acceptodds
Under review as a conference paper at ICLR 2027

Eliciting Latent Capabilities in Spoken Dialogue Models via On-Policy Self-Distillation

Abstract

Post-training has become increasingly important for unlocking and refining the capabilities of foundation models. Spoken dialogue models introduce an additional challenge, as post-training should improve semantic reasoning and paralinguistic responsiveness jointly. We find that both are limited not by a lack of capability but by insufficient elicitation from speech alone. Providing transcripts and structured rationales substantially recovers both. Further analysis reveals local redundancy in discrete audio representations. Based on these findings, we propose VoxOPD, a cross-modal hierarchical on-policy self-distillation framework for spoken dialogue models. During training, the student generates response trajectories from speech alone, while a frozen teacher initialized from the same model conditions on privileged context to provide dense supervision along those trajectories. Beyond token-level distribution alignment, the framework assigns shared credit to audio segments through chunk-level supervision, accommodating differences in information density and local redundancy between text and audio representations. This hierarchical supervision transfers the capabilities elicited under augmented contexts to the speech-only setting. On Step-Audio 2 mini, VoxOPD improves the VoiceBench capability-specific task average by 4.0 and VStyle by 0.71. Further experiments on Kimi-Audio and VITA-Audio consistently demonstrate its effectiveness in improving both reasoning and spoken interaction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
spoken dialogue model, post-training

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_LqZFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes VoxOPD, a cross-modal hierarchical on-policy self-distillation method for spoken dialogue models. It aims to improve both semantic ability and speech expression. The student generates mixed text–audio trajectories from speech-only inputs. A frozen teacher, initialized from the same model, also receives transcripts and task-specific rationales. It provides supervision at student-generated prefixes. The objective combines text-token distillation, audio-token distillation, and a chunk-level loss that shares feedback across neighboring audio tokens.

The authors train on 38,532 English spoken-interaction examples and evaluate three backbones: Step-Audio 2 mini, Kimi-Audio, and VITA-Audio. Evaluation includes VoiceBench, VStyle, human listening tests, and several ablations. On Step-Audio 2 mini, VB-Cap increases from 60.7 to 64.7, and VStyle increases from 3.14 to 3.85. Aggregate scores improve over all three base models. However, gains vary across subtasks and do not always exceed other post-training methods. The main contribution is joint distillation over mixed text–audio outputs, with an audio-chunk credit assignment objective.

Critical strengths

Joint text and audio supervision is practically useful, and the ablations help separate the roles of its main components.

Critical weaknesses
  1. The comparison settings and the reliability of the key gains are not sufficiently established. This is my main concern. There is nothing about baseline training budgets, tuning ranges, the GRPO reward, or the adaptations of CORD and ParaBridge. Using the same data pool does not establish comparability along these dimensions.

  2. The connection between the perturbation evidence and the chunk-level objective needs clarification. Figure 1(b), Appendix A, and Table 4 show that decoded speech changes little when a small fraction of audio tokens is replaced by nearby codebook entries. However, this may reflect codebook geometry or decoder smoothness, and does not directly explain why feedback should be shared across neighboring time steps. It is also unclear whether teacher–student discrepancies during training resemble these perturbations.

Questions for authors

You should clarify experiment settings and the meaning of "local redundancy" in the chunk-level loss, time locality or codebook locality? I also wandor a more distinguish experimental observations from design motivation and mechanism about the design of the chunk-level objective.

5 comments

Thank you for your constructive feedback. We appreciate your positive assessment of the practical value of joint text and audio supervision and the role of our ablations in disentangling the contributions of the main components.

W1 & Q1. Baseline Training Configurations

We summarize the training configurations of all baselines below. For CORD and ParaBridge, we follow the objective-related hyperparameters in the original papers. To ensure a fair comparison, we adjust the remaining training settings only within reasonable ranges while keeping the overall training conditions comparable. Because ParaBridge self-distills by comparing the same model with and without a scaffold, it converges rapidly and degrades with continued training; we therefore adjust its number of training steps accordingly.

Training configurationSFTGRPOVoxOPDCORDParaBridge
Training duration1 epoch1 epoch1 epoch1 epoch300 steps
Per-GPU batch size11111
GPUs / gradient accumulation8 / 48 / 48 / 48 / 48 / 4
Effective batch size3232 prompts; 128 responses per step3232 prompts; 128 responses per step32
Parameter updateFull fine-tuningFull fine-tuningLoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05LoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05LoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05
OptimizerAdamWAdamWAdamWAdamWAdamW
Learning rate1e-51e-62e-51e-62e-5
Scheduler / warmuplinear / ratio 0.03cosine / 20 stepslinear / nonecosine / 20 stepslinear / none
Number of rollouts—4 per prompt1 per sample4 per prompt1 per sample
Main objective / rewardSupervised token-level CEWithin-group standardized advantage; text/audio KL coefficients of 0.001/0Loss weights: λtext=λaudio=1\lambda_{\mathrm{text}}=\lambda_{\mathrm{audio}}=1,λchunk=0.3\lambda_{\mathrm{chunk}}=0.3Sequence-level GRPO + weighted reverse KL; K=20K=20, α=β=2\alpha=\beta=2, λtok=1\lambda_{\mathrm{tok}}=1Standard JSD objective; λ=1\lambda=1, τ=1.2\tau=1.2

For GRPO, we use Gemini 2.5 Pro as the reward model. We have added the scoring criteria to the appendix.

W2.1 & Q1. Mechanism of Chunk-Level Supervision

Our design principle is to match the granularity of supervision to the granularity at which speech properties emerge.

Semantic and acoustic attributes are typically realized jointly by audio tokens across multiple temporal positions, and changing a single local token need not alter the corresponding output. Therefore, as discussed in the Introduction, token-level alignment may overemphasize differences that do not affect the output and consequently yield only limited improvements in actual generation.

From this perspective, codebook locality and temporal locality characterize different, non-exclusive aspects of the same representation-to-generation process. Codebook geometry, decoder smoothness, and temporal context may all attenuate the effect of local representation changes on the decoded output. The perturbation study is intended to show that changes in audio tokens do not correspond one-to-one with changes in the output, motivating complementary supervision over segments composed of multiple tokens. Accordingly, we retain token-level alignment and add chunk-level supervision to account for the segment-level behavior jointly determined by multiple audio decisions.

We further clarify the shared-credit mechanism at the chunk level. For each audio position in the current sampled trajectory, let

\log \pi_T(y_t\mid a,x,r,y_{<t}) -\log \pi_\theta(y_t\mid a,y_{<t}).$$ The shared credit assigned to chunk $C_k$ is $$A_k=\operatorname{sg}\!\left[ \frac{1}{|C_k|}\sum_{t\in C_k}\delta_t \right].$$ This mean is also the least-squares solution obtained when approximating the token-wise signal by a constant within each chunk: $$\frac{1}{|C_k|}\sum_{t\in C_k}\delta_t = \underset{b}{\operatorname{arg\,min}}\sum_{t\in C_k}(\delta_t-b)^2.$$ For a fixed sampled trajectory, and applying stop-gradient as in Eq. (5), the update direction for a single chunk is $$-\nabla_\theta \ell_k = \frac{1}{|C_k|^2} \sum_{i\in C_k}\sum_{j\in C_k} \delta_i\,\nabla_\theta\log\pi_\theta(y_j\mid a,y_{<j}).$$ Thus, the teacher signal at every position influences the updates of all tokens within the same chunk through chunk-level aggregation. This shares feedback locally, reduces sensitivity to discrepancies at individual token positions, and promotes segment-level consistency. As shown in Table 3, adding $\mathcal{L}_{\mathrm{chunk}}$ to token-level distributional alignment improves VB-Cap and VStyle from 63.9 and 3.71 to 64.7 and 3.85, respectively. To further examine this mechanism, we compare our method with generic approaches that weaken or smooth the token-level signal on Step-Audio 2 mini. All controls set $\lambda_{\mathrm{chunk}}=0$ and either reduce the weight of the audio-modality token-level loss, $\lambda_{\mathrm{audio}}$, or increase the temperature $\tau$ to smooth the teacher distribution. All other settings remain unchanged. | Setting | VB-Open | VB-Cap | VStyle | | --- | --- | --- | --- | | $\tau=1.0,\ \lambda_{\mathrm{audio}}=1.0$ | 3.39 | 63.9 | 3.71 | | $\tau=1.5,\ \lambda_{\mathrm{audio}}=1.0$ | 3.34 | 64.3 | 3.74 | | $\tau=2.0,\ \lambda_{\mathrm{audio}}=1.0$ | 3.33 | 63.8 | 3.80 | | $\tau=1.0,\ \lambda_{\mathrm{audio}}=0.5$ | 3.38 | 64.1 | 3.81 | | $\tau=1.5,\ \lambda_{\mathrm{audio}}=0.5$ | 3.38 | 63.9 | 3.78 | | Ours | 3.42 | 64.7 | 3.85 | Reducing $\lambda_{\mathrm{audio}}$ or increasing $\tau$ improves VStyle to some extent, supporting the observation that per-token audio supervision can overemphasize local differences. However, simply weakening or smoothing token-level supervision remains inferior to the full method across all three dimensions. This shows that chunk-level shared credit is not merely a form of distribution smoothing or loss reweighting. Instead, it aggregates teacher signals from adjacent tokens to provide complementary segment-level supervision, yielding more balanced improvements in semantic capability and acoustic expressiveness. This result is consistent with our motivation: adjacent audio tokens jointly determine the semantic and acoustic realization of a local speech segment, and the shared-credit mechanism provides a simple and effective supervisory signal for such segment-level dependencies. Smoothing the teacher distribution or adjusting $\lambda_{\mathrm{audio}}$ can also be combined with our method as a potential avenue for further optimization.

W2.2. Teacher–Student Discrepancies and Perturbations

To examine how teacher–student discrepancies during training relate to the perturbation study, we consider positions where the teacher's preferred token differs from the token sampled by the student. We then compute the proportion for which the teacher's preferred token falls among the student's top-KK nearest codebook neighbors (Hit@KK):

StepHit@5Hit@10Hit@20Hit@50
033.28%49.54%64.25%76.61%
30033.26%49.71%64.34%76.50%
60033.03%49.61%64.24%76.61%
90033.21%49.19%64.05%76.31%

These results show that, when the teacher and student disagree, the teacher's preferred token often lies in the local codebook neighborhood of the student's token. Thus, teacher–student discrepancies during actual training frequently take the form of local substitutions within the codebook, consistent with the locality of the replacements considered in the perturbation study.

Reviewer_fP4dOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VoxOPD, a hierarchical on-policy self-distillation framework for improving the semantic reasoning and paralinguistic responsiveness of spoken dialogue models. The method uses a student that generates mixed text–audio response trajectories from speech-only input and a frozen teacher initialized from the same pretrained model. The teacher receives additional privileged information, including transcripts and task-specific rationales, and provides supervision on the student's generated trajectories.

VoxOPD combines token-level distillation for text and audio with chunk-level credit assignment over audio tokens. Experiments on Step-Audio 2 mini, Kimi-Audio, and VITA-Audio report improvements on VoiceBench and VStyle. The authors also provide objective-level ablations, analyses of privileged teacher inputs, chunk-size sensitivity studies, and human evaluations of audio quality.

Critical strengths
  1. A relevant motivation for spoken dialogue post-training. The paper examines the gap between capabilities elicited by speech-only input and those available when transcripts or structured rationales are provided. The privileged-teacher formulation offers a reasonable approach to transferring these capabilities to a speech-only student.

  2. Useful experimental comparisons and analyses. The paper compares VoxOPD with several post-training baselines and evaluates three spoken dialogue architectures. It also includes ablations of the text, audio, and chunk-level objectives, as well as analyses of teacher-side privileged information. The human evaluation provides additional evidence for the reported speech-quality improvements.

Critical weaknesses

While VoxOPD achieves substantial improvements on Step-Audio 2 mini, the gains on Kimi-Audio and VITA-Audio are relatively modest. For example, the VoiceBench capability-specific task average increases from 60.7 to 64.7 on Step-Audio 2 mini, but only from 72.3 to 72.5 on Kimi-Audio. Although these results demonstrate the applicability of VoxOPD across different architectures, I am curious about the factors underlying such differences in improvement magnitude.

Questions for authors

Q1. Could the authors provide further insights into why VoxOPD achieves substantially larger improvements on Step-Audio 2 mini than on Kimi-Audio and VITA-Audio? Additional analysis or discussion of these factors would help clarify the generalizability and applicability of the proposed approach.

1 comment

We are grateful for your encouraging assessment of our motivation, privileged-teacher formulation, and experimental evaluation across multiple spoken dialogue architectures.

W1 & Q1. Differences in Performance Gains

As stated in Section 4.1, we directly use the Step-Audio 2 mini hyperparameters for Kimi-Audio and VITA-Audio, without additional tuning, to evaluate the transferability of our method. We subsequently conducted a small-scale hyperparameter search on VITA-Audio, focusing on metrics for which the gains in Table 1 were modest:

SettingOBQABBHMMSUAcousticInstructionEmpathy
Base model76.053.246.52.592.964.72
c=3, λchunk=0.2c=3,\ \lambda_{\mathrm{chunk}}=0.276.152.647.73.002.714.80
c=5, λchunk=0.2c=5,\ \lambda_{\mathrm{chunk}}=0.276.954.147.93.122.784.80
c=10, λchunk=0.2c=10,\ \lambda_{\mathrm{chunk}}=0.275.853.147.73.062.784.81
c=3, λchunk=0.3c=3,\ \lambda_{\mathrm{chunk}}=0.376.752.347.53.222.774.68
c=5, λchunk=0.3c=5,\ \lambda_{\mathrm{chunk}}=0.376.554.247.73.072.764.78
c=10, λchunk=0.3c=10,\ \lambda_{\mathrm{chunk}}=0.376.554.3046.92.932.574.85

With c=5c=5 and λchunk=0.2\lambda_{\mathrm{chunk}}=0.2, VITA-Audio improves by 0.9, 0.9, 1.4, and 0.53 on OBQA, BBH, MMSU, and Acoustic, respectively. This result indicates that the modest gains in Table 1 are partly attributable to the absence of model-specific hyperparameter tuning.

As shown in Table 1, Kimi-Audio already provides a relatively strong baseline, leaving less room for further improvement. Nevertheless, VoxOPD improves OBQA and BBH from 83.5 and 67.8 to 85.5 and 68.8, respectively. The smaller aggregate gain on VB-Cap is primarily due to the decrease in IFEval from 61.2 to 58.9. We also evaluate the corresponding privileged teacher, which obtains 96.92, 94.20, and 68.70 on OBQA, BBH, and IFEval, respectively. These results provide preliminary evidence that the benefit of self-distillation may depend on the teacher–student capability gap: the larger gaps on OBQA and BBH provide more transferable signal, whereas the smaller gap on IFEval may limit the potential gain.

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

This paper proposes an on-policy self-distillation method "VoxOPD" for spoken dialogue models. A student generates from speech only, while the teacher additionally receives the transcript and a task-specific rationale. VoxOPD combines token-level distillation for text and audio tokens with a chunk-level audio objective, motivated by the observation that audio representations contain local redundancy. The method is evaluated on Step-Audio 2 mini, Kimi-Audio, and VITA-Audio.

Critical strengths
  • The paper clearly motivates the need to improve semantic reasoning and paralinguistic behavior jointly.

  • The results on Step-Audio 2 are very strong. VoxOPD improves VB-Cap from 60.7 to 64.7 and VStyle from 3.14 to 3.85, outperforming SFT, GRPO, CORD, and ParaBridge overall.

  • The paper includes ablations on the different loss terms, teacher-side privileged information, chunk size, and chunk-loss weight, as well as human evaluation. The human evaluation also supports the improvements on spoken style. I also liked the audio-token perturbation analysis, since it gives some empirical support for the claim that local audio-token changes can have limited perceptual impact.

Critical weaknesses
  • The paper builds on existing work on on-policy distillation, privileged self-distillation, and cross-modal speech distillation. The main methodological novelty appears to be the chunk-level supervision for audio tokens, together with the specific privileged teacher setup. I would like the paper to clarify more explicitly how VoxOPD differs from the closest prior distillation methods, in particular CORD/X-OPD and related privileged self-distillation approaches.

  • I am not fully convinced by the motivation for the chunk-level objective. The perturbation experiment shows local redundancy in audio tokens, but it does not directly show why assigning the same credit to all tokens in a fixed chunk is the right objective. The ablations show that it helps, but the connection is still somewhat a heuristic.

  • I would also like more details about the training and tuning of the baselines. The paper states that the methods use the same training-data pool, but especially GRPO can still be sensitive to reward design and hyperparameters.

  • The paper seems to report only single training runs. For the larger Step-Audio gains this is less concerning, but for the small differences on Kimi-Audio and VITA-Audio it is hard to judge whether the improvements are meaningful.

Questions for authors
  • What exact training configurations were used for SFT and GRPO? Were hyperparameters tuned independently for each baseline with comparable budgets? In particular, what reward formulation was used for GRPO?

  • How does chunk size c=10 map to actual time in milliseconds for Step-Audio 2? Is this temporal span similar across Kimi-Audio and VITA-Audio? If audio-token densities differ substantially across architectures, using the identical token-level c seems somewhat surprising.

  • Why should audio-token redundancy specifically imply shared credit rather than reduced or smoothed token-level supervision? I think this conceptual link deserves a more formal explanation.

4 comments

Thank you for your thoughtful and detailed review. We appreciate your positive assessment of our motivation, empirical results, and analyses.

W1. Distinction from Existing Distillation Methods

X-OPD uses an additional text teacher model to transfer textual reasoning capabilities to a dialogue model. ParaBridge instead compares the behavior of the same model with and without a scaffold, turning its latent paralinguistic awareness into observable dialogue behavior. CORD uses the text-input branch of the dialogue model as the teacher and combines a weighted reverse KL divergence with sequence-level GRPO based on the consistency between audio- and text-conditioned answers. Despite their relevance to OPSD, CORD and ParaBridge each have certain limitations. As discussed in Section 4.2, CORD relies exclusively on the text-input branch of the dialogue model as its teacher. This may propagate erroneous signals from the teacher branch to the student and cannot provide paralinguistic information. ParaBridge, by contrast, relies on a generic scaffold and thus provides neither direct semantic and reasoning enhancement from a teacher model nor explicit guidance for speech generation. More broadly, none of these three approaches provides supervision specifically tailored to the audio modality.

In contrast, our method addresses both semantic decision-making and acoustic generation. In addition to token-level distributional alignment in the audio modality, we introduce chunk-level supervision to better exploit the structure of audio representations. We further construct query-specific guidance to improve performance and systematically analyze alternative ways of constructing privileged information.

W2 & W3. Mechanism of the Chunk-Level Loss

As discussed in the Introduction, changing individual tokens does not necessarily affect the final output. Token-level alignment may therefore overemphasize output-insensitive differences and yield limited gains in actual generations. This motivates us to additionally constrain audio supervision at the chunk level.

Using Eqs. (5)–(6), we further clarify the mechanism of the chunk-level loss. For each audio position in the current sampled trajectory, let

\log \pi_T(y_t\mid a,x,r,y_{<t}) -\log \pi_\theta(y_t\mid a,y_{<t}).$$ The shared credit assigned to chunk $C_k$ is $$A_k=\operatorname{sg}\!\left[ \frac{1}{|C_k|}\sum_{t\in C_k}\delta_t \right].$$ This mean is also the least-squares solution obtained when approximating the token-wise signal by a constant within each chunk: $$\frac{1}{|C_k|}\sum_{t\in C_k}\delta_t = \underset{b}{\operatorname{arg\,min}}\sum_{t\in C_k}(\delta_t-b)^2.$$ For a fixed sampled trajectory, and applying stop-gradient as in Eq. (5), the update direction for a single chunk is $$-\nabla_\theta \ell_k = \frac{1}{|C_k|^2} \sum_{i\in C_k}\sum_{j\in C_k} \delta_i\,\nabla_\theta\log\pi_\theta(y_j\mid a,y_{<j}).$$ Thus, the teacher signal at every position influences the updates of all tokens within the same chunk through chunk-level aggregation. This shares feedback locally, reduces sensitivity to discrepancies at individual token positions, and promotes segment-level consistency. We further conduct experiments on Step-Audio 2 mini to determine whether simply weakening or smoothing token-level supervision can account for these benefits. Specifically, we remove the chunk loss and either reduce the weight of the audio-modality token-level loss, $\lambda_{\mathrm{audio}}$, or increase the temperature $\tau$ to smooth the teacher distribution. All other settings remain unchanged. | Setting | VB-Open | VB-Cap | VStyle | | --- | --- | --- | --- | | $\tau=1.0,\ \lambda_{\mathrm{audio}}=1.0$ | 3.39 | 63.9 | 3.71 | | $\tau=1.5,\ \lambda_{\mathrm{audio}}=1.0$ | 3.34 | 64.3 | 3.74 | | $\tau=2.0,\ \lambda_{\mathrm{audio}}=1.0$ | 3.33 | 63.8 | 3.80 | | $\tau=1.0,\ \lambda_{\mathrm{audio}}=0.5$ | 3.38 | 64.1 | 3.81 | | $\tau=1.5,\ \lambda_{\mathrm{audio}}=0.5$ | 3.38 | 63.9 | 3.78 | | Ours | 3.42 | 64.7 | 3.85 | The results show that weakening or smoothing token-level supervision can partially mitigate its overemphasis on output-insensitive differences, but its overall performance across the evaluation dimensions remains below that of our full method. This finding suggests that, compared with an individual speech token, a segment comprising multiple tokens carries more complete semantic and acoustic information. Chunk-level shared credit is therefore not merely a form of distribution smoothing or loss reweighting. By aggregating teacher signals from adjacent tokens, it provides complementary segment-level supervision and yields more balanced improvements in semantic capability and acoustic expressiveness. Smoothing the teacher distribution and adjusting $\lambda_{\mathrm{audio}}$ are compatible with our method; in practice, these hyperparameters may be further tuned to improve performance.

W3 & Q1. Baseline Training Configurations

We summarize the training configurations of all baselines below. Hyperparameters associated with each training objective follow the corresponding original work. To ensure a fair comparison, we adjust the remaining training settings only within reasonable ranges while keeping the overall training conditions comparable. Because ParaBridge performs self-distillation using a paralinguistic scaffold and converges rapidly, continued training leads to performance degradation; we therefore adjust its number of training steps accordingly.

Training configurationSFTGRPOVoxOPDCORDParaBridge
Training duration1 epoch1 epoch1 epoch1 epoch300 steps
Per-GPU batch size11111
GPUs / gradient accumulation8 / 48 / 48 / 48 / 48 / 4
Effective batch size3232 prompts; 128 responses per step3232 prompts; 128 responses per step32
Parameter updateFull fine-tuningFull fine-tuningLoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05LoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05LoRA r=16r=16, α=32\alpha=32, dropout =0.05=0.05
OptimizerAdamWAdamWAdamWAdamWAdamW
Learning rate1e-51e-62e-51e-62e-5
Scheduler / warmuplinear / ratio 0.03cosine / 20 stepslinear / nonecosine / 20 stepslinear / none
Number of rollouts—4 per prompt1 per sample4 per prompt1 per sample
Main objective / rewardSupervised token-level CEWithin-group standardized advantage; text/audio KL coefficients of 0.001/0Loss weights: λtext=λaudio=1\lambda_{\mathrm{text}}=\lambda_{\mathrm{audio}}=1,$ \lambda_{\mathrm{chunk}}=0.3$Sequence-level GRPO + weighted reverse KL; K=20K=20, α=β=2\alpha=\beta=2, λtok=1\lambda_{\mathrm{tok}}=1Standard JSD objective; λ=1\lambda=1, τ=1.2\tau=1.2

For GRPO, we use Gemini 2.5 Pro as the reward model. We have added the scoring criteria to the appendix.

W4 & Q2. Cross-Model Hyperparameters and Performance

Let the effective audio output rate be faudiof_{\mathrm{audio}} tokens/s. The temporal span represented by a chunk is then

Δt=cfaudio×1000 ms.\Delta t=\frac{c}{f_{\mathrm{audio}}}\times1000\ \mathrm{ms}.

For Step-Audio 2 mini, faudio=25f_{\mathrm{audio}}=25, so c=10c=10 corresponds to 400 ms of audio. For Kimi-Audio and VITA-Audio, faudio=12.5f_{\mathrm{audio}}=12.5, so the same chunk size corresponds to 800 ms.

As stated in Section 4.1, we did not perform additional tuning for Kimi-Audio or VITA-Audio in order to evaluate the transferability of our method. Both models therefore use the same settings, c=10c=10 and λchunk=0.3\lambda_{\mathrm{chunk}}=0.3.

We subsequently conducted a small-scale hyperparameter search on VITA-Audio, focusing on metrics for which the gains in the main table were modest:

SettingOBQABBHMMSUAcousticInstructionEmpathy
Base model76.053.246.52.592.964.72
c=3, λchunk=0.2c=3,\ \lambda_{\mathrm{chunk}}=0.276.152.647.73.002.714.80
c=5, λchunk=0.2c=5,\ \lambda_{\mathrm{chunk}}=0.276.954.147.93.122.784.80
c=10, λchunk=0.2c=10,\ \lambda_{\mathrm{chunk}}=0.275.853.147.73.062.784.81
c=3, λchunk=0.3c=3,\ \lambda_{\mathrm{chunk}}=0.376.752.347.53.222.774.68
c=5, λchunk=0.3c=5,\ \lambda_{\mathrm{chunk}}=0.376.554.247.73.072.764.78
c=10, λchunk=0.3c=10,\ \lambda_{\mathrm{chunk}}=0.376.554.3046.92.932.574.85

With c=5c=5 and λchunk=0.2\lambda_{\mathrm{chunk}}=0.2, OBQA, MMSU, and Acoustic improve from 76.0, 46.5, and 2.59 to 76.9, 47.9, and 3.12, respectively. These results indicate that the modest gains reported in the main table are partly attributable to the absence of model-specific hyperparameter tuning.

As shown in Table 1, Kimi-Audio already provides a relatively strong baseline, leaving less room for further improvement. Nevertheless, VoxOPD improves OBQA and BBH from 83.5 and 67.8 to 85.5 and 68.8, respectively. The smaller aggregate gain on VB-Cap is primarily due to the decrease in IFEval from 61.2 to 58.9. We also evaluate the corresponding privileged teacher, which obtains 96.92, 94.20, and 68.70 on OBQA, BBH, and IFEval, respectively. These results provide preliminary evidence that the benefit of self-distillation may depend on the teacher–student capability gap: the larger gaps on OBQA and BBH provide more transferable signal, whereas the smaller gap on IFEval may limit the potential gain.

Reviewer_UQVQOpenReview ↗

Final Response by Reviewer

Thanks for the detailed response, which answers my questions on the baselines, the GRPO reward, and the chunk-level loss, so I raise my soundness score to 3. I keep my overall rating at 3.

Thank you for your thoughtful review and for taking the time to follow up. Your feedback helps us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.