Most popular AI assistants are powered by conventional (autoregressive) language models. They receive a prompt and write text word by word, from left to right, with no ability to go back and correct a mistake. Such a model takes a long time to generate a lengthy response: the text appears sequentially, and an error made at the beginning affects the rest of the text. Deep research systems, which gather information from various sources and produce detailed reports, also run on these same models. In such systems, several AI agents sequentially plan the search, formulate queries, find and analyze materials, and then combine the results into a single document. Depending on the task, preparing such a report can take 15-30 minutes.
In order to accelerate this process, ITMO researchers have opted for a diffusion language model instead of the conventional one typically used in deep research systems. Diffusion models generate texts differently: not word by word, but in parallel in several places. This process can be compared with the static on a TV screen – the image does not appear line by line, but gradually across the whole screen. A diffusion model assembles the text from separate fragments and then goes through several generation cycles, refining previously generated parts of the text at each one. Moreover, while the report is still in the making, the model can change words and correct mistakes made at the beginning. All of this makes generation several times faster.
“To make sure that it was definitely the new model that accelerated the process, we ran an experiment. We took the same system, prompts, code, and evaluation, and changed only the ‘engine’ – either the diffusion or the comparable conventional model. We used Deep Research Bench as our speed benchmark – it’s a collection of 100 PhD-level problems in 22 subject fields. As a result, the diffusion-model based system completed them about six times faster, with planning and writing accelerated most significantly – by 10-12 times,” shares Alan-Barsag Gazzaev, the main developer of DiffResearch, a PhD student at ITMO’s Information Technologies and Programming Faculty.
Alan-Barsag Gazzaev. Credit: Viktoria Gazzaeva
Another advantage of DiffResearch is the option to start searching for information while the AI agent is still forming queries. Regular systems would need to complete the search query and only then submit it to a search engine. A diffusion model can use the already generated parts of the query without the need to wait for it to be generated completely. Thanks to that, the system starts searching for scientific papers and materials earlier, decreasing the time it takes to prepare a query nearly twofold.
This acceleration didn’t cause a drop in response quality. While selecting its sources, the system turns to scientific search engines, just like the conventional models do. DiffResearch was tested using RACE – an automatic evaluation of the response with another language model that assesses whether the topic is fully covered, whether there are informative conclusions, whether the user’s request has been fulfilled, and whether the text is comprehensible. The researchers compared three versions of the same system: one with an autoregressive model, another with a diffusion model, and a third one with a diffusion model in the early response extraction mode. Their average scores were 0.3904, 0.3895, and 0.3969 respectively on a 0-1 scale.
“Diffusion models are a brand-new topic – the first NeurIPS workshop that focused on them was held in 2026. We are the first team in Russia to have used them in deep research. Additionally, we have taught the model to use its drafted search queries before they are done generating, which is a function we haven’t found in any earlier publication,” adds Alan-Barsag Gazzaev.
Such systems will come in handy in situations when you need to quickly master a complex topic. Students and scientists will be able to assemble materials on a particular topic quicker, while businesses will cut down the time needed to produce analytical reports. This acceleration will be particularly useful in medicine, where it’s crucial to quickly receive updates on treatment standards and new studies to select the best approach for each patient.
Alexey Gavrilov, a PhD student at the Information Technologies and Programming Faculty, has also contributed to the development process. The project was headed by Sergey Muravyov, an associate professor at the Institute of Applied Computer Science. In October, the team presented their study in Budapest at EMNLP 2026, one of the world’s most prestigious conferences in natural language processing.
