Due to the latest advances in LLM, in-depth research (DR) agents are rapidly gaining popularity in research and the industry. However, most popular public Dr agents do not take into account human thinking and writing processes. They often lack structural steps to support human researchers, such as drafting, searching, and using feedback. Current Dr agents compile test time algorithms and various tools for uncohesive frameworks, highlighting the key needs for dedicated frameworks that can match or exceed human research capabilities. The cognitive process that is not artificially inspired in the current approach can have a gap between how humans are researched and how AI agents handle complex research tasks.
Existing works (e.g., test time scaling) use iteratively improved algorithms, debate mechanisms, hypothetical ranking tournaments, and self-criticism systems to generate research recommendations. Multi-institutional systems utilize planners, coordinators, researchers and journalists to generate detailed responses, while certain frameworks enable feedback integration of human co-pilot modes. The proxy adjustment method focuses on training in enhancing learning capabilities through multitasking goals, subtle adjustments made by components, and enhanced learning to improve search and browsing capabilities. The LLM diffusion model attempts to break the high-quality output of the self-rotating sampling assumption by generating a complete noise draft and iteratively lowering the tokens.
Google researchers introduced the Test Time Diffusion Researcher (TTD-DR), inspired by the iterative nature of human research, inspired by repeated searches, thinking and refining cycles. It conceptualizes the generation of the research report as a diffusion process that begins with the draft, which is an updated outline and evolving foundation to guide the research direction. The draft is iteratively improved through the “transformation” process and is dynamically informed through the search mechanism, which contains external information in each step. This draft-centric design allows reports to reduce information loss during the iterative search process, making reports more timely and coherent. TTD-DR achieves the latest results on benchmarks that require intensive search and multi-hop reasoning.
The TTD-DR framework addresses the limitations of existing DR agents that employ linear or parallel processes. The proposed backbone DR agent consists of three main stages: research plan generation, iterative search and synthesis, and final report generation, each containing unit LLM agent, workflow and agent status. Agents use self-development algorithms to improve performance at each stage, helping them find and retain high-quality environments. The proposed algorithm is inspired by the latest self-evolution work and is implemented in parallel workflows as well as sequential and circular workflows. The algorithm can be applied to all three stages of the agent to improve overall output quality.
In a side-by-side comparison with OpenAI’s in-depth study, the long-term study report generation task had a win rate of 69.1% and 74.5%, while the performance was better than 4.8%, 7.7% and 1.7% on three study datasets with short form base answers. It shows strong performance in assistive and comprehensive automatic scoring scores, especially on long-term research datasets. In addition, for OpenAI in-depth research on long-term research and in-depth consultation, the winning rates of self-evolution algorithms are 60.9% and 59.8%. Correctivity scores show 1.5% and 2.8% enhancements for the HLE dataset, although Gaia’s performance is still lower than 4.4% of Openai DR. Retrieval combined with diffusion leads to in-depth research on OpenAI for all benchmarks, thus achieving huge benefits.
In short, Google proposes TTD-DR, an approach that addresses basic limitations through human-inspired cognitive design. The framework’s approach conceptualizes the generation of research reports as a diffusion process using an updateable skeleton draft guiding the research direction. TTD-DR enhanced by self-evolution algorithms applied to each workflow component ensures high-quality context generation throughout the research process. Furthermore, evaluations demonstrated state-of-the-art performance of TTD-DR in a variety of benchmarks that required intensive searches and multi-hop inference, achieving excellent results in comprehensive long-term research reports and concise multi-hop inference tasks.
Check The paper is here. random View various applications in our tutorial page on AI Agents and Agent AI. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Sajjad Ansari is a final year undergraduate student from IIT Kharagpur. As a technology enthusiast, he delves into the practical application of AI, focusing on understanding AI technology and its real-world impact. He aims to express complex AI concepts in a clear and easy way.
