot et al. · arXiv:2608.20316

Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

A look at how large language models (LLMs) can best learn and reuse "skills" from past experiences, finding that breaking down learning into smaller steps and using natural language descriptions works much better than learning from whole tasks or using code.

Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents · Yiyang Feng, Biddut Sarker Bijoy, Niranjan Balasubramanian et al. · arXiv:2608.20274

MidTool: Mid-training Data Synthesis for Agentic Tool Use

A new training method, called "mid-training," that specifically teaches AI models to become much better at using external tools, like software programs or web services, to accomplish tasks.

MidTool: Mid-training Data Synthesis for Agentic Tool Use · Fengqing Jiang, Yite Wang, Boyi Liu et al. · arXiv:2608.20314

Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization

A look at how to make large AI models internalize specific knowledge from a set of documents, so they can answer questions accurately without needing to retrieve those documents during conversation. Their three-stage process first teaches the model the document content, then fine-tunes it for question-answering, and finally recovers its broader abilities.

Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization · Qian Kou, Xiaofeng Shi, Xiaosong Qiu et al. · arXiv:2608.20281

Inducing Task Models from Computer-Use Traces

A new way for computers to watch someone use a computer, by recording their clicks and screen activity, and automatically figure out the different tasks they were doing, even when multitasking. It then builds a structured model for each task, detailing both the overall goals and the specific steps taken to achieve them.

Inducing Task Models from Computer-Use Traces · Yucheng Jiang, Zora Zhiruo Wang, Ruishi Chen et al. · arXiv:2608.20319

The Forward-Backward Disconnect: State Dynamics, Credit Assignment, and Biological Grounding in Neural Computation

This paper points out a growing "disconnect" in AI: while our neural networks are becoming incredibly complex and brain-like in how they process information, the methods we use to teach them are still simple and less biologically plausible.

The Forward-Backward Disconnect: State Dynamics, Credit Assignment, and Biological Grounding in Neural Computation · Hadi Al Mubasher, Mariette Awad · arXiv:2608.19995

All Gazettes · Artificial Intelligence Collective