SteER: A Framework for Steerable Deep Research with LLMs

A novel framework named SteER (Steerable deEp Research) has been unveiled to overcome the shortcomings of existing deep research systems that utilize large language models (LLMs). Traditional systems often adhere to inflexible workflows characterized by one-time scoping and prolonged autonomous operations, which fail to adapt when user intentions shift. SteER introduces interpretable, mid-process control within extended research workflows. At each critical juncture, it employs a cost-benefit analysis to determine whether to seek user feedback or continue independently. This framework integrates diversity-aware planning with utility signals that incentivize alignment, novelty, and coverage, while also maintaining a dynamic persona model that evolves during the session. The study indicates that SteER surpasses leading open-source and proprietary benchmarks by as much as 22.80% in alignment. The findings are available on arXiv under the identifier 2605.24266.