acceptodds
Under review as a conference paper at ICLR 2027

GDPO-Listener: Expressive Interactive Head Generation via Auto-Regressive Flow Matching and Group Reward-Decoupled Policy Optimization

Abstract

Generating realistic 3D head motion for dyadic interactions is a significant challenge in virtual human synthesis. While recent methods achieve impressive results with speaking heads, they frequently suffer from the `Regression-to-the-Mean' problem in listener motions, collapsing into static faces, and lack the parameter space for complex nonverbal motions. In this paper, we propose GDPO-Listener, a novel framework that achieves highly expressive speaking and listening motion generation. First, we introduce an Auto-Regressive Flow Matching architecture enabling stable supervised learning. Second, to overcome kinematic stillness, we apply the Group reward-Decoupled Policy Optimization (GDPO). By isolating reward normalization across distinct FLAME parameter groups, GDPO explicitly incentivizes high variance expressive generations. Finally, we enable explicit semantic text control for customizable responses. Extensive evaluations across the Seamless Interaction and DualTalk datasets demonstrate superior performance compared to existing baselines on long-term kinematic variance, visual expressivity and semantic controllability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Interactive Head Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Tj9kOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GDPO-Listener, a two-stage framework for expressive listener head-motion generation. It combines autoregressive flow matching with reinforcement learning post-training, using group-wise reward normalization to improve motion expressiveness and temporal dynamics. Experiments demonstrate improvements over existing methods.

Critical strengths
  1. Relevant problem: Addresses the challenge of generating expressive and natural listener reactions in conversations.

  2. Effective framework: Combines flow matching with reward-driven post-training in a coherent pipeline.

  3. Promising results: Experiments and ablations demonstrate improvements over existing methods, including standard GRPO.

Critical weaknesses
  1. Limited methodological novelty: GRPO-based post-training has become increasingly common in generative modeling. Applying it to expressive head-motion generation, even with task-specific rewards, may represent a relatively incremental contribution. The methodological novelty of GDPO over existing policy optimization approaches could be better established.
  2. The necessity of reinforcement learning is not fully established. The proposed rewards are primarily based on explicitly defined motion statistics and reconstruction errors. It remains unclear whether the observed improvements require policy optimization or could be achieved through direct supervised fine-tuning with corresponding auxiliary losses. A comparison against such an alternative would help establish the specific value of the reinforcement learning stage.
  3. Limited validation of GDPO: The contribution of group-wise reward normalization and its sensitivity to design choices could be investigated further.
Questions for authors
  1. How does GDPO compare with supervised fine-tuning using the same reward components as auxiliary losses?
  2. How do you ensure that the reward design improves naturalness and conversational relevance rather than merely increasing motion variability?
  3. How sensitive is GDPO to the choice of parameter groups and reward weights?
  4. Why do speaking metrics improve if reinforcement learning is applied exclusively to listener frames?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Tj9k

We thank the reviewer for their time and valuable feedback. Please find the responses to the raised concerns below:

1. Novelty (W1)

We respectfully clarify that our contribution goes beyond applying GRPO-based post-training. First, to our knowledge, this is the first work to introduce RL post-training to dyadic listening head generation, treating listener expressiveness as an alignment problem. Second, unlike previous autoregressive methods that rely on discrete motion tokens, we build a continuous Auto-Regressive Flow Matching model, and adapt RL to it with SDE rollouts. Third, we design dynamics-aware rewards with group-wise normalization over an expanded FLAME space. With these designs, we demonstrate clear improvements in addressing the problem of static listening head reactions, as shown in Tab. 1, Tab. 5, Tab. 9, Tab. 10 and Tab. 12.

2. Necessity of reinforcement learning (W2 & Q1)

Thank you for the suggestion. First, regarding the necessity of RL, we compare with supervised alternatives in Tab. 8, including the supervised AR-Flow model and AR-Flow with SFT on an expressive subset. Both perform worse than our GDPO post-training.

Second, regarding auxiliary losses, our rewards cannot be directly used as auxiliary losses in supervised training. They are sequence-level statistics computed on the full decoded motion (200 frames), which requires a complete autoregressive rollout. In contrast, AR-Flow is trained with teacher forcing, where each step only predicts a single latent token given the reference history, so no full sequence is generated during training. Instead, RL post-training naturally optimizes such sequence-level rewards on complete rollouts.

3. Validation and sensitivity of GDPO (W3 & Q3)

Thank you for the question. These analyses are provided in the appendix. For the contribution of group-wise reward normalization, we compare GDPO with standard GRPO in Tab. 12 and Fig. 6 (Appendix C.6). For the sensitivity to parameter groups, we conduct leave-one-out ablations in Tab. 10 (Appendix C.3), where the reward weight of each FLAME group is set to zero in turn. For the reward design, we further ablate the reward components in Tab. 9 (Appendix C.2). These results show that group-wise normalization and each reward design contribute to the final performance.

4. Naturalness beyond motion variability (Q2)

This is a fair concern. Our rewards not only encourage expressive motion dynamics, but also preserve naturalness and conversational relevance. Among the four reward terms, the variance and velocity-variance terms are responsible for expressive dynamics, while the mean and MSE terms serve as regularization to keep the motion natural and grounded in the conversational context. This is further verified by our user study in Tab. 5, where participants prefer our method on naturalness, liveliness and diversity. We also provide video demos in the supplementary materials for visual comparison.

5. Scope of RL post-training (Q4)

Thank you for pointing this out. The statement in Sec. 3 that GDPO is “applied exclusively to listener frames” is a writing mistake, and we have corrected it in the revised paper. GDPO post-training is applied to all frames, both speaking and listening, consistent with the reward description in Sec. 3.3.

Since our rewards are designed for motion dynamics, the speaking dynamic metrics also improve substantially (Tab. 7: FDD 39.54 → 22.45, PDD 3.50 → 2.71, JDD 1.81 → 1.26). In contrast, the lip-sync metric remains largely unchanged (Tab. 7: LVE 3.07 → 2.95), as lip motion is largely determined by the audio and is largely unaffected by RL post-training.

Thank you for the response. I believe a score of 3 is appropriate, and I will maintain my positive rating.

Reviewer_iE6qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets expressive 3D speaking/listening head generation in dyadic conversation. It argues that supervised models collapse to low-variance mean listener motion. The method including (1) stage 1: an auto-regressive flow matching (AR-Flow) model over a VAE latent of an expanded FLAME space, conditioned on a text prompt, both participants' audio and partner motion, with time-aligned RoPE and CFG for intensity control; (2) stage 2: RL post-training with GDPO, where rewards are computed per FLAME group, normalized per group and then combined. Experiments on Seamless Interaction, DualTalk and zero-shot ViCo show better dynamic metrics.

Critical strengths
  • The expanded FLAME space is practical.
  • The experiments are extensive.
  • The RL implementation is described in detail.
Critical weaknesses
  • FDD/PDD/JDD are defined as the absolute difference between the predicted and GT std of frame-to-frame velocities, and the velocity-variance reward term is the same quantity. Table 9 shows this single term gives almost all of the gain. So the main improvements in Tables 1-4 and 7 largely measure how well the policy fits its own reward.
  • The authors extract FLAME parameters themselves with SHeaP and their own smoothing, while the competitors are not. Numbers computed on different GT representations should not appear in one table.
  • The claims about UniLS, i.e., biased prior, catastrophic forgetting, are not supported by experiments.
  • Flow matching is already a generative model, not regression. The rewards still compare each rollout with its single GT sequence, so the model is not really rewarded for diverse but valid reactions.
  • The main new part of this paper seems to be applying GDPO to this task with per-FLAME-group rewards, which I do not think is enough to support the paper.
Questions for authors

Please refer to weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer iE6q

We thank the reviewer for their time and valuable feedback. Please find the responses to the raised concerns below:

1. Bias of the metrics toward the reward (W1)

We agree that the velocity-variance term contributes most of the gain. However, the other terms also matter. As shown in Tab. 9, using the velocity variance term alone degrades lip-sync and geometric accuracy (LVE 3.07 → 3.23, MHD 1.12 → 1.17), while the full reward improves them (LVE 2.95, MHD 1.06). Moreover, without the mean and MSE terms as regularization, we observe visual artifacts such as unnatural head rotations, which are not reflected in these numbers. We therefore choose the full reward for the best overall performance and visual quality.

For metrics that are not part of our reward, Tab. 11 in Appendix C.5 reports FD, SID and rPCC, where GDPO-Listener also brings consistent improvements. We have also added a paragraph "Metrics Beyond the Rewards" in Sec. 4.1 of the revised main paper to emphasize them.

2. Comparison with baselines on the same representation (W2)

Thank you for raising this point. UniLS [R1] in Tab. 1 is evaluated on the same FLAME parameters extracted by SHeaP [R2], and it is also trained on the Seamless Interaction dataset. For a fairer comparison, we further finetune UniLS on our SHeaP-extracted and smoothed FLAME parameters:

MethodsSpeakingListening
LVE ↓MHD ↓FDD ↓PDD ↓JDD ↓FDD ↓PDD ↓JDD ↓
UniLS (official checkpoint)4.091.4434.975.394.5322.232.851.75
UniLS (finetuned on our data)2.961.4129.373.142.2221.272.181.34
GDPO-Listener2.951.0622.452.711.2618.851.900.87

After finetuning, UniLS improves notably and achieves comparable lip-sync accuracy, showing that it is a strong baseline. Our GDPO-Listener still shows advantages on the remaining metrics in both speaking and listening. These results are added to Tab. 15 and Appendix C.8 of the revised paper.

3. Claims about UniLS (W3)

Thank you for pointing this out. These claims were speculations based on our understanding of the UniLS two-stage training, and are not supported by experiments. We have removed them from the revised paper.

4. Flow matching and the reward design (W4)

That is an interesting question. We agree that flow matching is a generative model, not regression. We also want to mention that the discrete autoregressive model, as used in UniLS [R1], is also a generative model. However, both UniLS and our work observe the same phenomenon: under direct supervised training, the generated listening motion is stiff and low-variance. Even with supervised finetuning (SFT) on an expressive subset (Tab. 8), the problem is not resolved. In this sense, the difference only lies in the solution: UniLS chooses two-stage training, while we choose RL post-training. They are just different solutions to the same general problem.

Regarding the reward, it is not a per-frame regression. Three of its four terms are sequence-level statistics (variance, velocity variance and mean), so different reactions with similar dynamics receive similar rewards, and the MSE term only serves as regularization. The diversity of our results is supported by Tab. 11 and our user study in Tab. 5.

5. Novelty (W5)

We respectfully clarify that our contribution goes beyond applying GDPO. First, to our knowledge, this is the first work to introduce RL post-training to dyadic listening head generation, treating listener expressiveness as an alignment problem. Second, unlike previous autoregressive methods that rely on discrete motion tokens, we build a continuous Auto-Regressive Flow Matching model, and adapt RL to it with SDE rollouts and chunked replay. Third, we design dynamics-aware rewards with group-wise normalization over an expanded FLAME space. With these designs, we demonstrate clear improvements in addressing the problem of static listening head reactions, as shown in Tab. 1, Tab. 5, Tab. 9, Tab. 10 and Tab. 12.


[R1] Chu, X., Liu, R., Huang, Y., Liu, Y., Peng, Y., & Zheng, B. (2026). Unils: End-to-end audio-driven avatars for unified listening and speaking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 25142-25152).

[R2] Schoneveld, L., Chen, Z., Davoli, D., Tang, J., Terazawa, S., Nishino, K., & Nießner, M. (2025, October). Sheap: Self-supervised head geometry predictor learned via 2d gaussians. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 14162-14172). IEEE.

AuthorsOpenReview ↗

Response to Reviewer iE6q

Dear Reviewer iE6q,

As the discussion deadline approaches, we wanted to check that our previous reply has addressed your concerns. Should any further questions or specific points remain, we will be happy to provide further clarification.

We sincerely appreciate the time and detailed effort you have invested in reviewing our submission.

I have read the response and the revised paper. Thanks for removing the unsupported claims about UniLS and for adding the finetuned UniLS baseline. However, the new results make my first two concerns stronger rather than weaker. Once UniLS is finetuned on the same data, the lip-sync gap disappears (LVE 2.96 vs 2.95), and the remaining advantage is on FDD/PDD/JDD, which are the quantities the reward directly optimizes. The authors also agree that the velocity-variance term gives most of the gain; in Tab. 9 the full reward only moves speaking FDD from 23.02 to 22.45 over that single term. FD in Tab. 11 is not really independent either, since the reward already matches the mean and variance of the GT sequence. On W4, the reward is still computed against the statistics of one GT sequence, so it does not reward diverse but valid reactions, and Tab. 8 shows that simply using more denoising steps already brings listening FDD from 28.03 to 21.07 without any RL. On novelty, the response restates the contributions, and I still see the paper mainly as GDPO applied to this task with per-group rewards. Overall I think the paper is below the acceptance bar, and I keep my rating.

AuthorsOpenReview ↗

Further Clarification to Reviewer iE6q

We thank the reviewer for reading our response and the revised paper. We respect the reviewer’s judgment, and would like to clarify three points with the numbers in the paper.

  1. The remaining advantage is not only on the reward-related metrics. Against the finetuned UniLS, our method keeps a clear advantage on MHD (1.06 vs. 1.41 in Tab. 15), which measures geometric accuracy and already holds for our model without RL (1.12 in Tab. 7). Metrics that are not targeted by any reward term, such as SID and rPCC, also improve after GDPO post-training (Tab. 11). Moreover, our paper focuses on listening. Finetuning mainly improves the speaking lip-sync of UniLS, while its listening FDD changes only marginally (22.23 → 21.27, vs. 18.85 for ours), and all three listening metrics remain behind ours.

  2. More denoising steps do not replace RL, especially for listening. In Tab. 8, GDPO with the default 10 steps outperforms AR-Flow with more denoising steps on all eight metrics (e.g., listening FDD 18.85 vs. 21.07, listening PDD 1.90 vs. 2.37), without additional inference cost. The stiff listening problem cannot be resolved by more sampling steps alone.

  3. Novelty. To our knowledge, this is the first work to introduce RL post-training to dyadic listening head generation. Reviewer Uxod found this approach both principled and interesting and largely underexplored in this task, and Reviewer Tj9k recognized it as an effective framework.

We thank the reviewer again for the detailed discussion.

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

Task Description

  • This paper addresses dyadic 3D conversational head generation. The goal is to generate an actor's facial and head motion as FLAME parameters, covering both speaking and listening. The inputs are the actor's audio, the partner's audio and motion, and an optional text prompt.

Research Motivation

  • Existing methods predict only a restricted subset of FLAME parameters (expression, jaw, and neck), so they cannot express blinks or large head nods.
  • Listening motion is inherently one-to-many, but existing models are trained to match a single ground-truth trajectory. This leads to "Regression-to-the-Mean": static, low-variance listening heads. This paper therefore treats listener expressiveness as an alignment problem rather than a purely supervised one.

Main Contributions

  • The paper introduces GDPO-Listener, a two-stage framework. (i) An auto-regressive flow matching model is trained on an expanded FLAME space that adds eyelid, eye pose, and global head rotation. (ii) The model is then post-trained with GDPO, which normalizes rewards separately for each FLAME parameter group and rewards sequence-level motion variance.
Critical strengths
  • Well-motivated problem formulation and a sound, interesting approach. The paper builds on a compelling intuition: speaking facial motion is largely determined by the speaker's own audio, whereas listening facial motion is only weakly conditioned on the conversational context. Based on this observation, the authors formulate listening motion generation as a one-to-many problem. To address it, they propose an RL post-training stage with sequence-level rewards that primarily target the positional and velocity variance of the generated motion. I find this approach both principled and interesting.

  • Excellent presentation quality. The abstract and teaser figure alone clearly convey the task and what distinguishes this work from prior approaches. The introduction clearly explains the authors' perspective on the limitations of existing methods, and the specific problem formulation derived from this analysis makes the necessity of the work evident. On the limited budget, the method section describes the core components concisely, and the proposed model is easy to follow. The experiments are also well designed: ablation studies isolate the contribution of each key component, and the comparisons are extensive and convincing across multiple benchmark datasets.

Overall, the paper conveys its claims persuasively, introduces a strong problem formulation that can serve as a valuable reference for speaking and listening facial motion generation, and proposes an interesting RL-based post-training methodology that has been largely underexplored in this task.

Critical weaknesses
  • Mismatch between the RL rollout sampler and the inference sampler. Introducing SDE-based exploration is itself consistent with the non-deterministic nature of listener reactions. However, the RL rollouts use an Euler–Maruyama sampler without the score-dependent drift correction (Eq. 7), whereas inference uses a deterministic ODE sampler. Since the SDE rollouts do not preserve the marginals of the inference ODE, there is no guarantee that the distribution optimized by RL matches the distribution actually produced at inference. A theoretical justification or an empirical analysis of this gap would strengthen the method.

  • Unclear scope of RL post-training. Section 3 states that GDPO is applied exclusively to listener frames. Yet Section 3.3 describes the reward as encouraging both speaker and listener dynamics, and the speaking metrics change substantially after post-training (e.g., speaking FDD drops from 39.54 to 22.45 in Table 7). It is therefore unclear whether the final model used in the benchmark comparisons applies RL to speaking segments, where lip motion should be largely determined by the audio. The authors should clarify this point.

  • Limited contribution on the speaking side. I could not identify a new contribution for speaking motion generation. The paper itself characterizes speaking head generation as a deterministic problem that supervised learning solves easily, and it does not appear to claim any specific methodological contribution for it. Given that the paper is framed as an interactive (speaker & listener) framework, a clearer statement of what the paper contributes to speaking motion would be helpful.

  • Potential bias of the main listening metrics toward the proposed reward. The listening evaluation primarily relies on FDD/PDD/JDD. These metrics are defined as the absolute difference between the standard deviations of the predicted and ground-truth frame-to-frame velocities. This is nearly identical to the velocity-variance term of the GDPO reward, which directly penalizes its standard deviation. The main listening comparisons therefore largely measure the quantity the proposed model was explicitly optimized for, which may favor the proposed method by construction. The additional metrics in Table 11 partially address this concern, but they are reported limitedly. I encourage the authors to place more weight on such metrics that are not part of the reward.

  • Leakage of ground-truth information through text prompts. At evaluation, the text prompts are generated by captioning the target actor's ground-truth video with Qwen3-VL (Appendix E.2). This injects ground-truth information about the target reaction into the input, making it difficult to consider the comparison with baselines fully fair. Notably, UniLS is evaluated with a zero style embedding precisely because its style embedding is derived from ground-truth motion (Section 4.1), whereas the proposed model receives target-derived captions. Without text prompts (Table 7), the listening FDD (22.19) becomes comparable to that of UniLS (22.23, Table 1), although PDD and JDD remain better, and the speaking FDD (33.02) falls behind ARTalk (31.64). The authors should report the main comparisons without target-derived captions, and clarify whether such captions are also used in Tables 2–5 and the user study.

Questions for authors
  • The paper's central motivation rests on two premises: (1) speaking facial motion is relatively deterministic, and (2) listening motion is inherently a one-to-many correspondence. Could the authors provide stronger evidence, either statistical or empirical, supporting these premises for this work?

  • What is the distinctive novelty of this work for speaking motion generation, or what specific efforts were made to achieve a meaningful performance gain over prior methods?

  • Please specify the FLAME version used and the dimensionality of each parameter group. In addition, could the authors clarify how eyelid closure is parameterized (e.g., through the FLAME expression coefficients or through additional eyelid blendshapes provided by the tracking pipeline)?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Uxod (1/2)

We thank the reviewer for their time and valuable feedback. Please find the responses to the raised concerns below:

1. Mismatch between the RL rollout sampler and the inference sampler (W1)

Recall that the ODE sampler at inference is xt+Δt=xt+Δtvθ(xt,t)x_{t+\Delta t} = x_t + \Delta t v_\theta(x_t, t). Flow-GRPO [R4] uses the SDE sampler xt+Δt=xt+Δt[vθ(xt,t)+σt22t(xt+(1−t)vθ(xt,t))]+σtΔtϵx_{t+\Delta t} = x_t + \Delta t\left[ v_\theta(x_t, t) + \dfrac{\sigma_t^2}{2t}\big(x_t + (1-t)v_\theta(x_t, t)\big) \right] + \sigma_t\sqrt{\Delta t}\epsilon in RL rollouts. Ours is xt+Δt=xt+Δtvθ(xt,t)+σΔtϵx_{t+\Delta t} = x_t + \Delta t v_\theta(x_t, t) + \sigma\sqrt{\Delta t}\epsilon, without the drift correction term.

We agree that this does not guarantee the same marginal distribution as the ODE. However, it does not affect the effectiveness of our RL post-training, for three reasons.

(1) Similar design exists in prior work. ReinFlow [R5] adopts the same form ak+1=ak+vθΔtk+σθ′ϵa_{k+1} = a_k + v_\theta\Delta t_k + \sigma_{\theta'}\epsilon, which also injects Gaussian noise around each Euler step without the drift correction. It treats the noisy denoising chain as a discrete time Markov process, proves a policy gradient theorem for it, and keeps only the finetuned velocity field after training. Although it targets a different area, it still shows that this is a valid way to apply RL to flow models.

(2) The ODE/SDE mismatch also exists in Flow-GRPO. The drift correction only guarantees the same marginal distribution between ODE and SDE in continuous time. It does not guarantee the same conclusion in discrete time, especially Δt\Delta t in SDE rollout in RL training is different than the one in ODE rollout. For example, Flow-GRPO collects rollouts with 10 SDE steps in RL training and uses 40 ODE steps at inference. Its Appendix D and Fig. 19 show that the few-step SDE samples are visibly different from the inference samples. However, this does not collapse optimization, because the group relative advantage only relies on comparisons within a group. So the same holds for our method.

(3) All results in our paper are obtained with the ODE sampler, and RL post-training brings consistent improvements (Tab. 7 and Tab. 8).

In summary, the missing drift correction does not affect the effectiveness of our RL post-training in our paper.

2. Scope of RL post-training (W2)

Thank you for pointing this out. The statement in Sec. 3 that GDPO is “applied exclusively to listener frames” is a writing mistake, and we have corrected it in the revised paper. GDPO post-training is applied to all frames, both speaking and listening, consistent with the reward description in Sec. 3.3.

Since our rewards are designed for motion dynamics, the speaking dynamic metrics also improve substantially (Tab. 7: FDD 39.54 → 22.45, PDD 3.50 → 2.71, JDD 1.81 → 1.26). In contrast, the lip-sync metric remains largely unchanged (Tab. 7: LVE 3.07 → 2.95), as lip motion is largely determined by the audio and is unaffected by RL post-training.

3. Contribution on the speaking side (W3 & Q2)

Our focus is on improving the naturalness of facial expression and head motion in both speaking and listening, rather than lip synchronization. As shown in Tab. 1, our method shows superior performance over baselines on facial expression and head motion dynamics in both speaking and listening conditions, and lip-sync accuracy is still preserved.

4. Bias of the listening metrics toward the rewards (W4)

Thank you for the helpful suggestion. Metrics that are not part of our reward are already reported in Tab. 11 in Appendix C.5, where GDPO post-training also brings consistent improvements. And we have also added a paragraph "Metrics Beyond the Rewards" in Sec. 4.1 of the revised main paper to emphasize them. Additionally, our user study in Tab. 5 shows positive human judgement on naturalness, liveliness and diversity, beyond the quantitative metrics.

5. Information leakage through text prompts (W5)

We agree with this concern. To remove the leakage, we generate text prompts from the input audio only, without access to the target video (details in Tab. 14 and Appendix E.4 of the revised paper), and re-evaluate on the Seamless Interaction test set:

MethodsSpeakingListening
LVE ↓MHD ↓FDD ↓PDD ↓JDD ↓FDD ↓PDD ↓JDD ↓
Ours w/ motion-derived text prompt2.951.0622.452.711.2618.851.900.87
Ours w/ audio-derived text prompt3.051.1022.832.731.3019.641.950.92
Ours w/o text prompt2.731.0433.022.971.4422.192.110.87

With audio-derived text prompts, our model performs comparably to motion-derived prompts. This shows that our improvements do not rely on information leakage.

AuthorsOpenReview ↗

Response to Reviewer Uxod (2/2)

6. Evidence for the two premises (Q1)

These two premises are supported by the statistical analysis in Fig. 2 in UniLS [R1], which is conducted on the same Seamless Interaction dataset. It shows that speaking motion is strongly correlated with the speaker’s own audio, while listening motion is only weakly correlated with the partner’s audio, indicating a one-to-many correspondence that leads to the “Regression-to-the-Mean” problem. We have added this to the introduction of the revised paper.

7. FLAME version and parameterization (Q3)

We use FLAME 2020. The 67-dimensional motion consists of expression (50), jaw pose (3), neck pose (3), eyelid (2), eye pose (6) and global rotation (3). Eyelid closure is not part of the FLAME expression coefficients. It is parameterized by two additional scalars (one per eye) estimated by SHeaP [R2], each scaling a per-vertex eyelid blendshape from MICA [R3] that is added to the FLAME vertices. Unlike prior dyadic methods, which typically predict only expression, jaw and neck, our model also predicts eyelid, eye pose and global rotation, enabling natural eye blinking and head nodding. We have added these details to Appendix F.1 of the revised paper.


[R1] Chu, X., Liu, R., Huang, Y., Liu, Y., Peng, Y., & Zheng, B. (2026). Unils: End-to-end audio-driven avatars for unified listening and speaking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 25142-25152).

[R2] Schoneveld, L., Chen, Z., Davoli, D., Tang, J., Terazawa, S., Nishino, K., & Nießner, M. (2025, October). Sheap: Self-supervised head geometry predictor learned via 2d gaussians. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 14162-14172). IEEE.

[R3] Zielonka, W., Bolkart, T., & Thies, J. (2022, October). Towards metrical reconstruction of human faces. In European conference on computer vision (pp. 250-269). Cham: Springer Nature Switzerland.

[R4] Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., ... & Ouyang, W. (2026). Flow-grpo: Training flow matching models via online rl. Advances in neural information processing systems, 38, 40783-40818.

[R5] Zhang, T., Yu, C., Su, S., & Wang, Y. (2026). Reinflow: Fine-tuning flow matching policy with online reinforcement learning. Advances in Neural Information Processing Systems, 38, 106282-106319.

Thank you for the detailed and thorough response. All of my questions and concerns have been addressed. Please make sure to incorporate these clarifications into the final version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Reviews and Responses

We thank all the reviewers for their feedback and comments. We summarize the status of the discussion and our revisions below.

  • Reviewer Uxod (rating 4): stated that all questions and concerns have been addressed.
  • Reviewer Tj9k (rating 3): maintained the rating after our response.
  • Reviewer iE6q (rating 2): maintained the rating after our response. We posted further clarifications to the follow-up comments.

To the best of our ability, we addressed all the reviewer comments in the following areas:

  • Scope of RL post-training (Reviewers Uxod, Tj9k) - Corrected the statement in Section 3.
  • Metrics beyond the reward (Reviewers Uxod, iE6q) - Added a paragraph in Section 4.1, referring to Appendix C.5 Table 11.
  • Information leakage through text prompts (Reviewer Uxod) - Added the evaluation with audio-derived text prompts in Table 7, Appendix E.4 and Table 14.
  • Comparison on the same representation (Reviewer iE6q) - Added the comparison with finetuned UniLS in Appendix C.8 Table 15.
  • Claims about UniLS (Reviewer iE6q) - Removed from Section 1.
  • FLAME version and parameterization (Reviewer Uxod) - Added the details in Appendix F.1.
  • Necessity of RL and validation of GDPO (Reviewer Tj9k) - Clarified with Appendix C Tables 8, 9, 10 and 12.
  • Sampler mismatch, reward design and novelty (Reviewers Uxod, iE6q, Tj9k) - Clarified in our responses.

We thank the reviewers for engaging in the discussion to improve our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.