AdaSR: Adaptive Streaming Reasoning with Hierarchical Relative Policy Optimization

Researchers have introduced AdaSR, an adaptive streaming reasoning framework that allows large language models to reason while input arrives incrementally and then perform a final deliberation once the stream is complete [1]. The work targets dynamic real-world scenarios such as audio and video streams where information is continuous [2]. The framework departs from the standard read-then-think paradigm, in which models observe a complete input before reasoning over a static context [1]. In streaming settings, models must update and respond under partial observations, a challenge that existing streaming reasoning methods address primarily through supervised imitation of pre-constructed trajectories, limiting their flexibility [2]. AdaSR instead learns when to think and how much computation to allocate across different stages of the streaming process [3]. To optimize this hierarchical reasoning process, the authors propose Hierarchical Relative Policy Optimization, or HRPO [1]. HRPO decomposes policy optimization into streaming reasoning and deep reasoning phases, providing more fine-grained advantage assignment instead of uniformly distributing a single sequence-level advantage over all tokens [4]. The algorithm refines Group Relative Policy Optimization with stage-aware advantage assignment, assigning distinct advantages to streaming tokens, deep-reasoning tokens, and the full rollout [3]. This preserves the group-relative optimization while turning the coarse sequence-level advantage into finer-grained, hierarchy-aware credit signals [5]. HRPO integrates three reward types: format, accuracy, and adaptive thinking rewards [1]. The format reward enforces valid reasoning protocols, the accuracy reward preserves final task performance, and the adaptive thinking reward encourages latency-aware computation allocation [2]. Experiments reported in the paper show that AdaSR achieves a better balance among reasoning accuracy, computational efficiency, and streaming latency compared with a supervised fine-tuning baseline [6]. The paper was submitted to arXiv on June 12, 2026, under the Computation and Language category [1]. arXiv, an open-access repository of electronic preprints that is not peer-reviewed, hosts scientific papers across fields including computer science and passed the two-million-article milestone by the end of 2021 [10]. The authors have released their code on GitHub under the EIT-NLP organization [2].