Junlong Tong, Wenqi Xu, Yingqi Fan, Anhao Zhao, Xuan Lu, Yang Tan, Xiaoyu Shen
4 min
Abstract
Large reasoning models typically follow a read-then-think paradigm: they observe the complete input, reason over a static context, and then produce the answer. Yet many real-world scenarios are inherently dynamic, such as audio and video stream, where information arrives as a continuous stream and models must reason, update, and respond under partial observations. Recent streaming reasoning methods allow models to think while reading, but they largely rely on supervised imitation of pre-constructed trajectories, which limits their flexibility. In this paper, we propose AdaSR, an adaptive streaming reasoning framework that enables models to reason during input streaming and perform final deliberation once the stream is complete, learning when to think, and how much computation to allocate across different stages. To optimize this hierarchical reasoning process, we introduce Hierarchical Relative Policy Optimization (HRPO), which 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. HRPO integrates format, accuracy, and adaptive thinking rewards to enforce valid reasoning protocols, preserve final task performance, and encourage latency-aware computation allocation. Experiments show that AdaSR achieves a better balance among reasoning accuracy, computational efficiency, and streaming latency compared with supervised fine-tuning baseline. We release our code at https://github.com/EIT-NLP/StreamingLLM/tree/main/AdaSR.
Alex: So by grading each phase separately, the model can actually learn which specific parts of its thinking were genuinely useful, and which were just noise?
Sam: Precisely. And there's a second component to HRPO that addresses a different problem. The system includes what the authors call an "adaptive thinking reward." The idea is straightforward: longer outputs generally mean the model is spending more computation. So the reward gently discourages the model from rambling on when the answer is simple, while still giving it room to think at length when the problem genuinely demands it.
Alex: That's a sensible design. Spend effort where it counts, not everywhere equally. What does the evidence say about how well this actually works?
Sam: The experiments suggest AdaSR achieves a better balance between accuracy and response speed compared to earlier methods. By allocating computation more selectively—less time on straightforward or irrelevant input, more time on genuinely complex passages—the model becomes more responsive without sacrificing reliability. The paper frames this as a meaningful step toward AI systems that can handle real-time, dynamic information more practically.
Alex: What are the limits of this approach? Where does it break down?
Sam: That's a fair and important question. The current framework depends on having specific, verifiable correct answers to guide the training process. The system needs a clear "ground truth" to grade itself against. For open-ended tasks—summarising a live conversation, or responding to an unfolding video—there often isn't a single right answer to measure against. So the approach, as it stands, works best in domains where correctness is well-defined.
Alex: So the training wheels are still on, in a sense. It needs a clear target to learn from.
Sam: That's a reasonable way to put it. The authors are candid about this. Extending streaming reasoning to more open-ended or subjective tasks is a genuine open problem, and one they flag as a direction for future work.
Alex: It's a thoughtful piece of research. The core insight—that you need to grade the process, not just the outcome—feels like it could have broader implications for how we think about training AI systems generally. Thanks for walking us through it, Sam.
Sam: Thanks for having me. And thanks to everyone listening to ResearchPod.