acceptodds
Under review as a conference paper at ICLR 2027

AutoCAST: Automated Character-Attributed Speech Transcription in Long-Form TV Dramas

Abstract

Automated named-speaker transcription aims to determine the start and end times, text, and speaking character of each utterance in long-form TV dramas. This task remains challenging for existing audio-visual models: multiple visible characters, off-screen speech, and reaction shots create visual ambiguity, while the lack of reliable character-specific voice references and limited speaker discrimination complicate acoustic attribution. To address these challenges, we propose AutoCAST (Automated Character-Attributed Speech Transcription), a three-stage framework that takes raw audiovisual streams, a cast roster, and character portraits as input to produce timestamped, character-attributed transcripts. Timestamp-aware transcription first predicts utterance boundaries and text using explicit temporal markers. Agreement-guided reference bank construction then selects candidate speech segments whose initial identity predictions agree with face observations to build a character-specific voice reference bank. Finally, reference-guided multimodal attribution performs acoustic seed filtering and label propagation, then combines character-level voiceprint scores with visual and dialogue context to determine the speaker of each utterance. The complete framework requires neither manual transcripts nor manually verified voice references at inference. On DramaSR-532K, AutoCAST achieves 91.70% mean speaker-attribution accuracy across seven held-out Chinese-language dramas, outperforming the strongest evaluated closed-source baseline (83.06%) by 8.64 percentage points, and achieves 87.27% across three held-out English-language series. Ablation studies support the effectiveness of automatic segmentation, reference bank construction, and multimodal speaker attribution. Code is available at https://anonymous.4open.science/r/AutoCAST/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal learning, audio-visual speaker recognition, speaker-attributed transcription, long-form video understanding, speaker diarization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uWvSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses named-speaker transcription in long-form TV dramas: given raw video and audio, a cast roster, and character portraits, the system outputs the start and end times, transcript, and speaking character of each utterance. The proposed pipeline, AutoCAST, first uses a fine-tuned Qwen3-Omni with interleaved integer-second markers to transcribe and segment the audio. It then builds per-character voice reference banks automatically, keeping an utterance only if its predicted speaker is detected by face recognition nearby. Finally, voiceprint label propagation adopted from DramaSR-LRM yields candidate speakers with similarity scores, and a fine-tuned Qwen3-VL-32B combines these scores with face-annotated frames and dialogue context to decide the speaker. Trained on two Chinese dramas from DramaSR-532K, AutoCAST achieves 91.70% and 87.27% attribution accuracy on seven held-out Chinese dramas and three English series, versus 83.06% and 83.74% for the best general-purpose baselines. Ablations show that automatic segmentation and bank construction come within about 0.6 points of their manually annotated counterparts.

Critical strengths

Controlled ablations directly test the main claim that the manual inputs required by DramaSR-LRM can be replaced. With the other components fixed, the automatically built bank reaches 91.70% versus 92.27% for a manually labeled bank (Table 3), and predicted segmentation reaches 92.27% versus 92.92% for GT boundaries (Table 2b). The evaluation covers about 400 hours from ten held-out series in two languages, with a strict end-to-end metric and alignment and scoring rules specified in Appendix A. On this data, the best general-purpose baselines reach 83.06% (Chinese) and 83.74% (English), compared with 91.70% and 87.27% for AutoCAST.

Critical weaknesses

W1. No downstream evaluation. The introduction motivates the task with story understanding and video QA (citing MICap and TVQA), but no experiment tests whether AutoCAST transcripts improve these or other downstream tasks, so the practical benefit of the outputs is not quantified. W2. Single evaluation domain. All test data are professionally produced TV dramas; no other recording conditions (e.g., meetings, interviews, user-generated video) are evaluated. With GT segmentation and a manually labeled bank, removing voiceprint scores lowers accuracy from 92.92% to 76.55%, and removing images lowers it to 91.49% (Table 4). How voiceprint propagation behaves under noisy or far-field audio is not examined. W3. The Table 1 baselines are general-purpose models with no task-specific adaptation described, whereas AutoCAST is fine-tuned on two DramaSR-532K dramas and pools voice references across each entire drama, including later clips (Appendix C.1).

Questions for authors

Plz refer to the Weaknesses.

2 comments

We thank the reviewer for the careful assessment and constructive comments. We provide additional experiments and clarifications on downstream utility, cross-domain evaluation, and baseline comparisons below. 

Q1.Downstream utility W1W1. To directly evaluate the downstream benefit of AutoCAST, We evaluated story understanding using 896 QA pairs constructed from the seven held-out Chinese dramas, with Qwen3-VL-8B throughout (Table 1). Adding AutoCAST-generated named-speaker transcripts, with automatic segmentation and reference-bank construction, improves accuracy from 39.06% to 66.07% +27.01percentagepoints+27.01 percentage points. This is 0.56 points below the ground-truth-speaker condition 66.6366.63%. These results provide direct evidence that AutoCAST outputs improve story QA performance.

Q2.Evaluation beyond TV dramas W2W2. Without meeting-specific fine-tuning, AutoCAST transfers directly from TV dramas to AMI meeting recordings, achieving 25.40% DER (Table 2). This slightly outperforms TFAF 25.6225.62% and substantially outperforms the evaluated general-purpose models 60.7860.78%–87.06%, demonstrating transfer beyond the training domain.

Q3.Task adaptation and comparison fairness W3W3. We fine-tuned Qwen3-Omni on the same training data as AutoCAST. On Minning Town, AutoCAST achieves 90.97% attribution accuracy versus 72.18% for the adapted baseline +18.79points;Table3+18.79 points; Table 3.

We also compared against CineSRD and Character-aware under matched inputs and evaluation protocols, giving both methods in each comparison the same full-drama access, including later clips. With GT segmentation and automatically clustered reference banks, AutoCAST achieves 88.07% versus CineSRD's 60.89%; in the end-to-end setting, it achieves 90.97% versus Character-aware's 42.32%.

Table 1. Story QA results on 896 DramaVQA questions from seven held-out Chinese dramas. All conditions use Qwen3-VL-8B.

MetricImages OnlyImages + AutoCAST-Predicted Named Speakers + DialogueImages + Ground-Truth Speakers + Dialogue
Correct / Total (Accuracy ↑)350/896 39.0639.06%592/896 66.0766.07%597/896 66.6366.63%

Table 2. Speaker diarization results and error decomposition on the AMI meeting dataset.

MethodDER ↓Missed Speech ↓False Alarm ↓Speaker Confusion ↓
AutoCAST25.40%21.682.071.65
TFAF25.62%22.162.960.50
Gemini 3.5 Flash72.32%15.366.8250.13
Doubao60.78%13.9314.6532.20
Qwen3-Omni87.06%48.587.9330.55

Error components are contributions to DER in percentage points. Their sum may differ slightly from the reported DER due to rounding.

Table 3. End-to-end comparison with a task-specifically fine-tuned baseline on Minning Town.

MethodAttribution Acc. ↑CER ↓SSGC ↑AAS ↓
AutoCAST90.97%9.57%96.25%0.18 s
Qwen3-Omni task−specificfine−tuningtask-specific fine-tuning72.18%15.11%91.87%0.55 s
Qwen3-Omni55.83%32.77%61.78%4.90 s

Thank you again for your thoughtful comments. We have carefully addressed your concerns in our rebuttal and would greatly appreciate it if you could take a moment to review our responses.

We would be happy to provide further clarification or additional evidence if any of our responses remain unclear. Thank you for your time and consideration!

Reviewer_wyk1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies named-speaker transcription for long-form TV dramas, jointly producing utterance boundaries, transcripts, and character names from audiovisual streams, cast rosters, and portraits. It proposes AutoCAST, which combines timestamp-aware transcription, agreement-filtered automatic voice-reference banks, and reference-guided multimodal attribution. Experiments on DramaSR-532K evaluate end-to-end performance, reference-bank construction, segmentation, and modality ablations.

Critical strengths
  • The paper is well-organized and clearly presented. The overall framework and its three stages are easy to follow, and the figures effectively illustrate the proposed method and its key components.
  • The ablation studies provide support for the effectiveness of the proposed components, demonstrating the contributions of reference bank construction and voiceprint-guided multimodal speaker attribution.
Critical weaknesses
  • The proposed framework primarily integrates existing techniques, including timestamp-aware transcription, face-based reference filtering, acoustic label propagation, and multimodal LLM-based speaker attribution, without introducing substantial new algorithmic innovations.
  • The paper primarily compares AutoCAST with general-purpose multimodal models, while lacking end-to-end comparisons with existing methods specifically designed for character-aware audiovisual subtitling and speaker attribution. This makes it difficult to assess the actual improvements over prior task-specific approaches.
  • AutoCAST requires a predefined cast roster and reference portraits, which may not always be available in real-world scenarios, limiting its level of automation and practical applicability. In contrast, CineSRD (CVPR 2026) performs speaker registration and diarization without requiring such information, raising questions about the necessity of these assumptions. CineSRD is also a closely related work that should be discussed and compared with in the paper.
Questions for authors
  • Q1. Could the authors further clarify the key methodological innovations of AutoCAST beyond the integration of existing techniques? In particular, what distinguishes the proposed framework from previous multimodal speaker attribution pipelines?
  • Q2. How does AutoCAST perform when predefined cast rosters or reference portraits are unavailable or incomplete? Given that CineSRD (CVPR 2026) performs speaker registration and diarization without such information, could the authors clarify the necessity of these inputs and discuss the differences between the two approaches?
4 comments

We thank the reviewer for highlighting the need to clarify our methodological contribution, compare with task-specific approaches, and examine dependence on cast information.

Q1. Methodological contribution. AutoCAST addresses speaker identification across long-form dramas when visual, acoustic, and contextual cues are ambiguous or conflicting. Our contribution is a fully automated pipeline that acquires voice references from video and jointly interprets these complementary cues to produce timestamped, speaker-attributed transcripts, without manual transcripts or verified voice references. Cast rosters and portraits are optional: automatic face clustering supports anonymous identities Q2Q2. The complete pipeline substantially outperforms direct multimodal-model inference in our experiments.

  • Context and agreement guide reference selection. DramaSR-LRM forms candidates through visual neighborhoods; Character-aware uses audiovisual synchronization and acoustic filtering; CineSRD constructs initial prototypes through visual–acoustic clustering. AutoCAST makes contextual identity prediction and prediction–face agreement jointly determine reference admission.

  • Controlled ablations support both components. We fix predicted segmentation, the speaker encoder, propagation backend, and final attribution checkpoint. Removing identity initialization reduces SeedQual from 90.89% to 69.03% and final accuracy from 91.70% to 88.21%. Removing face agreement, while retaining the same initialization model, reduces them to 84.67% and 89.75%, respectively. These results support the contributions of both identity proposals and agreement-based admission.

  • Automatic references approach manually labeled performance. Under the same predicted segmentation, the automatic bank achieves 91.70% accuracy, versus 92.27% with a manually labeled bank. This 0.57-point gap shows that the proposed acquisition procedure approaches manually labeled reference performance without requiring manual verification of acoustic references at inference.

Comparison with CineSRD. CineSRD addresses speaker registration and diarization but does not produce transcripts. We therefore compare it with AutoCAST on their shared speaker-attribution and diarization tasks, using the same GT segmentation and full-drama access for both methods. AutoCAST uses automatic face clustering without predefined cast rosters or portraits. While CineSRD relies on threshold- and rule-based decisions, AutoCAST can also use weaker visual and voiceprint cues when supported by complementary evidence and dialogue context. Against the stronger evaluated CineSRD variant, without voting, AutoCAST achieves 87.98% attribution accuracy and 9.08% DER, compared with 83.99% and 14.29%(Table 4). Attribution accuracy improves on six of seven dramas, and DER decreases on all seven.

This controlled comparison demonstrates the effectiveness of AutoCAST's identity-processing pipeline with utterance boundaries held fixed and predefined cast information removed. Separately, we evaluate complete named-speaker transcription against D-ORCA, Character-aware, and task-specifically fine-tuned Qwen3-Omni. These two sets of experiments evaluate the shared identity-processing tasks and the complete transcription task, respectively.

Q2. Dependence on cast information. In practice, a face gallery for the main cast can often be readily assembled from cast names in end credits and portraits retrieved through web searches. Such a gallery is helpful but not required: without predefined cast rosters or portraits, AutoCAST can cluster faces directly from video and attribute speech to anonymous identities.

Under GT segmentation, replacing predefined visual references with automatic clustering reduces mean acoustic-reference coverage from 97.43% to 94.20% and attribution accuracy from 90.01% to 87.98% (Table 4). Cast information therefore improves coverage and enables explicit character-name grounding, rather than being a prerequisite for speaker attribution. This experiment evaluates complete metadata removal; robustness to partially missing or incorrect cast entries remains to be quantified.

Named outputs also support downstream use: we evaluated story understanding using 896 QA pairs constructed from the dramas, the AutoCAST-named transcript condition achieves 66.07% accuracy versus 60.49% for the ground-truth anonymous-speaker condition, using Qwen3-VL-8B throughout (Table 3).

Additional task-specific comparisons. We added end-to-end evaluations of D-ORCA-8B, Character-aware, and Qwen3-Omni fine-tuned on the same training data. On Minning Town, AutoCAST achieves 90.97% attribution accuracy, compared with 65.56%, 42.32%, and 72.18%, respectively, while also obtaining the best CER, SSGC, and AAS among these evaluated methods (Table 1). We will incorporate the related-work discussion, experimental protocols, and limitations into the revision.

We additionally evaluated MOSS and VibeVoice across seven dramas (Table 2). Across seven dramas, AutoCAST and MOSS perform similarly under one-minute evaluation 91.3291.32% vs. 91.22%, while AutoCAST leads at episode level (87.98% vs. 66.87%; Table 2). MOSS and VibeVoice receive optimal speaker-ID alignment per evaluation unit, whereas AutoCAST directly predicts character names.

We also compare AutoCAST with other methods on additional public datasets without further fine-tuning; see our response to Reviewer 3, “Q2.Evaluation beyond TV dramas W2W2.”

Table 1. End-to-end comparison on Minning Town.

MethodAttribution Acc. ↑CER ↓SSGC ↑AAS ↓
AutoCAST90.97%9.57%96.25%0.18 s
D-ORCA-8B (original prompt)65.56%32.39%69.88%1.01 s
Qwen3-Omni55.83%32.77%61.78%4.90 s
Qwen3-Omni task−specificfine−tuningtask-specific fine-tuning72.18%15.11%91.87%0.55 s
Character-aware42.32%18.44%37.76%0.71 s

Table 2. Attribution accuracy across seven Chinese-language test dramas with different processing units.

MethodOne-minute evaluation ↑Episode-level evaluation ↑
AutoCAST91.32%87.98%
MOSS91.22%66.87%
VibeVoice71.90%—

Audio-only speaker IDs are Hungarian-matched to GT labels per evaluation unit. AutoCAST directly predicts character names and, by design, retains one-minute inference windows in both settings.

VibeVoice does not consistently produce reliable episode-level outputs in our evaluation.

Table 3. Downstream question-answering results on 896 QA pairs from seven dramas, using Qwen3-VL-8B for all conditions.

MetricImages OnlyImages + Ground-Truth Anonymous Speakers + DialogueImages + AutoCAST-Predicted Named Speakers + DialogueImages + Ground-Truth Speakers + Dialogue
Correct / Total (Accuracy ↑)350/896 39.0639.06%542/896 60.4960.49%592/896 66.0766.07%597/896 66.6366.63%

Table 4. Comparison of predefined visual references and automatic face clustering, alongside CineSRD with and without voting.

DramaPredefined Bank CoverageClustered Bank CoverageAutoCAST (Predefined) Acc. ↑AutoCAST (Clustered) Acc. ↑AutoCAST (Clustered) DER ↓CineSRD Acc. ↑CineSRD DER ↓CineSRD w/o Voting Acc. ↑CineSRD w/o Voting DER ↓
The Long Night100.00% 37/3737/3789.19% 33/3733/3788.72%86.12%10.68%63.25%35.16%82.94%15.59%
Qin Empire 295.00% 57/6057/6095.00% 57/6057/6088.14%86.60%10.70%59.64%39.72%87.31%10.71%
Ode to Joy 193.22% 55/5955/5989.83% 53/5953/5994.12%91.41%5.36%63.08%36.34%85.96%13.31%
The Knockout97.33% 73/7573/7596.00% 72/7572/7590.71%88.94%8.21%62.70%36.00%84.08%13.96%
Minning Town96.47% 82/8582/8589.41% 76/8576/8588.96%88.07%10.06%60.89%36.86%79.93%18.08%
Stand by Me 1100.00% 30/3030/30100.00% 30/3030/3093.90%90.99%6.27%66.30%33.34%90.47%9.63%
Battle of Changsha100.00% 34/3434/34100.00% 34/3434/3485.55%83.72%12.27%55.57%42.38%77.27%18.72%
Overall97.43%94.20%90.01%87.98%9.08%61.63%37.12%83.99%14.29%

Coverage measures the proportion of characters represented in the acoustic reference bank constructed from the corresponding visual bank. Overall coverage is macro-averaged across dramas. AutoCAST results use ground-truth segmentation.

Thank you again for your thoughtful comments. We have carefully addressed your concerns in our rebuttal and would greatly appreciate it if you could take a moment to review our responses.

We would be happy to provide further clarification or additional evidence if any of our responses remain unclear. Thank you for your time and consideration!

Thank you for the detailed response and the additional experiments. The new comparisons with task-specific methods and the analysis of dependence on cast information partially address my concerns.

Regarding novelty, I view this as primarily a system/engineering contribution. The response does not dispute that the framework mainly integrates existing techniques. It demonstrates that this integration is effective. To be clear, I do not discount pipeline-style work, and I acknowledge the practical value of the system. However, I would leave the final judgment on whether the level of methodological novelty meets the bar of the venue to the AC.

I will maintain my original rating.

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

AutoCAST addresses automated named-speaker transcription in long-form TV dramas. Given raw audiovisual streams, a cast roster, and character portraits, it predicts utterance boundaries, text, and character identities. Its three stages are Timestamp-Aware Transcription, Agreement-Guided Reference Bank Construction, and reference-Guided Multimodal Attribution. On DramaSR-532K, it reports mean speaker-attribution accuracy of 91.70% across seven held-out Chinese-language dramas and 87.27% across three English-language series, with task-specific training on only two Chinese-language dramas. The method uses an offline, transductive protocol with references pooled across the target drama.

Critical strengths
  1. The complete pipeline improves speaker-attribution accuracy over the strongest evaluated closed-source baseline by 8.64 percentage points on Chinese-language dramas and 3.53 points on English-language series. It requires neither manual transcripts nor manually verified voice references during inference stage. The English results also support cross-language generalization within the evaluated series.

  2. The automatic reference bank achieves 91.70% final accuracy. The ablations keep predicted segmentation and the downstream attribution model unchanged. Accuracy drops to 89.75% without prediction-face agreement and to 88.21% without identity initialization. Under the same predicted segmentation, the manually labeled reference bank achieves 92.27%, only 0.57 percentage points higher than the automatic bank.

Critical weaknesses
  1. Table 2(a) does not isolate the contribution of explicit temporal markers from task-specific fine-tuning. A comparison which show the difference between using the same base model and training settings with and without the markers is not conducted. MOSS also has lower CER and AAS, although AutoCAST has higher SSGC. The paper does not report downstream attribution with MOSS under the same pipeline. This causes the contribution of the transcription design insufficiently established.

  2. Characters without references receive no acoustic scores, and the final prompt asks the model to choose from at most twenty candidates. Although SeedQual assigns zero to characters without seeds, the paper does not report how often the correct character is absent from the candidate list. The appendix also specifies error conditions, but does not report their frequency or how any affected cases enter the evaluation. These details concern about accessing the coverage of the automatic pipeline.

Questions for authors
  1. With the base model, training data, and training settings fixed, how does removing temporal markers affect CER, SSGC, AAS, and final speaker-attribution accuracy?
  2. How many characters have no seeds, and how often is the correct character included in the final candidate list?
  3. Did the error conditions in Appendices C.1 and C.4 occur during evaluation? If so, how were affected cases included in the reported metrics?
2 comments

We thank the reviewer for suggesting this controlled comparison. 

Q1.Contribution of explicit temporal markers. We separately fine-tune Stage-1 variants with/without markers during both training and inference, fixing the base model, training data/settings, and downstream checkpoints/settings. On Minning Town (Table 1), removing markers worsens all six metrics: AAS rises from 186.31 to 892.74 ms, and final accuracy falls from 90.97% to 60.28%. The larger decline after propagation than initialization is consistent with propagation's sensitivity to boundary errors.

Q2.Downstream comparison with MOSS. We evaluated both transcription methods within the same downstream pipeline on Minning Town (Table 2). MOSS achieves lower CER, whereas AutoCAST achieves higher SSGC, lower AAS, and 1.84 percentage points higher final attribution accuracy. Thus, lower CER does not directly translate into better attribution on this series. 

Q3.Reference and candidate coverage. Across the seven Chinese-language test dramas (Table 3), 12 of 380 characters, counted separately within each drama, lack acoustic references. Reference coverage ranges from 93.22% to 100%, while the correct character is included in the final candidate list for 96.98%–99.33% of evaluated utterances across dramas (Table 3). Coverage is therefore high but incomplete, and we will explicitly report the limitations caused by missing references and candidate omissions. 

Q4.Exception handling (Appendices C.1 and C.4). C.1 recovery was triggered in 466/10,256 non-final windows 4.544.54%, handled through window extension. In C.4, 26/166,294 subtitle entries 0.0160.016% had no acoustic candidates and received “Other.” All affected GT utterances remained in the attribution evaluation, with “Other” scored as incorrect for roster speakers. We will clarify these recovery procedures in the appendix.

Table 1. Ablation of input temporal markers on Minning Town, with the base model, training data, training settings, and downstream pipeline held fixed.

MethodCER ↓SSGC ↑AAS ↓Initial Attribution Acc. ↑Acc. after Acoustic Propagation ↑Final Attribution Acc. AACAAC ↑
AutoCAST (Full)9.57%96.25%186.31 ms75.08%88.16%90.97%
w/o Input Temporal Markers14.09%93.34%892.74 ms73.64%41.89%60.28%

Table 2. Comparison with MOSS on Minning Town using the same downstream attribution pipeline.

Transcription MethodCER ↓SSGC ↑AAS ↓AAC ↑
AutoCAST9.57%96.25%186.31 ms90.97%
MOSS8.79%95.16%194.25 ms89.13%

Table 3. Acoustic-reference coverage, final candidate recall, and Top-k accuracy across seven Chinese-language test dramas.

DramaAcoustic-Reference CoverageFinal Candidate RecallTop-1Top-2Top-3
The Long Night100.00% 37/3737/3796.98% 6,713/6,9226,713/6,92286.03%94.70%96.17%
Qin Empire 295.00% 57/6057/6097.93% 23,131/23,62123,131/23,62192.27%96.71%97.52%
Ode to Joy 193.22% 55/5955/5998.93% 31,544/31,88531,544/31,88593.23%97.30%98.20%
The Knockout97.33% 73/7573/7598.91% 31,228/31,57231,228/31,57290.23%96.82%98.13%
Minning Town96.47% 82/8582/8599.33% 20,917/21,05820,917/21,05893.08%96.63%97.88%
Stand by Me 1100.00% 30/3030/3099.14% 30,005/30,26430,005/30,26495.70%97.95%98.53%
Battle of Changsha100.00% 34/3434/3498.51% 16,691/16,94316,691/16,94392.63%96.41%97.55%

Table 4. Exception statistics across seven Chinese-language test dramas.

MetricThe Long NightQin Empire 2Ode to Joy 1The KnockoutMinning TownStand by Me 1Battle of ChangshaOverall
C.1 Trigger Rate67/613 10.9310.93%90/2,294 3.923.92%92/1,795 5.135.13%18/1,602 1.121.12%11/1,043 1.051.05%95/1,501 6.336.33%93/1,408 6.616.61%466/10,256 4.544.54%
C.4 Empty Acoustic Candidates5/7,75716/27,9320/33,2862/29,8840/20,5690/27,3753/19,49126/166,294 0.0160.016%

Thank you again for your thoughtful comments. We try to response your concerns in our rebuttal carefully. We would be happy to provide further clarification or additional evidence if any of our responses remain unclear. Thank you for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.