acceptodds
Under review as a conference paper at ICLR 2027

Learning When to Think While Listening in Large Audio-Language Models

Abstract

Recent advances in Large Audio-Language Models (LALMs) have made real-time, streaming spoken interaction increasingly practical. In this setting, reasoning quality and responsiveness are tightly coupled: reasoning that starts after the speech endpoint remains on the response path, while reasoning under partial audio risks missing decisive late evidence. We introduce a learnable wait-think-answer control formulation for LALMs. Motivated by the incremental nature of human conversation, the controller decides under partial audio evidence when to wait, and when to produce a compact reasoning update. Using Qwen2.5-Omni-7B as the base model, we construct aligned wait-think-answer traces from spoken reasoning data, train the controller with supervised fine-tuning (SFT), and then apply Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO). The reward combines answer correctness, action validity, update timing, post-end residual work, reasoning quality, and chain consistency, optimizing the complete wait-think-answer trajectory and not the final answer alone. On a six-task synthetic spoken reasoning question answering (SRQA) benchmark, the six-reward DAPO controller improves the row-weighted accuracy from 67.6% to 70.3% while reducing post-endpoint residual reasoning from 10.44 to 8.99 tokens under the same Qwen streaming replay protocol. These tokens measure remaining explicit reasoning, not elapsed response time. Native-direct accuracy is essentially unchanged after controller post-training. On a 186-item human-recorded Real Audio Bench, a transfer check beyond text-to-speech (TTS)-rendered speech, the controller family remains functional: SFT achieves the strongest accuracy, while the six-reward DAPO controller is the only learned variant whose post-end residual falls below the base. These results show that, in the endpoint-gated Qwen setting, a streaming model can learn when to make intermediate reasoning explicit during the audio stream.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large audio-language models, omni model, streaming speech reasoning, multimodal reasoning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_Jt78OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed the idea to incorporate stateful controls into speech LLMs. The states control the models when to wait, when to think, and when to answer. Furthermore, the author proposed 6 types of rewards to optimize the performance. The model is evaluated on several spoken reasoning question-answering tasks. With DAPO with 6 rewards the model gets improvement on average accuracy. They also evaluate their method on a real human recording of spoken question reasoning.

Critical strengths
  1. The novel design of integrating states prediction in the streaming speech LLMs
  2. The design of applying DAPO to improve the model to learn better state control as well as reasoning.
  3. Provide analysis on human-recorded real audio of spoken reasoning QA.
Critical weaknesses
  1. The improvement on the main results are marginal (Synthetic SRQA, Table3).

    The accuracies of all variants of the proposed methods are very close (<5%). With a basic controller, the average accuracy is already 67.6%. With all 6 rewards by DAPO, the accuracy is 70.3%. Comparing with the latency, in table 9 (correct me if this is not the right table to refer to), the latency (numbers of tokens after question finish) is almost the same. The maximum gap between the best and the worse settings is smaller than 2 tokens.

  2. The performance differences on Table 4 (real audio) is mixed and the differences are marginal.

    In table 4, DAPO is only good at decreasing the latency but the accuracy also falls. Furthermore, the accuracies and latencies differences are marginal. Gap of Acc. is still < 5% and Latency is still < 2 tokens.

Questions for authors
  1. What is "Base controller"? I cannot find a formal definition in the paper.
  2. It is weird that the latencies are not shown together with the accuracies in Table 3. Reader have to look into the Appendix for table 9.
  3. I think it would be better to emphasize Qwen2.5-Omni-7B as a reference (top line for accuracy, baseline for latency) in Table 3. Shih et al. is using Moshi as their backbone, so there isn't too much to compare with them.
  4. It's better to show the confidence interval in the tables given that the numbers are already very close.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Jt78

We sincerely thank the reviewer for recognizing the stateful control design, the use of DAPO to learn state control and reasoning together, and the analysis on human-recorded audio. We adopt all four presentation suggestions (Q1–Q4) and then address the size of the gains (W1, W2).


Q1. What is the "Base controller"?

The Base controller is the unmodified Qwen2.5-Omni-7B, with no adapter, run under the same streaming protocol as the trained controllers: the same prompts, 0.5 s decision grid, action parser, token caps and scorer. Only the weights differ, which is what makes Base→DAPO a matched comparison. The submission uses the term in its tables without defining it; we will add this definition to Section 4.

Q2, Q3, Q4. Table 3: latency beside accuracy, Qwen as the reference, confidence intervals

We agree with all three. The post-end tokens, the Qwen reference and the Real Audio intervals are in the submission but split between the main text and the appendix (Table 9); the intervals for the main benchmark are new. We will bring them together in Table 3, with Qwen2.5-Omni-7B as the reference row and the Moshi rows moved to the appendix:

Protocol and checkpointAccuracy (95% CI)Post-end tokensMean RTF
Qwen2.5-Omni-7B, native direct (reference)71.85%––
Complete-audio reasoning, Base70.8%10.27–
Streaming Base67.6% (66.6–68.5)10.441.137
Streaming SFT66.1% (65.1–67.0)9.821.353
Streaming DAPO (six rewards)70.3% (69.4–71.3)8.991.134

W1. Size of the gains on the main benchmark

The gain is small in absolute points and large relative to the gap it addresses: DAPO closes 64% of the gap between the streaming Base and the Qwen reference, and the result is statistically reliable.

  • Accuracy. +2.7 points on 8,959 items is 245 more questions answered correctly (6,299 against 6,054). The Base and DAPO intervals above do not overlap, and the exact McNemar bound is p ≤ .0011 against Base and p ≤ 5.2×10⁻⁷ against SFT. (Intervals are normal-approximation binomial intervals over items, for fixed checkpoints and scorer; they do not reflect training-seed variation.)
  • Against the reference you suggest. The distance from streaming Base (67.6%) to native Qwen (71.85%) is 4.25 points. DAPO closes 2.7 of them, 64% of the observed gap to that reference.
  • Across tasks. DAPO improves five of the six task families over Base: +1.8 (ARC-E), +0.9 (ARC-C), +2.4 (SIQA), +5.7 (PIQA) and +3.1 (GSM8K); LLaMA-QS is unchanged.
  • Latency. "Fewer than 2 tokens" is a 13.9% reduction of the reasoning left after the user stops speaking (10.44→8.99), obtained together with the accuracy gain. Accuracy and post-endpoint work normally trade against each other; here both improve. Reported replay RTF is 1.137 for Base and 1.134 for DAPO.
  • Relative to its starting point. DAPO is initialized from SFT. Against that initialization it gains 4.2 points and cuts the residual by 8.5%, with reported mean replay RTF 16.2% lower.

W2. Real Audio (Table 4)

Real Audio Bench is a 186-item transfer check on human speech. Every learned controller is above Base on it; SFT is the most accurate, and six-reward DAPO is not the best configuration on this set.

ControllerAccuracy95% CI
Base64.0%[57.0, 71.0]
SFT68.8%[61.8, 75.3]
DAPO, 5 rewards67.7%[60.8, 74.2]
DAPO, 6 rewards65.1%[58.1, 72.0]
  • What the table shows. All three learned controllers are above Base, which supports transfer from synthetic to human speech. One item is 0.54 points here: SFT is 9 items above Base, and five- and six-reward DAPO are 2 and 7 items below SFT.
  • Evidence on the larger SRQA benchmark. On the 8,959-item benchmark the same DAPO-against-SFT comparison is 70.3% against 66.1%, with p ≤ 5.2×10⁻⁷.
  • Reproducibility. Our rerun for this response reproduces the reported six-reward DAPO accuracy exactly (121/186).

Summary of revisions

  • Formal definition of the Base controller in Section 4 (Q1).
  • Table 3 with post-end tokens and RTF beside accuracy, Qwen2.5-Omni-7B as the reference, and Moshi rows moved to the appendix (Q2, Q3).
  • Confidence intervals in Tables 3 and 4 (Q4).

We hope these clarifications and revisions address the concerns, and respectfully ask the reviewer to reconsider the assessment.

AuthorsOpenReview ↗

Revised PDF uploaded

Dear Reviewer Jt78,

We have uploaded the revised paper; all changes are marked in blue. The items from our response are in place:

  • Q1. The Base controller is defined in Section 4.
  • Q2, Q3. Table 3 now shows post-end tokens and RTF beside accuracy and uses native Qwen2.5-Omni-7B as the reference; the literature Moshi rows moved to Appendix D.4.
  • Q4. Confidence intervals for the main benchmark (Table 3) and for Real Audio Bench (Table 4).

We would be glad to answer any remaining questions. If the response and the revision address your concerns, we would be grateful if you would consider updating your assessment.

I appreciate the authors' effort for the explanation.

They addressed most of my concerns in the question section. For Q4, the author only shows the confidence interval for the accuracies. Could they also provide the confidence intervals for the latencies (Post-end tokens)? Given that the numbers are also very close across the setups.

Regarding the weakness section: From the results of Table 3 and Table 4 (Real Audio), we see that DAPO variants are only outperforming SFT on Table 3 but not Table 4. I wonder what the explanation for this is. Does it mean that DAPO is optimized on the synthetic data and degrades the model's generalization ability to real audio?

AuthorsOpenReview ↗

Response to Reviewer Jt78's follow-up

We thank the reviewer for the follow-up and are glad that most of the questions are resolved. We answer the two remaining points below.

Confidence intervals for post-end tokens

For the streaming Base, the 95% interval is ±0.13 tokens on the main benchmark and ±0.55 tokens on Real Audio Bench.

  • Main benchmark. For Base, the six-task macro-average post-end length is 10.44 tokens, with a 95% task-stratified bootstrap interval of [10.31, 10.58] (10,000 resamples, items resampled within each task).
  • Real Audio Bench (186 items). Here the interval is wider: 6.52 for Base with [5.99, 7.09]. Six-reward DAPO has the lowest observed mean in Table 4 (6.33, against 6.52 for Base and 6.64 for SFT), and we do not claim a significant residual difference among the three on this set.

Intervals for the remaining rows, with paired intervals for the differences, will be added in the final version using the same procedure.

Why DAPO outperforms SFT in Table 3 but not in Table 4

Every DAPO variant has higher observed accuracy than Base on human recordings. What does not carry over to this set is DAPO's additional gain over SFT.

Real Audio Bench (186 items)CorrectDifference from BaseDifference from SFT
Streaming Base119––
SFT128+4.8 [+1.6, +8.6]–
DAPO, 4 rewards122+1.6 [−1.6, +5.4]−3.2 [−7.0, +0.5]
DAPO, 5 rewards126+3.8 [0.0, +8.1]−1.1 [−4.8, +2.7]
DAPO, 6 rewards121+1.1 [−2.7, +4.8]−3.8 [−8.1, 0.0]

Differences are in accuracy points, with 95% paired-bootstrap intervals (10,000 resamples of the 186 items).

  • The comparison does not isolate robustness to human speech. SFT and DAPO are both trained on synthesized speech only. Real Audio Bench changes the acoustic domain and the item distribution at the same time, so a difference between the two stages on this set cannot be attributed to human audio alone.
  • The differences on 186 items are small. Relative to SFT, the DAPO variants differ by 2 to 7 items. No significant difference is detected for the four- and five-reward variants; for the six-reward variant the interval ends at zero. Relative to Base, the three variants are 3, 7 and 2 items higher.
  • One plausible explanation is specialization. DAPO optimizes the reward on its own training distribution, which includes a small share (below 2%) of training-split examples in the benchmark task formats. Its gain over SFT is largest on those formats: 4.2 points in Table 3 (p ≤ 5.2×10⁻⁷). The results show that this gain does not transfer uniformly to a different evaluation distribution. We have not tested this explanation directly.
  • This is how the paper states it. Section 5.2 says that SFT has the highest observed accuracy on this set and that six-reward DAPO is not the most accurate configuration there; the primary claim rests on the 8,959-item benchmark.

We hope this resolves the remaining points, and would be grateful if the reviewer would consider updating the assessment.

Reviewer_juxHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tries to improve a large audio LLM ability to answer questions in a faster and more dynamic approach during streaming. Essentially, it trains the model (Qwen-Omni model) to wait, think, or answer the question. The authors synthesized the training data through TTS and gathered the ground truth from GPT-4o annotations, then used LoRA SFT followed by DAPO with six rewards. The authors claim results on a synthetic speech reasoning QA benchmark improved from 67.6 to 70.3 in streaming setting, and residual tokens dropped from 10.44 to 8.99, which is supposed to reduce latency.

Critical strengths

The research question proposed by this paper is useful, and the overall end-to-end engineering work is pretty complete. The statistical analysis is very rigorous (McNemar tests, bootstrap CIs, and a judge-agreement audit). The authors also honestly reported negative results, such as native direct answers performing better, with no observed improvement on real speech data.

Critical weaknesses

The key idea for this paper was to reduce latency, because the accuracy after the audio LLM reads the full sentence is higher than streaming accuracy. However, the paper never really tests the real impact of the wall-clock latency. That is the biggest miss to me for this paper. Second, even for the streaming accuracy reported, I don't see a good baseline. For example, in the original paper they refer to, I believe they use fixed-interval thinking. That could easily serve as a baseline. Third, I still see a risk of reward hacking. The evaluation and the reward are both using Qwen 3.6 for some of the rewards. Finally, even from a controller training perspective, the paper does not provide any stats on what the improvement is from their training (for example, the stats distribution on wait, think, or answer). I also don't get a clear picture of how this work helps the model react to different traffic at different times.

Questions for authors

1.Can you report end-to-end TTFT or wall-clock latency? 2. How do fixed-interval policy perform on the same backbone? 3. What is the distribution of think timings relative to the arrival of key information?

3 comments
AuthorsOpenReview ↗

Response to Reviewer juxH (Part 1 of 2): Q1 and Q2

We sincerely thank the reviewer for recognizing the research question, the completeness of the engineering, the statistical rigor, and the reporting of negative results. Part 1 answers Q1 (wall-clock) and Q2 (fixed-interval baseline). Part 2 answers Q3 (controller statistics), reward hacking, and traffic.


Q1. End-to-end TTFT or wall-clock latency

The paper's claim is about how much explicit reasoning remains after the user stops speaking, and it measures exactly that. The abstract states the boundary: "These tokens measure remaining explicit reasoning, not elapsed response time." The wall-clock cost of the whole run is reported as replay RTF.

ControllerAccuracyPost-end tokensMean RTF
Streaming Base67.6%10.441.137
SFT initialization66.1%9.821.353
DAPO (six rewards)70.3%8.991.134
  • The metric follows the paper that defined this setting. Can Speech LLMs Think While Listening? (Shih et al., ICLR 2026) defines latency as the number of tokens between the end of the user's question and the start of the system's response, and reports that count, not seconds. We adopt its SRQA evaluation and measure the same boundary on Qwen, where those tokens are the final think: 10.44→8.99, a 13.9% reduction.
  • Why a token count. In Moshi every step lasts a fixed 80 ms, so their count maps to time. A Qwen text token has no fixed duration; it depends on the serving stack, cache, batching and hardware. Our evaluation harness also re-encodes the full audio prefix on every call instead of reusing a cache, which adds repeated-encoding overhead to any endpoint-to-answer time measured in it. We therefore report the token count, which is determined by the policy. The current results validate the accuracy and residual-reasoning improvement under a matched streaming protocol; direct endpoint-to-first-answer-token timing on a cache-native stack would further establish deployment utility, and the limitations section lists it as the next step.
  • End-to-end wall-clock. Replay RTF (Appendix B.1) is the total wall-clock time over every pre-endpoint call and thought, the final think and the answer, divided by audio duration. In the reported measurements it is 1.137 for Base, 1.353 for SFT and 1.134 for DAPO. Residual tokens, replay RTF and endpoint-to-answer time are different quantities; the paper's claims concern the first two.
  • What is claimed. Under the same streaming protocol, post-training improves accuracy and shortens post-endpoint reasoning together. This matched-protocol result is what the reported quantities support. Full-sentence native direct answering is reported as the reference operating point (71.85%); it has no reasoning span, and DAPO closes 64% of the gap from streaming Base to it. On reasoning-heavy GSM8K the controller is above the same checkpoint answering directly (25.9% against 21.08%).

Q2. Fixed-interval policy on the same backbone

We ran it for this response, on the same backbone, harness and scorer, for both Base and DAPO weights. The learned controller beats fixed-interval thinking.

Base / DAPOReal Audio (186)GSM8K (300)LLaMA-QS (300)
Accuracy (%), fixed interval (think every 2.0 s)57.0 / 59.118.3 / 17.070.3 / 71.3
Accuracy (%), learned controller64.5 / 65.121.3 / 22.771.0 / 71.3
Early thoughts per item, fixed interval6.76 / 6.7011.66 / 10.831.10 / 1.10
Early thoughts per item, learned controller7.31 / 1.090.83 / 0.070.18 / 0.00
  • Accuracy. Pooled over the three sets, the learned DAPO controller is 3.6 points above the fixed interval (exact paired McNemar p = .015); Base's own schedule is 3.2 points above it (p = .025). On Real Audio the gap is 5.9 points for DAPO (p = .043) and 7.5 for Base (p = .004); on GSM8K it is 5.7 points for DAPO (p = .024).
  • More thoughts, lower accuracy. The fixed interval writes six times as many thoughts as the trained controller on Real Audio, at lower accuracy. It does shorten the final think (DAPO on GSM8K: 12.8 → 4.8 tokens), which shows that the residual alone can be lowered trivially and why the paper reports it only together with accuracy. The control changes the number and content of thoughts as well as their timing, so it compares scheduling strategies as a whole.
  • Short questions. LLaMA-QS questions last about 3 s. Accuracy is the same under both schedules there, although the behavior differs (1.10 against 0.00 early thoughts per item for DAPO).

(GSM8K: fixed random 300-item sample.)

AuthorsOpenReview ↗

Response to Reviewer juxH (Part 2 of 2): Q3, reward hacking, traffic and revisions

Q3. What training changes: action statistics and think timing

Post-training makes the controller think far less often, and later in the utterance.

Base → DAPOReal Audio (186)GSM8K (300)LLaMA-QS (300)
Wait decisions per item22.3 → 28.548.4 → 49.26.4 → 6.5
Early thoughts per item7.31 → 1.090.83 → 0.070.18 → 0.00
Items with no early thought0% → 15.6%40.0% → 93.0%87.7% → 100%
Median position of the first thought, items with a thought (share of utterance)0.31 → 0.620.59 → 0.650.74 → –

Every item ends with one final think and one answer, by protocol.

  • Fewer thoughts. On Real Audio the untrained controller writes 7.31 early thoughts per item and often repeats the same state; the trained controller writes 1.09.
  • Later thoughts. On Real Audio the median position of the first thought moves from 31% to 62% of the utterance.
  • Relative to key information. The benchmark items carry no annotation of when key information arrives, so we report position within the utterance. On the constructed training data, where evidence-update points are aligned to the audio, R_u is defined as the precision and recall of think placement against those points. It is the supervision designed for evidence-aligned timing; we do not isolate it as the cause of the shift above.

Reward hacking and the shared judge

Three of the six reward terms are fully deterministic, the evaluation scorer applies deterministic rules first, and the schedule comparisons above hold with no judge at all.

  • Deterministic terms. R_f, R_s and R_u are computed by rule.
  • Rule first, judge as fallback. R_a applies task rules and uses the judge only for semantic equivalence.
  • Judged terms. Only R_t and R_c are judge-scored, with prompts and criteria different from the evaluation fallback.
  • Eligibility for the compactness bonus. R_s pays it only for a correct answer supported by the final state.
  • Judge sensitivity. Re-scoring 2,916 decisions with a different model (Qwen2.5-72B) gives 99.11% agreement (κ = 0.978).
  • No judge. Under rule-only scoring the accuracies in Q2 are 0.3 to 3.2 points lower and the schedule comparison holds: learned controller against fixed interval, p = .0015 for Base on Real Audio and p = .024 for DAPO on GSM8K.
  • Answer-only ability is preserved. Native-direct accuracy is the same before and after post-training (71.85% vs 71.89%, McNemar p = .731).

Reacting to different traffic

The controller decides at every 0.5 s tick from the audio received so far, so its behavior follows the incoming speech. The statistics above show how it varies with the input: no early thought on 3-second factual questions and about one on 15-second multi-step speech. When a call overruns a tick, the harness skips the stale decision points and resumes at the current audio position, which avoids a backlog of outdated decisions. We have not evaluated varying input rates or compute load.


Summary of revisions

  • Table 3: post-end tokens and RTF next to accuracy, with the metric's definition and source in the caption.
  • The fixed-interval baseline added to the main results.
  • The action statistics and think-position table added to the paper.
  • Rule-only scores reported beside the judge-assisted ones.

We hope these clarifications and the added baseline address the concerns, and respectfully ask the reviewer to reconsider the assessment.

AuthorsOpenReview ↗

Revised PDF uploaded

Dear Reviewer juxH,

We have uploaded the revised paper; all changes are marked in blue. The items from our response are in place:

  • Metric. Post-end tokens and RTF beside accuracy, with the definition and its source in the caption (Table 3).
  • Fixed-interval baseline. Added to the main results for both Base and DAPO (Table 5 in Section 5.3).
  • Q3. Action counts and first-thought positions (Table 5, Appendix D.6).
  • Scoring. Rule-only scores beside judge-assisted scores (Appendix D.6).
  • Stale ticks. Handling described in Section 4; varying input rates are listed as untested in the limitations.

We would be glad to answer any remaining questions. If the response and the revision address your concerns, we would be grateful if you would consider updating your assessment.

Reviewer_VeygOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a learnable wait-think-answer controller for streaming speech reasoning in LALMs, enabling models to decide when to wait for additional audio evidence or generate intermediate reasoning updates. Built on Qwen2.5-Omni-7B, the approach uses approximately 75K synthetic spoken reasoning examples for supervised fine-tuning, followed by DAPO optimization with six reward components targeting correctness, reasoning quality, timing, and consistency. Evaluated on six synthetic spoken QA tasks with 8,959 samples and a human-recorded benchmark of186 samples, the proposed controller improves synthetic streaming accuracy from 67.6% to 70.3% while reducing post-endpoint reasoning tokens from 10.44 to 8.99.

Critical strengths
  • The paper addresses an important problem for real-time LALMs: deciding when to perform intermediate reasoning under incomplete audio evidence.
  • The experiments show improvements in accuracy (67.6% to 70.3%) while reducing post-endpoint reasoning tokens (10.44 to 8.99).
  • Reward ablations and evaluations on both synthetic and human-recorded audio are provided well.
Critical weaknesses
  • Limited evidence of actual latency improvement: The paper measures the number of reasoning tokens after the audio ends rather than actual response time.
  • The model is trained to improve both reasoning quality and reasoning timing; it would be better to clarify and ablate which contributes to the accuracy improvement.
  • Limited generalization to real-world audio: On the human-recorded benchmark, the six-reward DAPO model performs worse than SFT (65.1% vs. 68.8%) and shows only a small reduction in post-endpoint reasoning tokens.
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Veyg

We sincerely thank the reviewer for recognizing the importance of the problem, the joint improvement in accuracy and post-endpoint reasoning, and the ablations on synthetic and human audio.


W1. Latency: tokens after the endpoint versus response time

Post-end residual tokens are the quantity this paper studies: how much explicit reasoning is still left when the user stops speaking. The wall-clock cost of the whole run is reported separately, as replay RTF.

ControllerAccuracyPost-end tokensMean RTF
Streaming Base67.6%10.441.137
SFT initialization66.1%9.821.353
DAPO (six rewards)70.3%8.991.134
  • The metric follows the paper that defined this setting. Can Speech LLMs Think While Listening? (Shih et al., ICLR 2026) studies the same question on Moshi and defines latency as the number of tokens between the end of the user's question and the start of the system's response. That work reports this token count, not seconds. We adopt its SRQA evaluation and measure the same boundary on Qwen, where those tokens are the final think: 10.44→8.99, a 13.9% reduction.
  • Why a token count. In Moshi every step lasts a fixed 80 ms, so their count maps to time. A Qwen text token has no fixed duration; it depends on the serving stack, cache, batching and hardware.
  • Wall-clock is also reported. Replay RTF (Appendix B.1) is the total wall-clock time of the streaming run over audio duration. In the reported measurements it is 1.137 for Base, 1.353 for SFT and 1.134 for DAPO. Residual tokens, replay RTF and endpoint-to-answer time are different quantities, and the paper claims the first two.

W2. Reasoning quality versus reasoning timing

Two experiments give complementary evidence: the reward ablation shows what the quality terms add, and a new inference-time control shows that the learned schedule beats a fixed one.

  • Reward ablation. Adding the thought-quality term to the four-reward controller raises accuracy from 68.5% to 69.2% with a similar residual (10.87→10.94); adding chain consistency gives 70.3% and 8.99.
  • New schedule control. We held the checkpoint, harness and scorer fixed and replaced the policy's own schedule with a forced think every 2.0 s:
Base / DAPOReal Audio (186)GSM8K (300)LLaMA-QS (300)
Accuracy (%), forced think every 2.0 s57.0 / 59.118.3 / 17.070.3 / 71.3
Accuracy (%), policy's own timing64.5 / 65.121.3 / 22.771.0 / 71.3
Early thoughts per item, forced6.76 / 6.7011.66 / 10.831.10 / 1.10
Early thoughts per item, own timing7.31 / 1.090.83 / 0.070.18 / 0.00

With the weights unchanged, the policy's own schedule is 3.6 points above the forced one pooled over the three sets for DAPO (exact paired McNemar p = .015) and 3.2 for Base (p = .025); on Real Audio the gap is 5.9 points (p = .043). The forced schedule writes six times as many thoughts on Real Audio and is less accurate. The control changes the number and content of thoughts as well as their timing, so it compares scheduling strategies as a whole.

  • What post-training changes. On Real Audio the trained policy writes 1.09 early thoughts per item against 7.31 for Base, and, among items with a thought, the median position of the first one is 62% of the utterance against 31%.

(New runs; GSM8K is a fixed random 300-item sample, and Base on Real Audio is 120/186 here against 119 in the paper's table below. The full control table is in our response to Reviewer ruEA.)

W3. Real audio

Real Audio Bench is a 186-item transfer check on human speech. Every learned controller is above Base on it; SFT is the most accurate, and six-reward DAPO is not the best configuration on this set.

ControllerAccuracy95% CI
Base64.0%[57.0, 71.0]
SFT68.8%[61.8, 75.3]
DAPO, 5 rewards67.7%[60.8, 74.2]
DAPO, 6 rewards65.1%[58.1, 72.0]
  • What the table shows. All three learned controllers are above Base, which supports transfer from synthetic to human speech. Among them the order is SFT, five-reward DAPO, six-reward DAPO, with 2 and 7 items separating them from SFT (one item is 0.54 points).
  • Evidence on the larger SRQA benchmark. On 8,959 items the DAPO-against-SFT comparison is 70.3% against 66.1%, with an exact McNemar bound of p ≤ 5.2×10⁻⁷.
  • Reproducibility. Our rerun reproduces the reported six-reward DAPO accuracy exactly (121/186).

Summary of revisions

  • Table 3: post-end tokens and RTF next to accuracy, with the metric's definition in the caption (W1).
  • The schedule control and thought statistics, next to the reward ablation (W2).
  • Confidence intervals in the Real Audio table (W3).

We hope these clarifications and the added evidence address the concerns, and respectfully ask the reviewer to reconsider the assessment.

AuthorsOpenReview ↗

Revised PDF uploaded

Dear Reviewer Veyg,

We have uploaded the revised paper; all changes are marked in blue. The items from our response are in place:

  • W1. Post-end tokens and RTF beside accuracy, with the metric defined in the caption (Table 3).
  • W2. Schedule controls and thought statistics (Table 5 in Section 5.3), linked from the reward-ablation discussion in Section 5.1.
  • W3. Confidence intervals for Real Audio Bench (Table 4, Appendix D.10) and for the main benchmark (Table 3).

We would be glad to answer any remaining questions. If the response and the revision address your concerns, we would be grateful if you would consider updating your assessment.

Reviewer_ruEAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies when a large audio-language model should produce intermediate reasoning while getting sound signals. It introduces a wait–think–answer controller that observes the available audio prefix and accumulated reasoning states. Answers are permitted only after a supplied speech endpoint. The authors construct aligned synthetic controller trajectories, fine-tune Qwen2.5-Omni-7B, and apply DAPO-style reinforcement learning with rewards for correctness, protocol validity, timing, residual reasoning length, thought quality, and consistency. On 8,959 synthetic spoken reasoning questions, the six-reward controller improves accuracy over the matched streaming Base from 67.6% to 70.3% and reduces post-endpoint reasoning from 10.44 to 8.99 tokens. A 186-item human-recorded benchmark provides a limited transfer check. Native-direct accuracy remains approximately unchanged. The evidence concerns accuracy and explicit residual reasoning under full-prefix replay, it does not establish improved response latency.

Critical strengths
  1. A concrete formulation of a relevant control problem. The explicit WAIT/THINK interface separates acquiring additional evidence from committing an intermediate reasoning state. The aligned data pipeline and trajectory rewards provide a practical starting point for studying this problem, particularly when late-arriving evidence can invalidate an early interpretation.

  2. Useful within-backbone controls and appropriately restrained reporting. The paper evaluates Base, SFT, and RL controllers under a common streaming protocol, includes native-direct evaluation, and adds human-recorded speech. The synthetic accuracy improvement is supported by the reported item-level significance bounds. The explicit acknowledgment that runtime is essentially unchanged helps delimit the demonstrated contribution.

Critical weaknesses
  1. Causal input isolation is not established, and the public rollout implementation exposes future question content. In public commit c7582fcd, build_omni_chunk_prompt inserts a question stem under “Question focus” regardless of question_visible=False. The rollout passes sample.question into this builder and sends the resulting text to the model alongside the audio prefix [1, 2]. A minimal reproduction of this prompt builder shows that a late correction appears in the first-step prompt despite text visibility being disabled. The prompt also exposes the total number of audio chunks before the endpoint [1]. These inputs exceed the observation specified in Equation (1) and could undermine the central claim of reasoning under partial audio evidence. The inspected release predates the attached submission, so I cannot conclude that the reported evaluation used this path. Nevertheless, establishing the exact experimental provenance and causal isolation is essential.

  2. The policy-optimization objective contains a substantive clipping error and differs from the released training configuration. Equation (4) uses max⁡{rA,clip(r)A}\max\{rA,\mathrm{clip}(r)A\} for negative advantages, whereas the DAPO clipped surrogate uses min⁡{rA,clip(r)A}\min\{rA,\mathrm{clip}(r)A\} for both signs [3, Eq. (8)]. For example, with A=−1A=-1, r=0.5r=0.5, and lower clipping bound 0.80.8, the paper's surrogate is −0.5-0.5, while the standard surrogate is −0.8-0.8. The published expression removes the intended protection against excessive probability reduction. The released loss implements the same branch, so this is not merely an isolated notation concern [4]. Furthermore, the implementation averages within turns rather than globally over completion tokens [4, 5], and the paper launcher enables mixed local/outcome advantages despite Section 3.4 describing a shared trajectory advantage [6, 7]. These differences affect what algorithm was evaluated and require reconciliation.

References

All implementation references below point to the same fixed public commit, c7582fcd3e6d39033e3e11c25b1da818edaa8bb9.

[1] Public implementation, prompts/policy_prompts.py, lines 82–97. Pre-endpoint prompt construction.

[2] Public implementation, training/policy.py, lines 1548–1587. Construction and transmission of the policy prompt alongside prefix audio.

[3] Qiying Yu et al. (2025). DAPO: An Open-Source LLM Reinforcement Learning System at Scale. arXiv:2503.14476v2, Equation (8).

[4] Public implementation, training/dapo_loss.py, lines 37–45. Clipping branch and within-turn token averaging.

[5] Public implementation, training/omni_updater.py, lines 1686–1689. Loss aggregation and normalization.

[6] Public implementation, scripts/train_dapo_paper.sh, lines 85–88. Paper launcher configuration.

[7] Public implementation, training/omni_updater.py, lines 1377–1388. Mixing of outcome and local process advantages.

Questions for authors
  1. Clarification of experimental provenance and information access: Which code revision produced the reported checkpoints and evaluation results? Did any pre-endpoint policy prompt contain the complete question stem or total utterance length? Please provide representative serialized prompts and explain whether changing only an unheard suffix leaves the preceding policy inputs unchanged. If the released leakage path was used, corrected training and evaluation would materially affect my assessment.

  2. Clarification of the actual optimization: Was Equation (4)'s negative-advantage branch used during training? Please specify the implemented clipping expression, token/turn normalization, KL term, and credit-assignment configuration. If these differ from the manuscript or require correction, how do the results change?

  3. Evidence for the controller's specific benefit: Do checkpoint- and budget-matched results exist for forced all-WAIT, fixed periodic updates, and an external completeness/semantic trigger? These comparisons would distinguish useful learned timing from improved reasoning content or length regularization. This matters because native-direct accuracy exceeds streaming DAPO accuracy (Table 9), while the reported Base–DAPO runtime difference is negligible (Table 8).

  4. Clarification of evaluation independence: What cross-split exact and semantic overlap checks were performed between generated training data, benchmark-derived training examples, and evaluation items? How much of the accuracy improvement remains under deterministic-only scoring or blinded human adjudication? Reusing the reward judge for evaluation raises an independence concern (Section 4 and Appendix B.3), although it does not itself establish test-set contamination. The second-judge diagnostic in Appendix A.1 helps assess judge sensitivity but does not establish human validity.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ruEA (Part 1 of 2): Q1 and Q2

We sincerely thank the reviewer for the careful review of the implementation and for recognizing the formulation, the within-backbone controls and the restrained reporting. Part 1 answers Q1 and Q2; Part 2 answers Q3 and Q4.


Q1 / W1. Provenance and information access

Which code revision produced the results? Did any pre-endpoint prompt contain the question stem or total utterance length? Does changing only an unheard suffix leave the preceding inputs unchanged?

No question text is injected into any pre-endpoint prompt in the logged training run or in our evaluation runs. The total length N is shown, and removing it produces no significant change.

  • Code revision. The results come from our research revision, not the public export. There the stem is gated: stem = _question_stem_for_think(question) if question_visible else "", with a unit test asserting that audio-only mode hides the question. The gate entered the research repository on 17 April 2026; the training run that produced the reported checkpoint (step 484) is dated 4 May 2026. The logged prompts below show what that run sent. The public commit you inspected is a separate export that lost this gate. Your reading of it is correct, and we have corrected it.
  • Logged prompts. That run stores every serialized policy prompt. We scanned all 265,712 pre-endpoint prompts of its 1,000 steps for any five-word sequence from the question, excluding only the model's own earlier thoughts: 0 matches. The same scan over the 95,482 pre-endpoint prompts of the new evaluation runs below: 0 matches.
  • A serialized prompt (training step 484, instruction lines abridged):
Evidence unit 4/26.
Running observations so far:
(none)
Current audio window: 0.00s to 2.00s.
This audio is the full speech prefix from the start up to the current boundary.
[...]
The spoken question and any answer options remain available only through the attached audio.
Continue listening and output exactly one XML action: <wait/> or <think>...</think>
Do not output <answer> yet.
  • Suffix test. Three recordings share the first 4.0 s and then differ. With a suffix of equal length and different content, all 8 pre-boundary policy inputs (prompt text and audio samples) are byte-identical. With a suffix of different length, the audio inputs are identical and the prompts differ only in N.
  • Total length N. "Evidence unit k/N" is shown to Base, SFT and DAPO alike, so the matched comparison is on equal inputs. We re-evaluated with N removed, without retraining:
Accuracy (%)Real Audio (186)LLaMA-QS (300)GSM8K (300)
DAPO, N shown (new run)65.171.322.7
DAPO, N hidden66.771.323.3
Base, N shown64.571.021.3
Base, N hidden65.170.720.0

Exact paired McNemar gives p ≥ .52 for all six comparisons; on LLaMA-QS the DAPO outputs are identical item by item. (GSM8K: fixed random 300-item sample. On the full Real Audio and LLaMA-QS sets, the N-shown rows match the paper to within one item.)

Q2 / W2. The optimization that was actually run

Was Eq. (4)'s negative-advantage branch used? Specify the clipping expression, normalization, KL term and credit assignment. How do the results change?

Yes, Eq. (4) as printed is what was trained, and it differs from DAPO Eq. (8) as you describe. The reported results are the results of this implemented variant. The two forms differ only on tokens whose ratio leaves the clip range; the logged out-of-range fraction averaged about 0.4% per step.

ItemPaperImplementationEffect
Clip branch, A < 0max, Eq. (4)max, as printedDiffers from DAPO only when the ratio leaves [1−ε_l, 1+ε_h]: logged fraction (mean within each sequence, then over the batch) about 0.4% per step, 1.15% at most
NormalizationGlobal token meanToken mean per turn, then mean over turnsEquation corrected
Credit assignmentOne trajectory advantageHybrid, α = 0.5Section 3.4 corrected
KL; clip; lr; rollouts; steps0.01; 0.20/0.28; 4e-7; 8; 1,000SameNone
  • Why the branch was rarely active. The updater runs one epoch per rollout batch and computes π_old with the current weights just before the update, so the ratio stays near 1. The fractions above are logged over all 1,000 steps. On every other token Eq. (4), DAPO Eq. (8) and the unclipped surrogate have the same gradient. We report this as the measured extent of the deviation, not as a retraining comparison.
  • Credit assignment. For wait and think turns, A_turn = 0.5·A_traj + 0.5·A_local. A_traj is the group-relative trajectory advantage. A_local is the per-turn process score, normalized over the group's wait and think turns: the timing score for a wait, the thought-quality score for a think. Answer turns use A_traj alone.
  • KL. Coefficient 0.01, toward the frozen SFT adapter.
AuthorsOpenReview ↗

Response to Reviewer ruEA (Part 2 of 2): Q3, Q4 and summary of revisions

Q3. Evidence for the controller's specific benefit

Do checkpoint-matched results exist for forced all-WAIT, fixed periodic updates and an external trigger? These would distinguish learned timing from improved reasoning content or length regularization.

We ran forced all-WAIT and fixed periodic updates on the same checkpoints, harness and scorer, changing only the schedule. The learned schedule beats periodic updates by 3.6 points, and post-training cuts early thoughts by 85% and moves them later.

Accuracy (%), Base / DAPOReal Audio (186)GSM8K (300)LLaMA-QS (300)
Fixed periodic: forced think every 2.0 s57.0 / 59.118.3 / 17.070.3 / 71.3
Policy's own schedule (new runs)64.5 / 65.121.3 / 22.771.0 / 71.3
Forced all-WAIT (final think kept)71.0 / 69.925.3 / 22.070.7 / 71.3
Early thoughts per item, own schedule7.31 / 1.090.83 / 0.070.18 / 0.00
  • Not periodic thinking. The policy's own schedule beats the periodic one on Real Audio (p = .004 Base, p = .043 DAPO) and GSM8K (p = .024 DAPO). Pooled: +3.6 points for DAPO (p = .015), +3.2 for Base (p = .025).
  • A shorter residual alone does not buy accuracy. Forcing early thoughts shortens the final think (DAPO on GSM8K: 12.8 → 4.8 tokens) and loses 5.7 points. The periodic schedule changes the number and content of thoughts as well as their timing, so this control compares scheduling strategies as a whole.
  • Learned timing. On Real Audio the trained policy writes 1.09 early thoughts per item against 7.31 for Base, and, among items with a thought, the first one at a median of 62% of the utterance against 31%.
  • All-WAIT. It is a legal trajectory of the same policy. These sets put the question at the end, and Real Audio Bench is built so that a late cue overrides earlier content, so waiting is strong: 2.9 points above Base's own schedule pooled (p = .004). After post-training the difference is 0.9 points and not significant (p = .26); the trained policy gives identical outputs to all-WAIT on LLaMA-QS, shows no significant difference on GSM8K (p = .77), and is below it on Real Audio (p = .049).
  • External trigger. We did not run a completeness or semantic trigger. It needs a separate detector to decide when to fire, and the paper lists this comparison as a limitation.
  • Native direct (Table 9). It has no reasoning span and is reported as the reference operating point; DAPO closes 64% of the gap from streaming Base to it.

Q4. Evaluation independence

What cross-split overlap checks were performed? How much of the improvement remains under deterministic-only scoring?

The lexical check over the DAPO training prompts finds four coincident question wordings and one shared passage, too few to explain the gain. The control comparisons hold under deterministic-only scoring.

  • Overlap. We compared all 1,000 prompts consumed by the DAPO run with all 8,959 synthetic and 186 recorded evaluation items by normalized exact match and 8-gram overlap. Evaluation uses the ARC and GSM8K test splits and the PIQA and SocialIQA validation splits; benchmark-derived training examples come from the training splits. Four training prompts share their wording with an evaluation item: three generic factual questions that coincide with LLaMA-QS items (e.g. "What is the largest planet in our solar system?") and one ARC stem that belongs to a different ARC item. One further ARC training item shares its introductory passage with an evaluation item. LLaMA-QS is the one task where DAPO shows no gain over Base; two ARC items contribute directly less than 0.06 points of ARC accuracy and less than 0.03 points of the full benchmark.
  • Answer-only ability is unchanged. Native-direct accuracy is the same before and after post-training (71.85% vs 71.89%, McNemar p = .731).
  • Deterministic-only scoring. With no LLM judge, the accuracies in the control tables are 0.3 to 3.2 points lower (the rule rejects "11:00" for "11 AM") and these comparisons hold: own vs periodic schedule, p = .0015 (Base, Real Audio) and p = .024 (DAPO, GSM8K); hiding N, p ≥ .52 throughout.
  • Judge. The reward judge and the answer scorer use different prompts and criteria. A second judge agrees on 99.11% of 2,916 final decisions (κ = 0.978).

Summary of revisions

  • State the code provenance; add the serialized prompt, the prompt scan and the suffix test to the appendix. The public export is corrected.
  • Correct Eq. (1) to include N and report the N-hidden results.
  • State the implemented objective exactly: clip branch and its measured activity, per-turn normalization, hybrid credit assignment, KL target.
  • Add the schedule controls, the overlap check and the deterministic-only scores.

We hope this evidence resolves the provenance and objective concerns, and respectfully ask the reviewer to reconsider the assessment.

AuthorsOpenReview ↗

Revised PDF uploaded

Dear Reviewer ruEA,

We have uploaded the revised paper; all changes are marked in blue. The items from our response are in place:

  • Q1. Equation (1) now includes the unit index and total count; the prompt scan, suffix test and code provenance are in Appendix D.7, and the total-length ablation is in Appendix D.6.
  • Q2. The implemented clipping branch, normalization and KL target are stated in Section 3.4 and Appendix B.3 (Equation 4).
  • Q3. Schedule controls are in Table 5 (Section 5.3), with paired tests in Appendix D.6.
  • Q4. The training–evaluation overlap check is in Appendix D.8, and rule-only scores sit beside judge-assisted scores in Appendix D.6.

A summary of the isolation and scoring checks is also in the main text (Section 5.3). We would be glad to answer any remaining questions. If the response and the revision address your concerns, we would be grateful if you would consider updating your assessment.

Reviewer_nMjuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a WAIT/THINK controller that allows an audio-language model to decide when to wait for more audio and when to generate intermediate thoughts while listening. It uses DAPO to optimize these decisions, with a reward function that focuses on both answer quality and reducing reasoning after the speech ends. The reward design includes timing-specific rewards to encourage the model to think when useful information arrives. The generated intermediate thoughts are stored and reused as more audio arrives, helping the model build on earlier reasoning and improve its final answers. The approach shows incremental improvements over the baseline models in answer accuracy and post-endpoint reasoning efficiency, although these gains are more consistent on synthetic speech than on real-world audio.

Critical strengths
  • The paper builds on the idea of reasoning while information is still arriving and applies it to spoken language. Similar ideas have been explored in other domains [Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models]. This is a useful extension of existing streaming reasoning approaches to audio, where important information may arrive later in a sentence.
  • The paper proposes a detailed reward model that considers answer correctness, reasoning quality, update timing, and remaining reasoning after the speech ends. These choices strike a good balance between answer quality and response latency.
  • The paper tests its approach on both synthetic and real human-recorded speech. Including real audio is a strength, as it helps assess how well the method works beyond synthetic examples.
Critical weaknesses
  • The authors mention that the reward weights were chosen through validation experiments, but they do not show the results or explain why the final values were selected. In particular, it is unclear why the update-timing reward is given a higher weight (3.0) than the others. More experiments with different reward weights would help show whether the results are reliable.
  • The paper lacks a reward baseline ablation like R_a + R_f + R_t, something that could fully show the importance of R_u & R_s, rewards centered more around the paper’s claims. The authors partially address this by mentioning that cumulative ablations cannot isolate individual reward contributions.
  • The paper uses GPT-4o to generate the initial training examples, so the model may learn when to think from these examples rather than learning the best timing decisions on its own. Synthetic data dependence may be an issue.
  • The authors do a good job of citing inspiration papers [AudioThinker, SHANKS etc.], there is a similar paper in terms of reward modelling [ CESAR: Incentivizing Consistent, Effective and Scalable Reasoning Capability in Audio LLMs via Reasoning Process Rewards ] that is worth a mention and could possibly be used for baseline work
  • Even though it sounds generic, the paper results are incremental over the baseline at best and do not concretely justify the novelty the paper is trying to point out. Some limitations presented by the paper, like the no direct wall-clock latency evaluation, would make the reasoning for this approach defensible.
Questions for authors
  • How much do the intermediate thoughts actually contribute to final answer accuracy? Could the authors evaluate performance when these thoughts are removed or replaced during inference?
  • Any specific reason for choosing DAPO amongst other optimization algorithms?
5 comments
AuthorsOpenReview ↗

Response to Reviewer nMju (Part 1 of 2): W1–W5

We sincerely thank the reviewer for recognizing the extension of streaming reasoning to audio, the reward design, and the evaluation on human-recorded speech. Part 1 answers W1–W5. Part 2 reports the inference-time test requested in Q1 and answers Q2.


W1. Reward weights and why λ_u = 3

The weights are hyperparameters, selected by training runs judged on the validation split. The reported values are the best configuration, and the one evaluated on all 8,959 test items.

  • Why λ_u is the largest. R_u is the term that directly scores when a thought is written, the behavior this paper studies. A trajectory with no thought can still be correct, well-formed and short, so the other terms give no direct feedback on timing.
  • What was tuned. Development runs varied λ_u (1.5, 3), λ_s (0.05, 0.3, 1) and λ_f (0.5, 1). λ_u and λ_s moved the operating point most (Appendix B).
  • Why the result is reliable. The selected configuration improves over Base on five of six task families, and the aggregate accuracy gain has an exact McNemar bound of p ≤ .0011.

W2. Importance of R_u and R_s

The ablation builds on what prior work has established. Audio-Thinker showed, with its own ablations, that format, accuracy, consistency and think-quality rewards improve reasoning in audio-language models. We take that as given and ablate what this paper adds.

  • Four terms are inherited. R_f, R_a, R_c and R_t follow the four kinds of reward that Audio-Thinker introduced and ablated.
  • R_u and R_s are what this paper adds. R_u gives direct supervision on update timing and R_s scores the reasoning left after the endpoint. An R_a + R_f + R_t objective provides no explicit target for evidence-aligned update timing.
  • What our ablation tests: once reasoning is interleaved with listening, do process rewards still help? They do. Accuracy rises at every step, and the full objective also reduces post-endpoint work.
ObjectiveAccuracyPost-end tokens
Streaming Base67.6%10.44
DAPO, 4 rewards (R_a, R_f, R_s, R_u)68.5%10.87
DAPO, 5 rewards (+R_t)69.2%10.94
DAPO, 6 rewards (+R_c)70.3%8.99

W3. Dependence on GPT-4o traces

The traces initialize the controller by supervised imitation. The gain comes from reinforcement learning on the model's own trajectories.

  • Imitation alone does not help. SFT on the GPT-4o traces gives 66.1%, below Base (67.6%).
  • Reinforcement learning produces the gain. DAPO does not ask the model to reproduce a teacher continuation. The model generates its own trajectories, they are scored by rewards, and the policy is updated from that feedback. This stage reaches 70.3%.
  • Five of six rewards score only the model's own output. R_u additionally uses the aligned update points as timing references, which we will state.

W4. CESAR and Think-as-You-See

Both are relevant and complementary; we will cite and discuss them.

CESARThink-as-You-SeeOurs
ModalityAudioVisionAudio
Reasoning runsAfter complete audioDuring the streamDuring the stream
What is learnedReasoning quality (GRPO)Streaming CoTWhen to wait or update state (trajectory DAPO)

Comparing with CESAR in streaming would require adapting it to partial-audio input and the controller's action interface, so its complete-audio results are not interchangeable with our streaming rows.

W5. Size of the gains and wall-clock evaluation

The gain is measured under a matched protocol, is statistically reliable, and improves accuracy and post-endpoint reasoning together.

ControllerAccuracyPost-end tokensMean RTF
Streaming Base67.6%10.441.137
SFT initialization66.1%9.821.353
DAPO (six rewards)70.3%8.991.134
  • The metric follows the paper that defined this setting. Can Speech LLMs Think While Listening? (Shih et al., ICLR 2026) defines latency as the number of tokens between the end of the question and the start of the response, and reports that count, not seconds. We measure the same boundary on Qwen, where those tokens are the final think: 10.44→8.99, a 13.9% reduction.
  • Why tokens and not seconds. A Moshi step has a fixed 80 ms clock. A Qwen text token does not: its duration depends on the serving stack, cache, batching and hardware, none of which belongs to the policy.
  • Wall-clock is already reported. Replay RTF is total wall-clock time over audio duration. In the reported replay-harness measurements it is 1.137 for Base and 1.134 for DAPO.
  • Contributions. A learned wait/think/answer formulation over partial audio; an SFT-to-DAPO pipeline with 75,723 aligned records; a six-term timing-aware reward; Real Audio Bench (186 human recordings).
AuthorsOpenReview ↗

Response to Reviewer nMju (Part 2 of 2): Q1, Q2 and planned revisions

Q1. How much do the intermediate thoughts contribute?

We ran this test for the response, on the same checkpoints, harness and scorer: thoughts replaced by a fixed schedule, and thoughts removed. The learned scheduling strategy outperforms forced periodic updates, and post-training makes intermediate thoughts fewer and later. The trained controller beats a fixed schedule by 3.6 points and writes 85% fewer early thoughts than the untrained one.

Accuracy (%), Base / DAPOReal Audio (186)GSM8K (300)LLaMA-QS (300)
Thoughts replaced: forced think every 2.0 s57.0 / 59.118.3 / 17.070.3 / 71.3
Policy's own schedule (new runs)64.5 / 65.121.3 / 22.771.0 / 71.3
Thoughts removed: wait until the endpoint, final think kept71.0 / 69.925.3 / 22.070.7 / 71.3
Early thoughts per item, own schedule7.31 / 1.090.83 / 0.070.18 / 0.00

The last row is a count, not accuracy. New runs; GSM8K is a fixed random sample of 300 test items. Two cells differ from the paper by one item each (Base on Real Audio 120/186 vs 119; DAPO on LLaMA-QS 214/300 vs 213). Tests are exact paired McNemar.

  • Replacing the learned schedule costs accuracy. A forced thought every 2.0 s is 3.6 points below the trained policy's own schedule pooled over the three sets (p = .015) and 5.9 points below on Real Audio (p = .043), with six times as many thoughts. More thoughts are not better.
  • Post-training makes thoughts fewer and later. On Real Audio the trained policy writes 1.09 early thoughts per item against 7.31 for Base, and, on items with a thought, the first one comes at a median of 62% of the utterance against 31%. Base often repeats the same state; the trained policy typically writes once. It learns this although the training recipe resamples all-wait groups.
  • Removing the thoughts. Waiting is a legal trajectory of the same policy, not another system. These sets put the question at the end, and Real Audio Bench is built so that a late cue overrides earlier content, so waiting is strong: 2.9 points above Base's own schedule pooled (p = .004). After post-training the difference is 0.9 points and not significant (p = .26); on Real Audio alone waiting remains ahead (p = .049). The trained controller has closed most of the distance the untrained one leaves.
  • Scope. These utterances are short (about 3 s, 15 s and 25 s on average) and hold one question. Longer, multi-part speech, where state must be carried, is the setting the formulation targets and is future work, as the limitations state.

Q2. Why DAPO?

Two properties of DAPO fit this problem, and our trainer builds on both.

  • Dynamic sampling. Many controller rollout groups are unusable for learning when to think: too few format-valid rollouts, no valid final think, or no rollout that writes an early thought. We adapt DAPO's dynamic resampling to filter on protocol validity and on the presence of an intermediate update (Section 3.4), so every update step trains on rollouts that exercise the controller.
  • Decoupled clipping. Wait is the majority action at every tick, so a useful think starts at low probability. DAPO's larger upper clip bound (0.28 against 0.20) lets a rare action that earned reward gain probability faster than a symmetric clip allows.

Planned revisions

  • Weight rationale and explored values next to the reward definition (W1).
  • Correspondence with Audio-Thinker and the ablation design beside the ablation table (W2).
  • SFT→DAPO comparison stated in the main text (W3).
  • CESAR and Think-as-You-See cited and discussed (W4).
  • Table 3 with residual tokens and RTF next to accuracy (W5).
  • The Q1 table and thought statistics added to the paper (Q1).

We hope these clarifications and the added evidence address the reviewer's concerns, and respectfully ask the reviewer to reconsider the assessment.

AuthorsOpenReview ↗

Revised PDF uploaded

Dear Reviewer nMju,

We have uploaded the revised paper; all changes are marked in blue. The items from our response are in place:

  • W1. Weight rationale and explored values, next to the reward definition (Section 3.5).
  • W2. Correspondence with Audio-Thinker and the ablation design (Section 5.1).
  • W3, Q2. SFT versus DAPO, and why DAPO fits this controller (Section 3.4).
  • W4. CESAR and Think-as-You-See cited and discussed (Section 2, Appendix D.5).
  • W5. Post-end tokens and RTF beside accuracy (Table 3).
  • Q1. Schedule controls and thought statistics (Table 5 in Section 5.3; details in Appendix D.6).

We would be glad to answer any remaining questions. If the response and the revision address your concerns, we would be grateful if you would consider updating your assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion for the Area Chair and reviewers

We thank the Area Chair and all five reviewers for the discussion. We summarize the contribution, the evidence added and the status of each concern.

Contribution

A spoken question can change meaning in its last few words. Even a model that answers only after the speaker finishes faces a choice while listening: update its reasoning now, or wait? We make this choice a trainable action of the audio model. Building on prior work on thinking while listening and on process rewards, the paper contributes:

  • Data. 75,723 synthetic audio–text records with reasoning updates aligned to the speech.
  • Training. SFT teaches the action protocol; DAPO, in a custom streaming-controller trainer, then optimizes a six-term trajectory reward that scores when an update occurs and how much reasoning remains after the endpoint, beyond answer correctness. Table 3 gives the cumulative reward ablation.
  • Evaluation. Matched controls under one backbone and one streaming protocol, and Real Audio Bench, 186 questions we recorded with five human readers and released under CC BY 4.0 (Appendix E).

Main result. On 8,959 synthetic spoken questions, post-training raises the controller's accuracy from 67.6% to 70.3% (exact McNemar bound p ≤ .0011), 4.2 points above its SFT initialization (p ≤ 5.2×10⁻⁷), while mean post-end reasoning falls from 10.44 to 8.99 tokens (Table 3). No loss of native-direct accuracy is detected (71.85% vs. 71.89%, p = .731).

Strengths noted by reviewers.

  • Veyg: an important problem for real-time speech models; reward ablations; synthetic and human-recorded evaluation.
  • ruEA: a concrete formulation; useful within-backbone controls; restrained reporting.
  • Jt78: a novel state-prediction design; DAPO training of the controller; analysis on human recordings.
  • nMju: rewards covering reasoning quality and timing; evaluation on real audio.
  • juxH: a useful research question; end-to-end engineering; very rigorous statistical analysis.

Evidence added during the discussion

ConcernRaised byWhat was added
Learned schedule vs. fixed periodic updatesnMju, ruEA, Veyg, juxHWith checkpoint, harness and scorer fixed, the learned schedule is 3.6 points above forced periodic updates (three sets, p = .015). Post-training cuts pre-endpoint thoughts on Real Audio from 7.31 to 1.09 per item. All-WAIT is reported beside it (Table 5; residual and RTF in our reply to ruEA); the paper does not claim that early thoughts beat waiting on the short utterances tested.
Does the policy see the question or future audio?ruEAIn 265,712 training and 95,482 control-run pre-endpoint prompts, no five-word sequence from the question appears outside the model's own earlier thoughts. An equal-length suffix swap leaves all pre-boundary inputs byte-identical. Hiding the total unit count gives p ≥ .52 in all six comparisons.
Training–evaluation overlapruEAAppendix D.8, extended in our reply to the full 73,675-record training split (lexical and embedding-based). With all 104 flagged items excluded, DAPO retains at least 226 more correct answers than Base on the remaining 8,855 items (+2.55 points).
Metric, margins and real audionMju, Veyg, juxH, Jt78Table 3 reports post-end tokens and replay RTF beside accuracy; tokens measure residual reasoning, not elapsed time. Tables 3 and 4 give 95% accuracy intervals; paired differences on the human recordings are in our reply to Jt78.
ScoringruEA, juxHRule-only scores beside judge-assisted scores on the control sets (Appendix D.6). A second judge agrees on 99.11% of 2,916 final rule-or-judge decisions (κ = 0.978) and on 97.13% of the 907 cases that reach the judge.
Objective, reward design, related workruEA, nMju, Jt78Implemented objective stated exactly (Section 3.4, Appendix B.3); weight rationale (Section 3.5); Base controller defined (Section 4); CESAR and Think-as-You-See discussed (Section 2, Appendix D.5).

Status of the discussion

  • Reviewer nMju raised the rating from 2 to 3 after the response and revision.
  • Reviewer ruEA wrote that soundness would be raised from 1 to 2. Our follow-up adds the requested all-WAIT measurements and overlap analysis.
  • Reviewer Jt78 wrote that most concerns in the question section are addressed; we responded to the two remaining points.
  • Reviewers Veyg and juxH: each item is addressed in our responses and listed in the "Revised PDF uploaded" comments.

This work makes the timing of reasoning a decision that a speech model can learn from partial audio. The aligned training data, matched scheduling controls, and human-recorded benchmark give other researchers a way to test when intermediate reasoning helps and when waiting is enough. We hope the Area Chair and reviewers will take these points into account in the final assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.