What is streaming?

This repository contains the research code for a paper exploring a new approach to how large language models, the AI systems behind tools like ChatGPT, process information. Normally these models work in a single, sequential stream of thought, generating one token of text at a time. This research investigates running multiple parallel streams simultaneously so the model can, for example, continue reading new input while simultaneously solving a problem, or monitor its own reasoning for safety issues while generating a response. The code is organized into three sections matching the paper's experiments. The first looks at whether parallel streams improve efficiency on reasoning and question-answering benchmarks. The second examines whether this approach can make models more resistant to prompt injection attacks, where malicious instructions hidden in documents try to hijack the model's behavior. The third explores whether having dedicated monitoring streams makes the model's internal reasoning more observable and auditable from the outside. This is a research-grade Python codebase aimed at AI researchers who want to reproduce results from the paper or build on the multi-stream training approach.