bwr-hhh/tflow

TFlow is the official code release for a research paper about a new way for multiple AI language models to share information with each other. Normally, when several AI agents work together on a problem, one agent passes information to another by writing it out as plain text, which the second agent then has to read as part of its input. TFlow replaces that text passing with something more direct: it takes the internal numerical representation the sending agent produced while thinking, and turns it into a small, temporary adjustment to the receiving agent's own internal parameters. That adjustment only applies while the receiving agent is generating its answer, then disappears, so the underlying model is never permanently changed. The paper's experiments use three copies of a smaller open language model called Qwen3-4B, each acting as an agent. Compared to a single agent working alone, using TFlow to combine information from the other two agents improves accuracy by up to 8.5 percentage points while processing up to about a third fewer tokens. Compared to the usual approach of having agents write text messages to each other, TFlow cuts the number of tokens processed by up to 83 percent and finishes up to 4.6 times faster, while keeping accuracy close on four of the five tasks tested, which cover math word problems, general knowledge questions, and two kinds of coding tasks. This release includes only the code needed to run and evaluate the already trained model, not the training code itself, though the paper describes how the released checkpoint was produced. Setup uses a conda environment with Python 3.10, plus a requirements file and a local package install. The five benchmark datasets download automatically the first time they are used. Users can evaluate the model on all five benchmarks with a single script, or run inference on one question at a time from the command line, a text file, or piped input, saving the prediction and metadata to a JSON file if wanted. The project is aimed at researchers working on multi-agent language model systems, and is released under the Apache 2.0 license.