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: Welcome to another episode of ResearchPod. Today, we're looking at a new way to help AI models think more effectively. Sam, what's the core challenge this paper is tackling?
Sam: The paper introduces a framework called AdaSR. It addresses a fundamental limitation in how AI models handle information: they usually wait to receive an entire document or question before they start thinking. The authors call this the "read-then-think" approach.
Alex: So this paper is asking whether we can get models to reason in real-time, even before they have the full picture?
Sam: Exactly. In the real world, information often arrives in a continuous stream—like listening to a live speech or watching a video unfold. If a model waits until the very end to start processing, it creates unnecessary delay. AdaSR tries to solve this by enabling what the authors call "streaming reasoning," where the model thinks as it reads, rather than after.
Alex: That sounds like a significant shift. But if the model starts forming conclusions before it has all the facts, doesn't it risk going down the wrong path?
Sam: That is the central technical hurdle. Previous approaches tried to solve this by having the model copy a human's step-by-step reasoning style. But the authors argue that just mimics the surface appearance of good thinking without building genuine flexibility. What they wanted instead was a system that could decide for itself—when to pause and think carefully, when to move on quickly, and how much mental effort to spend at each stage.
Alex: So it's not just about thinking faster. It's about being smart with the "thinking budget." How does the model actually learn to make those decisions?
Sam: They use a training method called Hierarchical Relative Policy Optimization, or HRPO. Think of it like a teacher grading a student's work. Instead of just looking at the final answer, the teacher also grades the rough notes and working-out separately. HRPO does the same thing—it assigns different "credit" to the model's quick streaming thoughts versus its deeper, more deliberate reasoning at the end.
Alex: Why does that separation matter so much?
Sam: In standard AI training, you usually give one single reward signal to the entire sequence of thoughts and outputs. But in a streaming task, that's too blunt. Imagine a student who gets a lucky correct answer despite having messy, confused working-out notes. If you reward the whole thing equally, the student learns that the messy notes were fine—even though they weren't. The authors call this the "cross-phase credit paradox." A correct final result can accidentally reinforce poor reasoning habits from earlier in the process.
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.