acceptodds
Under review as a conference paper at ICLR 2027

OmniOPSD: Rationale-Privileged On-Policy Self-Distillation for Affective Computing

Abstract

Multimodal large language models (MLLMs) have advanced affective computing by enabling nuanced reasoning over visual, acoustic, and linguistic evidence. A key approach to improving their generalization is outcome-based reinforcement learning, which converts existing label-level annotations into rewards for final-answer correctness. However, its sparse sequence-level supervision provides limited fine-grained guidance for interpreting multimodal affective cues and may reward correct labels even when those cues are misinterpreted. To address this challenge, we introduce OmniOPSD, the first on-policy self-distillation framework for omni-modal affective reasoning. At its core, OmniOPSD transforms label-level supervision into evidence-rich, teacher-only rationales by leveraging reference affective judgments as semantic priors for offline multimodal evidence elicitation from frontier MLLMs. Through rationale-privileged self-distillation, it couples on-policy student rollouts with dense, evidence-informed token-level supervision from a local self-teacher, allowing the student to learn along its own trajectories rather than imitate fixed reasoning traces. This decoupling of offline evidence acquisition from local policy learning enables training without additional human annotation, frontier-model logits, or online frontier-model calls, while preserving student-only inference from the original task input. At both 3B and 7B scales, OmniOPSD outperforms supervised fine-tuning (SFT) and group relative policy optimization (GRPO) in mean macro-F1 on both in-domain and out-of-domain evaluations. Among the compared models, it ranks first by MER-UniBench mean across all three input-modality settings, reaching 84.13 with audio–video–text input, and achieves the highest MME-Emotion recognition score (Rec-S) of 42.4.

Primary area:
reinforcement learning
Keywords:
Affective Computing, Multimodal Large Language Models, Multimodal Emotion Recognition, Multimodal Intent Recognition, On-Policy Self-Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YkY7OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces OmniOPSD, a rationale-privileged on-policy self-distillation framework for multimodal affective understanding, covering emotion recognition, sentiment analysis, and intent recognition. Existing task annotations guide frontier models to generate multimodal descriptions of affective cues offline. These descriptions are paired with reference answers to form privileged context available only to the teacher during distillation. The student generates responses from the original multimodal inputs and task instructions, while a local teacher sharing its architecture and updated through an exponential moving average provides dense next-token distributions along the same student-generated prefixes. This design allows the student to learn along its own evolving affective reasoning trajectories, with the teacher supplying fine-grained, rationale-conditioned supervision. The distillation stage requires neither frontier-model logits nor online frontier-model calls, and inference uses only the student with the original task input. Experiments on MER-UniBench and MME-Emotion, together with controlled comparisons against SFT and GRPO at the 3B and 7B scales, provide evidence of the framework’s effectiveness across in-domain and out-of-domain affective tasks.

Critical strengths
  • A well-motivated framework for dense supervision of multimodal affective reasoning. OmniOPSD makes a clear methodological contribution by using multimodal affective rationales as teacher-only privileged information to supervise the student’s own reasoning trajectories. The distillation stage requires neither frontier-model logits nor online frontier-model queries. Sharing the architecture and tokenizer between the local teacher and student avoids the logit-access and tokenizer-alignment barriers of direct frontier-model distillation. EMA smooths the teacher parameters, while the reported time-to-target comparison demonstrates faster local training than OPD in the evaluated configuration.

  • Broad experimental coverage and informative analyses support the core design choices. The experiments examine both the effectiveness of OmniOPSD and how its supervision operates on affective reasoning.

  • The Label-only versus Label + Rationale comparison is particularly persuasive. Both teachers receive the correct answer, allowing the comparison to assess the additional value of affective evidence beyond answer conditioning alone.

  • This paper is clearly-written and well-organized.

Critical weaknesses

While the paper reports competitive recognition results across multiple benchmarks, the evaluation of reasoning quality relies primarily on automated metrics and model-based judgments. A systematic, independent assessment by affective computing experts would strengthen the comparisons with SFT, GRPO, and representative open- and closed-source models. In particular, the current evidence leaves some uncertainty about whether the performance gains are accompanied by better interpretation of auditory and visual cues, cross-modal integration, and affective inference.

Questions for authors
  1. Evidence of improved cue interpretation. Beyond the token-level diagnostics, are expert evaluations or other independent assessments available showing that OmniOPSD reduces errors in visual and acoustic cue descriptions relative to SFT, GRPO, or Label-only OPSD?

  2. Training comparability and data separation. Could the authors clarify the differences between the main-experiment and ablation protocols, particularly the training data, initialization, and training budgets? Which factors are controlled within each comparison, and what checks were performed to identify and address overlapping or near-duplicate samples across training and evaluation sources?

  3. Uncertainty in the ablation results. Are results from multiple training seeds available for the key comparisons, particularly Label-only versus Label+Rationale and OmniOPSD versus GRPO?

  4. Scope of the efficiency comparison. For Figure 4, could the authors specify the OPD teacher model and size, hardware allocation, batch and sequence settings, and timing boundaries? In particular, were comparable compute resources used, and does the 5.6-hour versus 17.7-hour comparison exclude offline rationale generation? These details are needed to interpret the reported approximately 3.2× speedup.

6 comments

We thank the reviewer for the constructive feedback. Our responses are as follows.

Q1: Evaluation of Multimodal Cue Interpretation and Affective Reasoning

A1 (Multimodal Interpretation and Reasoning). To assess multimodal evidence interpretation beyond recognition performance and token-level diagnostics, we directly evaluate the explanations generated by OmniOPSD, SFT, and GRPO on the same videos.

Evaluation protocol. We sample 40 videos from each of the eight MME-Emotion tasks, yielding 320 videos. Each explanation is assessed on six criteria: audio coverage, audio factual accuracy, visual coverage, visual factual accuracy, cross-modal cue integration, and reasoning linkage, which measures how well the cited evidence supports the affective conclusion. Each criterion is rated on a 0–2 scale, with higher scores indicating better performance.

For each explanation, we additionally compute the audio and visual cue scores as the minimum of the corresponding coverage and factual accuracy scores:

\qquad s_{\mathrm{visual}} = \min(c_{\mathrm{visual}}, a_{\mathrm{visual}}).$$ This rule requires both sufficient coverage and factual correctness for a high cue score. We compute the minimum for each explanation before aggregation and report results as mean ± standard deviation. | Dimension | OmniOPSD | SFT (7B) | GRPO (7B) | |---|---:|---:|---:| | Audio coverage | **1.54 ± 0.55** | 1.40 ± 0.57 | 1.36 ± 0.59 | | Audio factual accuracy | **1.28 ± 0.74** | 1.02 ± 0.78 | 1.13 ± 0.77 | | Audio cue score | **1.19 ± 0.70** | 0.90 ± 0.71 | 0.98 ± 0.70 | | Visual coverage | **1.50 ± 0.53** | 1.26 ± 0.60 | 1.13 ± 0.64 | | Visual factual accuracy | 1.14 ± 0.75 | 1.13 ± 0.78 | **1.22 ± 0.75** | | Visual cue score | **1.05 ± 0.71** | 0.91 ± 0.71 | 0.89 ± 0.68 | | Cross-modal cue integration | **1.25 ± 0.62** | 1.04 ± 0.61 | 1.07 ± 0.56 | | Reasoning linkage | **0.93 ± 0.73** | 0.78 ± 0.66 | 0.85 ± 0.64 | **Results and interpretation.** OmniOPSD achieves the highest mean in **seven of the eight dimensions**, including both derived cue scores. Audio factual accuracy improves by 0.26 over SFT and 0.15 over GRPO. The audio and visual cue scores also improve over both baselines, indicating better joint coverage and accuracy under the minimum-based scoring rule. Higher cross-modal integration and reasoning-linkage scores further support more effective use of multimodal evidence in the generated explanations. **Q2: Training Protocol Comparability and Train–Evaluation Data Separation** **A2 (Training Comparability and Data Separation).** We distinguish the main benchmark evaluation from the controlled comparisons, as these experiments use different training protocols and serve different purposes. **Main benchmark evaluation.** As described in Appendix E.2, the training data comprise the training splits of MELD, MIntRec 2.0, IEMOCAP, MC-EIU, CH-SIMS_v2_s, MAFW, MER2023, DFEW, CAER, BoLD, CMU-MOSI, and CMU-MOSEI. We first perform SFT cold-start training at a learning rate of $10^{-5}$. We then initialize OmniOPSD from the **SFT step-1,000 checkpoint** and train it for **one epoch, corresponding to 2,181 optimizer steps**, at a learning rate of $10^{-6}$. We evaluate the same resulting 7B checkpoint on both MER-UniBench and MME-Emotion. **Controlled comparisons.** As described in Appendix E.1, the controlled experiments examine both 3B and 7B models trained jointly on MELD, MIntRec 2.0, and IEMOCAP, using a common learning rate of $10^{-6}$. These experiments use the released training splits of MELD and MIntRec 2.0 for training and their test splits for evaluation. For IEMOCAP, Sessions 1–4 are used for training, while Session 5 is held out for evaluation. **Train–evaluation separation.** Both MER-UniBench and MME-Emotion construct their evaluation sets from the test or validation partitions of existing public datasets, whereas our training uses the training partitions. MER-UniBench follows the evaluation subsets listed in Appendix D.2. We also account for known overlap across dataset releases when selecting the training sources. Specifically, the MER2024 training set contains examples from the MER2023 test set that are included in MER-UniBench and MME-Emotion. **We therefore exclude MER2024 entirely from training** to prevent these evaluation examples from entering the training data. With these split assignments and dataset exclusions, the training samples and the reported evaluation samples are disjoint.

Q3: Multi-Seed Results and Uncertainty

A3 (Multi-Seed Reliability). We assess variation across training seeds through repeated runs of the main method comparison and the reasoning-mode ablation.

Main method comparison (Table 3).

MethodID Avg.OOD Avg.
SFT39.89 ± 0.8323.27 ± 0.51
GRPO38.16 ± 1.0523.56 ± 0.20
OmniOPSD40.67 ± 0.6424.67 ± 0.08

OmniOPSD achieves the highest mean on both ID and OOD evaluations. Compared with GRPO, the improvements are 2.51 points on ID and 1.11 points on OOD; compared with SFT, they are 0.78 and 1.40 points, respectively.

Reasoning-mode comparison (Table 4).

MethodReasoning modeID Avg.OOD Avg.
GRPOw/o Think38.55 ± 1.4523.67 ± 0.28
GRPOw/ Think38.20 ± 1.1223.10 ± 0.46
OmniOPSDw/o Think39.13 ± 2.0223.57 ± 0.80
OmniOPSDw/ Think40.62 ± 0.5824.66 ± 0.08

With reasoning enabled, OmniOPSD exceeds GRPO by 2.42 points on ID and 1.56 points on OOD. Within OmniOPSD, enabling reasoning improves the corresponding means by 1.49 and 1.09 points, respectively.

These multi-seed results are consistent with the experimental trends and conclusions reported in the paper and further suggest that OmniOPSD’s advantage is reproducible across independent training runs.

Q4: Scope and Configuration of the Efficiency Comparison

A4 (Efficiency Comparison). The approximately 3.2× speedup in Figure 4 concerns post-training wall-clock time to the reported target score, with rationales cached in advance. The student-side GPU allocation is matched, while the total GPU allocation differs because OPD uses an external teacher service.

Models and resource allocation. Both runs start from Qwen2.5-Omni-3B Base and use the same student-side settings, including four 80 GB GPUs and a per-GPU batch size of 1. OPD obtains top-64 logits from Qwen3-Omni-30B-A3B-Instruct, deployed on four additional 80 GB GPUs.

ConfigurationOmniOPSDOPD
Initial studentQwen2.5-Omni-3B BaseQwen2.5-Omni-3B Base
TeacherEMA self-teacherQwen3-Omni-30B-A3B-Instruct
Student-side GPU allocation4 × 80 GB4 × 80 GB
Additional teacher-service GPUsNone4 × 80 GB
Total GPU allocation4 GPUs8 GPUs
Per-GPU student batch size11

Target score and timing boundaries. OmniOPSD first reaches the OPD step-1,000 reference score at step 550; both scores round to 0.546 macro-F1. The corresponding plotted times are 5.6 hours for OmniOPSD and 17.7 hours for OPD, giving a wall-clock time ratio of 17.7/5.6≈3.217.7/5.6 \approx 3.2.

Offline rationale generation is excluded: OmniOPSD begins the measured post-training stage with the rationales already cached.

Figure 4 therefore supports a reduction in measured post-training time to the reported target score under the stated resource allocations.

I thank the authors for their response and additional results, e.g., the explanation-level evaluation and the results from multiple independent training runs provide valuable evidence, as well as the clarification of Figure 4, which substantially improves the interpretability of the efficiency claim. I keep my positive ratings, and want the authors to further solve my following concerns:

  1. Although the additional results further support the effectiveness of OmniOPSD, the multi-seed comparison between Label-only and Label + Rationale that I requested is still missing.
  2. While the ratings address my primary concern, further details of the rating procedure are needed: the background and number of the raters, whether they were blind to method identity.
Reviewer_HFxdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OmniOPSD addresses limited reasoning supervision in multimodal affective computing through rationale-privileged on-policy self-distillation. Offline GPT-4o rationales condition a local EMA self-teacher, which supplies token-level JSD targets on student-generated prefixes alongside answer and format rewards; inference requires only the student and original inputs. The paper evaluates benchmark performance and examines rationale conditioning, inference-time thinking, and prefix refresh.

Critical strengths

The approach offers a practical way to exploit existing labels while separating offline frontier-model evidence acquisition from local optimization. Two model scales, ID/OOD comparisons, and controlled rationale and thinking ablations support several design choices. The appendices clearly qualify the semi-gradient and post-hoc diagnostics, including a 600-response analysis with cluster bootstrap.

Critical weaknesses
  1. Table 12 reports gains of 1.38, 0.56, and 1.34 percentage points over label-only conditioning. However, answer-conditioned training rationales lack a reported independent factual audit, error rates, or quality-filtering results, and there are no same-label rationale swaps, generic-label explanations, or corrupted-rationale controls. Thus, sample-specific multimodal evidence has not been distinguished from answer-consistent narratives.

  2. Efficiency and reproducibility are incompletely documented. Figure 4 supports faster local optimization in its configuration, but does not fully specify the OPD teacher, matched hardware/GPU-hours, training budget, or costs of offline rationale acquisition and SFT cold start. The complete evidence-elicitation prompt and quality-control procedure are also missing, leaving end-to-end efficiency and the supervision pipeline insufficiently reproducible.

Questions for authors
  1. Were training rationales independently audited, and can you report the protocol and error rates together with same-label swaps or generic-explanation controls?

  2. What teacher, hardware, budget, and cost boundaries underlie Figure 4, and what is the total cost including rationale acquisition and cold start?

  3. What explains the 7B OOD losses, and do the small advantages persist across seeds or source-clustered paired confidence intervals?

  4. Can you provide label-independent evidence-quality measurements and reasoning-only/answer-only distillation results to isolate the contribution of process supervision?

2 comments

We thank the reviewer for the constructive feedback. Our responses are as follows.

Q1: Factual Quality and Specificity of Training Rationales

A1 (Rationale Quality and Specificity). We conduct a manual audit of 130 training rationales, randomly sampling 10 examples from each of 13 data sources. Each rationale is assessed against its corresponding video, with reference labels and model names hidden from the evaluators. Three dimensions are scored from 0–2.

DimensionCriterion for a score of 2Mean score
Factual groundingFactual descriptions are supported by the source material.1.93 / 2
Evidence specificityThe rationale identifies concrete, observable cues specific to the sample.1.87 / 2
Inference soundnessThe evidence reasonably supports the conclusion, with appropriate certainty.1.73 / 2

These ratings support the quality of the audited rationales, particularly their factual grounding and sample-specific evidence. The inference-soundness score further indicates generally reasonable connections between the cited evidence and affective conclusions.

Q2: Efficiency, Computational Cost, and Reproducibility

A2 (Efficiency, Cost, and Reproducibility). We clarify the Figure 4 comparison separately from offline rationale acquisition and the main-benchmark cold start.

Figure 4 configuration. Both runs start from Qwen2.5-Omni-3B Base, with four 80 GB student GPUs and a per-GPU batch size of 1. OPD additionally obtains top-64 logits from an external teacher deployed on four more 80 GB GPUs.

ConfigurationOmniOPSDOPD
TeacherEMA self-teacherQwen3-Omni-30B-A3B-Instruct
Student GPUs4 × 80 GB4 × 80 GB
Additional teacher GPUsNone4 × 80 GB
Steps to the reported reference score5501,000
Measured time5.6 h17.7 h
Student GPU-hours22.470.8
Total allocated GPU-hours22.4141.6†

†Assuming the four OPD teacher GPUs remain allocated throughout the 17.7-hour interval.

Both reported MELD scores round to 0.546 macro-F1. The time ratio is 17.7/5.6≈3.217.7/5.6 \approx 3.2, supporting an approximately 3.2× speedup in post-training to the reported target under these resource allocations.

Rationale acquisition. OpenFace selects frames at facial action-unit intensity peaks; Qwen2.5-VL-72B-Instruct extracts visual cues, and Qwen2-Audio-7B-Instruct extracts prosodic cues. GPT-4o then integrates the action-unit information, visual and acoustic descriptions, subtitles, and reference labels through the Chat Completions API.

For 24,023 examples from MELD, MIntRec 2.0, and IEMOCAP, the GPT-4o text-integration API cost is approximately US$60–90.

SFT cold start. Separately, the 7B main-benchmark cold start takes 50 min 55 s, approximately 6 GPU-hours, to reach the selected step-1,000 checkpoint.

Q3: Contribution of Reasoning-Span Supervision

A3 (Reasoning-Span Supervision). To examine the contribution of reasoning supervision during training, we additionally train three variants that differ only in the JSD mask over content tokens: Full supervises both reasoning and answer spans, Answer-only supervises answer content, and Reasoning-only supervises reasoning content. Format supervision and the reward objective remain identical.

Task (macro-F1, %)FullAnswer-onlyReasoning-only
MELD — Sentiment Analysis63.4361.8665.30
MIntRec 2.0 — Intent Recognition35.8633.1232.04
IEMOCAP — Emotion Recognition24.2822.4422.36
MC-EIU — Intent-ZH25.8321.9421.72
MC-EIU — Emotion-ZH27.9827.5825.28
MAFW — Emotion-SL20.9319.6315.47
MAFW — Emotion-ML23.9322.1217.80
ID mean (3 tasks)41.1939.1439.90
OOD mean (4 tasks)24.6622.8220.07

Full outperforms Answer-only on all seven tasks, with an ID mean improvement of 2.05 points, supporting the benefit of including reasoning-span supervision. It also outperforms Reasoning-only on six of seven tasks, with an ID mean improvement of 1.29 points. Reasoning-only performs best on MELD, while Full achieves the highest ID and OOD means.

These results support the complementary value of supervising reasoning and answer content under the shared format-supervision and reward objectives.

Q4: 7B OOD Performance and Generalization

A4 (7B OOD Performance and Generalization). Our initial ablation studies were conducted on Qwen2.5-Omni-3B. We subsequently transferred the selected hyperparameters directly to Qwen2.5-Omni-7B without a dedicated 7B hyperparameter search. Scale-specific sensitivity to these settings may therefore contribute to the OOD degradation and modest gains over GRPO reported in Table 3.

To further assess OOD generalization under matched conditions, we trained GRPO using the same main-benchmark training data and matched common training settings as OmniOPSD. Both models were evaluated with audio, video, and text inputs (AVT) under the same evaluation protocol on OV-MERD+, the fine-grained emotion recognition dataset included in MER-UniBench. OV-MERD+ is not included in our training data and serves as an OOD evaluation dataset in this study. The results are:

DatasetInputGRPOOmniOPSDDifference (pp)Pointwise paired clip-bootstrap 95% CI (pp)
OV-MERD+AVT57.9159.98+2.07[0.38, 3.80]

Scores are EW-F1 percentages. We computed the interval using 30,000 paired bootstrap replicates over the 532 evaluation clips, applying identical sampling weights to both methods and recomputing EW-F1 in each replicate. Conditional on the seed-42 checkpoints and the evaluation protocol, the interval excludes zero, providing evidence of an OmniOPSD gain over GRPO in this OOD evaluation under the assumption of independent clips.

For completeness, our responses to the other reviewers also include additional multi-seed evaluations of generalization, human assessments of OmniOPSD’s multimodal affective reasoning, and previously omitted experimental settings. Should you have any additional questions, we would be happy to discuss them further.

Reviewer_fK6KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Summary: The paper proposes OmniOPSD method, which uses GPT-4o offline to generate multimodal evidence rationales conditioned on the reference answer, and treats these rationales as teacher-only privileged context. An EMA self-teacher then performs full-vocabulary JSD distillation on rollout prefixes sampled by the student itself, optionally combined with answer/format rewards. At inference time, the student uses only the original multimodal input. On Qwen2.5-Omni-3B/7B, the authors report comparisons with SFT/GRPO, ID/OOD ablations, token-level diagnostics, and case studies across three modality settings on MER-UniBench and eight tasks on MME-Emotion.

Contributions: This paper introduces OmniOPSD, the first on-policy self-distillation framework for omni-modal affective reasoning, which leverages reference judgments as semantic priors to elicit multimodal evidence descriptions from frontier MLLMs offline and combines both into teacher-only privileged context rather than imitation targets; it formulates rationale-privileged on-policy self-distillation that couples on-policy student rollouts with dense token-level supervision from an EMA self-teacher conditioned on privileged context, requiring neither frontier-model logits nor additional human annotation, and at 3B and 7B scales it outperforms SFT and GRPO in mean in-domain and out-of-domain macro-F1.

Critical strengths
  1. This paper addresses a meaningful concern: label-level affective supervision is sparse, and outcome-based RL may reward responses that are "correct in answer but wrong in evidence".

  2. The paper's narrative and presentation are very clear.

Critical weaknesses
  1. The main concern on the novelty of this paper is that OmniOPSD's methodological skeleton substantially overlaps with recent work on privileged-context on-policy self-distillation. The paper itself cites the multimodal variants Vision-OPD [1], ViCuR [2], and [3], which already establish the recipe of sampling student rollouts on the original input, re-scoring the same prefixes with a teacher conditioned on privileged context, and applying token-level distillation on student-visited states. OmniOPSD's student rollout, teacher re-scoring, full-vocabulary JSD, and EMA teacher are standard components of this recipe; JSD and EMA are not new either. The genuine increment is to replace the privileged context (reference solution or visual cues) with an offline GPT-4o-generated, label-conditioned multimodal evidence rationale, and to apply the recipe to omni-modal affective reasoning. This is a useful domain transfer and system implementation, but it is an application/context transfer rather than a new learning method. In my view, both the novelty and the new knowledge that can be gained are very limited.

  2. The paper claims dense token-level supervision improves evidence interpretation, but the evaluation never measures it. Rec-S, HIT, and WAF are final-answer metrics; Rea-S improves by only 1.0 point and, since the judge sees the ground-truth label, 94.1% of reasoning steps are judged correct when the answer is correct versus 35.3% when wrong. So, Rea-S largely mirrors the final prediction. The JSD diagnostic only shows the teacher's distribution shifts, not that it becomes more faithful or correct. The paper thus infers "better evidence interpretation" from "better final answers" without any mediating measure; a more accurate claim is that final-answer accuracy improves, while evidence interpretation remains weakly supported.

References

[1] Yuan Q, Lou J, Yu X, et al. Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation[J]. arXiv preprint arXiv:2605.18740, 2026.

[2] Tian K, Liu S, Yan Z, et al. Vicur: Visual cues as recoverable privilege for multimodal on-policy distillation[J]. arXiv preprint arXiv:2606.05718, 2026.

[3] Wang S, Liu X, Liu L, et al. Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation[J]. arXiv preprint arXiv:2606.19120, 2026.

Questions for authors
  1. The combined objective in Eq. (8) is L=LD+αLRL = L_D + αL_R, with α=0.1α = 0.1 for the main benchmarks and α=0.2α = 0.2 for the ablations, yet no table reports the α=0α = 0 condition in which only token-level distillation is used. As a result, the reader cannot determine how much of the headline gains, 84.13 on MER-UniBench and 42.4 Rec-S on MME-Emotion, comes from distillation and how much comes from the answer/format reward, nor whether the central claim that "dense token-level supervision" independently improves performance is actually supported. The OPSD baseline in Figure 4 is not a substitute: it is a separate configuration, evaluated only on MELD, and is not necessarily equivalent to α=0α = 0. The authors should add α=0α = 0 (distillation only) and reward-only controls on the main benchmarks, trained and evaluated under the same protocol, to isolate the contribution of distillation.

  2. In the paper, GPT-4o serves simultaneously as the rationale generator and the MME-Emotion judge, which may introduce self-preference; this setup is unreasonable.

5 comments

We thank the reviewer for the constructive feedback. Our responses are as follows.

Q1: Contribution and New Empirical Insights

A1 (Contribution and Empirical Insights). We build on established OPSD, JSD, and EMA components. Our contribution concerns supervision for affective learning when outcome labels are available but the supporting multimodal evidence is unannotated. OmniOPSD uses existing labels and original inputs to construct offline rationales that condition a local self-teacher during on-policy updates.

The central empirical finding is that rationale context adds value beyond the reference label: Appendix Table 12 retains the same labels in both conditions and reports improvements across all three evaluated tasks when rationales are added. The analyses in A2 and A3 further examine the quality of generated explanations and the role of reward optimization.

After rationale caching, optimization requires no frontier-model logits or online frontier-model calls; inference uses only the student. Offline generation remains a separate cost. We present these findings and practical properties explicitly while distinguishing them from the inherited algorithms.

Q2: Direct Evaluation of Explanation Quality

A2 (Explanation Quality). Answer accuracy and teacher-distribution shifts alone do not establish better evidence interpretation. We therefore conduct two distinct expert assessments.

320-video comparison. Experts evaluate OmniOPSD, SFT, and GRPO on 40 videos from each MME-Emotion task. Six criteria—audio/visual coverage and factual accuracy, cross-modal integration, and evidence-to-conclusion linkage—are scored from 0–2. Two additional cue scores take the minimum of the corresponding coverage and accuracy scores. OmniOPSD achieves the highest mean in seven of eight dimensions, including integration and evidence-to-conclusion linkage.

160-question comparison. Three affective-computing experts assess explanations against the original audio and video. Scores indicate fully satisfied (2), partially satisfied (1), incorrect or insufficiently supported (0), and missing content (−1); inapplicable integration items receive 0.

ModelVisual accuracyAcoustic accuracyKey-cue coverageMultimodal integrationInference plausibility
OmniOPSD1.1591.1091.3361.0391.082
Qwen2.5-Omni-7B1.493−0.6240.7050.0840.155
AffectGPT0.3290.3970.2240.2350.170
Emotion-LLaMA−0.276−0.286−0.515−0.374−0.340
VideoLLaMA0.614−0.2640.3120.1230.410
VideoLLaMA20.105−0.309−0.346−0.443−0.225
OPPO0.3650.6210.6920.5680.659
Gemini 3.8 Flash0.9800.8030.9610.8680.961

OmniOPSD achieves the highest expert-rated mean scores in four of the five dimensions: acoustic accuracy, key-cue coverage, multimodal integration, and inference plausibility. These results highlight complementary strengths in interpreting acoustic cues, covering relevant affective evidence, integrating information across modalities, and forming plausible affective conclusions. Together, they support the quality and completeness of OmniOPSD’s explanations on the evaluated samples.

Full-benchmark comparison. We also re-evaluate Qwen2.5-Omni-7B Base on all eight tasks and 6,500 examples, matching OmniOPSD’s video-frame count and audio-input procedure on the same hardware.

ModelRec-S (%)Rea-S (%)CoT-S (%)
Qwen2.5-Omni-7B Base36.9854.9645.97
OmniOPSD42.4560.2851.36

The matched-input Rea-S gain is 5.32 points, compared with the original 1.0-point gain over the published baseline.

Q3: Contributions of Distillation and Reward Optimization

A3 (Distillation and Reward Optimization). The OPSD baseline in Figure 4 omits frontier-generated rationales, whereas the α=0\alpha=0 control requires the same rationale-conditioned teacher as the full model.

All conditions start from the same Qwen2.5-Omni-7B SFT step-1,000 checkpoint. Full (LD+0.1LR\mathcal{L}_D+0.1\mathcal{L}_R) and distillation-only (LD\mathcal{L}_D) share the data, teacher context, on-policy rollouts, JSD, EMA, learning rate of 10−610^{-6}, and 2,181 updates over one epoch.

Reward-only (0.1LR0.1\mathcal{L}_R) disables distillation and teacher operations. Its initial 10−610^{-6} run produces malformed outputs and reward collapse. The reported run uses 10−710^{-7} from the same initialization, without auxiliary KL or SFT losses.

All models use the same MER-UniBench AVT evaluation protocol. Full denotes the newly trained control run.

TaskMetricFullDistillation-onlyReward-only†
CMU-MOSIWAF88.3689.1786.28
CMU-MOSEIWAF88.5488.4985.12
CH-SIMSWAF91.5189.7685.01
CH-SIMS v2WAF90.1388.7085.30
MER2023HIT83.6583.1678.58
MER2024HIT91.7792.1984.90
MELDHIT72.5672.7968.22
IEMOCAPHIT89.5789.3885.04
OV-MERD+EW-F160.5660.6956.22
Nine-task mean84.0783.8279.41

†Lower-learning-rate run; supplementary comparison.

The distillation-only model achieves near-full performance on MER-UniBench, highlighting the effectiveness of our online self-distillation with fine-grained token-level supervision in this configuration.

Q4: Sensitivity to the Evaluator

A4 (Evaluator Sensitivity). Using GPT-4o for both rationale generation and evaluation risks self-preference. We rescore fixed outputs on 320 randomly selected MME-Emotion examples using the benchmark protocol with four judges. Each cell below reports OmniOPSD / OPPO; each judge scores 320 outputs per model.

JudgeRec-SRea-SCoT-S
GPT-4o36.88 / 18.4457.52 / 52.3547.20 / 35.39
GPT-6.1 Sol35.00 / 20.0068.94 / 66.9651.97 / 43.48
Gemini 2.5 Flash40.31 / 29.3854.11 / 44.5547.21 / 36.96
Gemini 3.8 Flash37.19 / 22.1972.04 / 63.4954.62 / 42.84

Although absolute scores vary across judge models, the comparative trend remains consistent: OmniOPSD outperforms OPPO on all three metrics under all four judges, with CoT-S advantages of 8.49–11.81 points. Together with the human assessments in A2, this indicates that the observed advantage on the evaluated samples is maintained across evaluator choices and is not specific to GPT-4o scoring.

AuthorsOpenReview ↗

Follow-up: Additional Experiments and Clarifications

Thank you again for your constructive feedback. We clarified OmniOPSD’s contribution as, to our knowledge, the first OPSD framework for omni-modal affective computing, offering a simple pipeline without additional human evidence annotation (A1). We added expert assessments to directly evaluate multimodal evidence and explanation quality (A2), distillation-only and supplementary reward-only comparisons on MER-UniBench to examine their respective contributions (A3), and multi-judge evaluations to assess whether the observed advantages depend on GPT-4o scoring (A4). Our responses to the other reviewers also provide additional experimental results, detailed settings, and expert assessments for reference. We would be grateful for your thoughts on whether these additions address your concerns, and would welcome any further questions or discussion.

Comments on the paper

2 comments

Thank you again to all reviewers for your time and thoughtful feedback. As the discussion period is approaching its deadline, we would greatly appreciate it if you could let us know whether there are any remaining questions or concerns regarding our response or the paper. We are happy to provide further clarification, additional details, or supporting evidence as needed. We hope our rebuttal has adequately addressed the points raised, and we sincerely appreciate your consideration.

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive feedback. We summarize the main feedback, additional evidence, and discussion outcomes from this rebuttal for your consideration.

First, the reviews recognized the value of our research motivation, practical framework, and presentation. YkY7 and fK6K acknowledged the importance of addressing insufficient evidence supervision in affective reasoning. YkY7 and HFxd recognized the practical value of using existing labels and separating offline evidence acquisition from local optimization, together with the experimental support for our design choices. fK6K and YkY7 also praised the clarity of the paper.

Second, we added experiments and clarifications addressing specific reviewer concerns.

  • Supervision mechanisms and component contributions: To address fK6K’s question about the contributions of distillation and rewards, we added matched Full and distillation-only experiments, alongside a supplementary reward-only comparison. For HFxd’s concern about reasoning supervision and YkY7’s request for training-stability evidence, we added reasoning-/answer-span supervision ablations and multi-seed results for key comparisons. These findings further support the value of evidence conditioning and fine-grained reasoning supervision.
  • Explanation quality and evaluation reliability: To address fK6K’s and YkY7’s requests for direct measurements of evidence interpretation quality, we added expert assessments of cue accuracy, coverage, cross-modal integration, and evidence-to-conclusion linkage. For HFxd’s concern about training-rationale reliability, we audited sampled rationales for factual grounding, sample specificity, and inference soundness. To examine fK6K’s concern about evaluator preference, we rescored fixed outputs using four judges. OmniOPSD’s advantage over OPPO was maintained across all reported metrics under every judge.
  • Efficiency and experimental transparency: To address HFxd’s and YkY7’s questions about Figure 4, we clarified the teacher configuration, hardware allocation, timing boundaries, and costs associated with offline evidence acquisition. We also clarified training protocols and train–evaluation splits for YkY7’s comparability and data-separation concerns. These details support faster post-training to the target performance with reduced external-teacher resource requirements in the evaluated configuration, with rationales cached beforehand.

Third, the additional evidence received positive feedback during the discussion. YkY7 acknowledged the explanation-level evaluation, independent training results, and Figure 4 clarifications, and raised their rating after receiving further controlled results and human-evaluation details. HFxd did not participate in the follow-up discussion. We would emphasize that OmniOPSD uses multimodal affective rationales as teacher-only privileged information to provide dense, token-level supervision along the student’s own sampled trajectories for omni-modal affective reasoning.

We welcome specific remaining questions about the experiments, evaluation protocols, or interpretation of the results, and respectfully ask that the contribution be assessed in light of the additional evidence and discussion.

Thank you again for your time, careful assessment, and constructive engagement.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.