Pith.

Computer Vision and Pattern Recognition · arXiv:2610.00825 v1 · submitted 30 Sep 2026 · 51 references

Align Then Reason: A Multimodal Lip-Sync Judge for Dubbing

Rui Liu, Bhavin Jawade, Haoqi Li, Shivam Mehta, Karan Saxena, Yinghong Lan, Cameron R. Wolfe

The paper says

“Dubbing quality control requires a reference-free judge that can determine whether a candidate text line matches a speaker's visible articulation in both content and timing, using only silent video and text because dubbed audio may not yet exist.”

more of the abstract

Existing visual speech recognizers and video-language models are poorly suited to this setting: even when fine-tuned to recover spoken content from lip motion, they remain largely insensitive to temporal errors. We introduce Align Then Reason (ATR), a multilingual lip-sync judge that first establishes a monotonic alignment between frame-level lip representations and the phonetic units of the candidate line, then reasons over this alignment to make the final judgment. An alignment scorer provides the LLM with both local evidence for each phonetic unit and a calibrated global alignment score, enabling it to reason jointly about content and timing. On a seven-language benchmark, our method improves mean AUC over the corresponding Qwen3.5 SFT baselines by 59.4%, 50.2%, and 50.8% with 2B, 4B, and 9B reasoners, respectively. The gains generalize across LLM families, reaching mean AUC improvements of 45.9% and 46.6% over the best baseline for LLaMA-3.1-8B and Mistral-7B, respectively. They also transfer across datasets to three unseen MuAViC languages. Furthermore, we evaluate on two downstream tasks built from real dubbing lines. On dub-line reranking, ATR-9B outperforms the best lip-reading baseline by 52.0%, while on script-to-clip assignment, ATR-9B improves over the best lip-reading baseline by 17.7%.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 2 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A genuinely novel candidate-conditioned CTC judge, honestly ablated, but the temporal-sensitivity claim is only tested on corruptions that duplicate the training negatives, never on realistic order-preserving onset lag. This line is wrong

3 major objections · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 51 references say.

The paper says, at §3.1–3.2, Eq. (2)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The four temporal test axes (reverse, shift, freeze, swap) in Table 1 are exactly the R-family negatives in the contrastive loss L_align of Eq. (2). The temporal AUC therefore measures in-distribution discrimination against the scorer's own training targets, not generalization to timing-error families the model never saw in training.

more of objection 1

The ablation (Table 4) confirms that all temporal sensitivity lives in the calibrated scalar: removing it drops reverse from 0.964 to 0.525, shift from 0.931 to 0.534, and swap from 0.952 to 0.517, i.e., in the CTC score trained on those very negatives. Moreover, these axes partly change content as well as order (freeze duplicates a frame; circular shift wraps content across the boundary), consistent with freeze staying at 0.805 without the scalar. As the abstract and §1 present these numbers as broad 'temporal grounding' and 'sensitivity to temporal orders', the headlined temporal claim is unsupported beyond the trained corruption families.

The paper says, at §1 and §2.2, Eq. (1)

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The paper's own motivation (§1: a poor dub may 'begin too early, lag behind the mouth') is an order-preserving timing error, yet no evaluation axis or downstream task measures it. In Eq. (1), the CTC scorer has a blank state, so a constant onset lag or an early end is absorbed by leading or trailing blank emissions, and because the score is length-normalized by the number of units N rather than by frames, the log-likelihood changes little.

more of objection 2

Since the reasoner is trained only with content negatives (§2.3) and the ablation shows timing is carried solely by the calibrated scalar, the judge would plausibly score a delayed or shortened line as genuine. The downstream tasks in Tables 5–6 test content discrimination, not timing. Before the temporal-sensitivity claim — the paper's key differentiator over content-only baselines — can be considered verified, I ask for order-preserving timing experiments: onset-lag, early-end, and duration-mismatch negatives added to Eq. (2) with corresponding test axes, or an evaluation on real mis-timed dubs.

The paper says, at §3.4

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

A consequence of the above that should be stated explicitly: the LLM reasoner never sees a temporal negative during training and its temporal behavior is therefore entirely inherited from the scalar signal produced by a scorer trained on the four order-breaking families.

more of objection 3

The ablation (Table 4, 'w/o LLM reasoner') shows the reasoner neither adds nor preserves temporal sensitivity on its own. This means that any fix to the temporal evaluation (new order-preserving negatives, real mis-timed dubs) must also update the reasoner-training protocol if the model is to leverage the new timing evidence, and the paper currently provides no mechanism for the reasoner to use timing evidence beyond the scalar.

4 minor objections

The paper says, at Table 5 caption

Pith says: minor objection 1 of 4

The caption attributes the 0.479 Top-1 result to 'the alignment scorer', but the row and the main text attribute it to ATR-9B, which is the full model including the reasoner; the naming should be aligned.

The paper says, at Tables 1, 3 vs Table 9

Pith says: minor objection 2 of 4

Tables 1 and 3 report single numbers (0.920 for the seven-language mean; 0.751 and 0.875 for MuAViC zero-shot and recalibrated), while Table 9 reports three-seed means for the same quantities (0.910±0.015; 0.778±0.060; 0.852±0.033). The paper should state explicitly whether the main tables report a single primary run or the seed average, since 0.875 versus 0.852±0.033 is a nontrivial discrepancy for the headline cross-lingual claim.

show 2 more

The paper says, at §3.3

Pith says: minor objection 3 of 4

The MuAViC recalibration procedure requires 300 labeled genuine target-language samples; the paper describes this but does not discuss the deployment assumption that labeled genuine pairs are available in a target language at application time. A sentence on what the realistic uncalibrated performance (0.751) implies for a deployment without such labels would help.

The paper says, at §1

Pith says: minor objection 4 of 4

Typo: 'validate a effective judge' should read 'validate an effective judge'.

What would settle it

Construct a 'delayed onset' corruption: for each test sample in the seven-language benchmark, prepend K frames of a neutral/silent face (or zero-pad the video onset) before the original mouth motion, with K corresponding to 100–500 ms of delay, while keeping the candidate text unchanged. Measure pooled AUC of ATR-9B on this new axis. If AUC falls below ~0.8, the temporal sensitivity claim does not cover the most common dubbing timing error and the headline 0.920 mean AUC is not a reliable predictor of production performance. Also report AUC for an 'early offset' variant (truncate the tail). This directly tests whether the CTC blank label masks order-preserving timing shifts.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 2 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 27 s · 19,132 tokens read, 4,026 written
  3. Editor wrote the verdict above 18 s · 20,071 tokens read, 2,667 written
  4. Referee major revision: 3 major objections, 4 minor 1 min 15 s · 19,489 tokens read, 12,166 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Worth a real read, with one load-bearing caveat. The architecture is new: a candidate-conditioned CTC whose labels are positions in the candidate line rather than a fixed phoneme vocabulary, plus per-unit soft tokens and a calibrated scalar feeding an LLM reasoner. That is a genuine departure from standard CTC and from SyncNet-style audio-based sync, and it closes a real gap — nothing else scores arbitrary candidate text against silent video on both content and timing.

The evaluation is careful in ways that matter. The 59.4/50.2/50.8% relative gains match the Table 1 means exactly, the ablations behave as the design predicts (removing the calibrated scalar craters temporal axes to 0.672 while content barely moves), multi-seed variance is reported in Appendix D, and the MuAViC transfer restricted to refitting two scalars is a clean experiment. The ablation honesty deserves credit: the paper does not pretend the LLM does temporal work; it shows the global scalar is what carries it.

Now the soft spots, in proportion. The stress-test concern lands and is the big one. Temporal sensitivity is measured on reverse, shift, freeze, swap — exactly the R-family negatives of Eq. 2 — so the test is in-distribution against the model's own training signal. More specifically, the CTC in Eq. 1 includes a blank state, and the length normalization is by number of units, not frames, so a constant onset lag can be absorbed by leading blanks without much score change. The paper never evaluates an order-preserving timing error — a line that starts late or ends early with otherwise matching content — which is arguably the most common real mis-dub. Both downstream tasks (reranking, script-to-clip) test content discrimination, not timing. So the timing claim is real for order-breaking corruptions but narrower than the abstract implies, and it is the differentiator over content-only baselines. Second: the dataset is proprietary, no code or checkpoints, so the headline numbers are not independently reproducible; for a method paper that puts a real ceiling on verification. Third, minor: the reference list leans heavily on a self-citation cluster in the 2026 RL/reasoning literature; padding, not fatal.

My take: the content-discrimination results and the candidate-conditioned CTC design hold up, and the paper deserves referee time. The revision that would settle it is one experiment — an onset-lag or offset-early corruption that preserves order. Without it, the temporal generalization claim stays unverified.

Recommendation: yes, send to peer review. Expect the reviewers to ask for exactly that corruption test, and the paper will be stronger for it.

The strongest objection

The paper's central claim that ATR provides temporal sensitivity rests on pooled AUC measured on four synthetic temporal corruptions (reverse, circular shift, freeze, swap) in Section 3.2. These are exactly the R-family negatives used in the contrastive loss (Eq. 2), so the test measures in-distribution discrimination against the model's own training targets. All four corruptions break the monotonic path that CTC scores: reverse and swap violate order, shift wraps the sequence, and freeze repeats a frame. But the most common real dubbing timing error—a line that starts late or ends early relative to the mouth—preserves ordering. The CTC formulation in Eq. 1 includes a blank label, so a constant onset lag can be absorbed by leading blank emissions without reducing the alignment log-likelihood; the model may therefore score a delayed line as genuine. The paper never evaluates such an order-preserving temporal error. The downstream tasks (dub-line reranking, Section 3.5; script-to-clip assignment, Table 6) measure content discrimination, not timing. Consequently, the temporal-sensitivity claim is unverified for realistic mis-timing, which is the paper's key differentiator over content-only baselines. This is a load-bearing gap because ATR's advantage over Auto-A VSR on temporal axes is precisely what it claims beyond existing content discriminators.

The reader's weakest_assumption correctly identifies that the temporal-sensitivity claim is built on synthetic corruptions that duplicate the training negatives and that the downstream tasks do not exercise timing. My stress-test sharpens this into a concrete technical failure mode: the CTC alignment scorer includes a blank label, so an order-preserving delay (the most typical real dubbing error) can be absorbed by leading blanks and may score as genuine. None of the four evaluated temporal corruptions tests this class of error, and the paper provides no evidence that the temporal AUC generalizes to onset/offset mis-timing. This is load-bearing because ATR's claimed advantage over content-only baselines is precisely temporal judgment. The proposed delay-corruption test would settle whether the concern lands; if AUC remains high, the temporal claim is supported for realistic timing errors, and the reader's conditional accept stands. I therefore recommend no change to the reader's CONDITIONAL verdict, but with the explicit condition that the onset-delay test be run before the temporal-sensitivity claim is generalized to production dubbing.

The referee's report

If the temporal claim holds, this is a clean and useful contribution: decoupling monotonic alignment from LLM reasoning, with internally consistent ablations (Table 4), multi-seed statistics (Appendix D), public arithmetic that I verified (59.4/50.2/50.8% relative gains and the 52.0/17.7% downstream gains all check out), and a light-touch cross-lingual recalibration. However, the headline differentiator — temporal sensitivity — is measured only on the four corruption families that are simultaneously the training negatives of Eq. 2 and the test axes of Table 1, and none of them is an order-preserving timing error. The paper's own introduction motivates exactly such errors (a dub that 'begin[s] too early, lag[s] behind the mouth'). Because the CTC formulation of Eq. 1 includes blanks that absorb a constant onset lag or early end, and because the temporal signal in the full model is carried entirely by the calibrated scalar, the method would plausibly score a delayed line as genuine.

Pith's reading of the paper (a machine's summary, not the authors')

The paper proposes Align Then Reason (ATR), a reference-free lip-sync judge that decides whether a candidate text line matches a speaker's visible mouth motion — in both content and timing — from silent video and text alone, since dubbed audio may not exist at review time. It claims that visual speech recognizers and video-language models, even fine-tuned, stay near chance on temporal errors such as reversed, shifted, or frozen frames, and that the fix is to split the problem: first score a monotonic alignment between frame-level lip embeddings and the phonetic units of the candidate line, then let an LLM reason over that alignment. On a seven-language benchmark ATR-9B reaches mean pooled AUC 0.920 versus 0.610 for the strongest baseline, a 50.8% relative improvement, and the gains persist across LLM families (LLaMA-3.1-8B, Mistral-7B), transfer to three unseen languages after a two-scalar recalibration (0.875 mean AUC), and appear on real-line downstream tasks (52.0% relative gain on dub-line reranking, 17.7% on script-to-clip assignment). The load-bearing assertion is that explicit monotonic alignment supplies temporal grounding while LLM reasoning improves content discrimination.

What carries the argument. The central object is a candidate-dependent CTC scorer: frame-level lip embeddings and phonetic-unit embeddings of the candidate line form a cosine-similarity matrix, and because the labels are the positions of the units within the line, the Connectionist Temporal Classification (CTC) log-likelihood scores only monotonic paths through the matrix — reordering, shifting, or freezing the frames breaks the path. Two encoders (Auto-A VSR lip embeddings and XPhoneBERT phonetic units, projected into a shared 256-dimensional space) are trained contrastively against content negatives (mismatch, shuffle, dub) and temporal negatives (reverse, shift, freeze, swap), with an auxiliary phoneme-recognition CTC head shaping the visual features. The scorer then emits, for each phonetic unit, a soft token — a projection of the unit embedding, its frame-attended visual summary, and its strongest local match — plus a calibrated global score, and these are fed to an LLM reasoner fine-tuned with LoRA whose Yes/No logit difference is the judge's score.

Core claim. The central claim is that lip-sync judgment for dubbing reduces to a two-stage procedure: a candidate-conditioned CTC alignment scorer establishes which phonetic units of the candidate line match which video frames in the correct monotonic order, then feeds per-unit local evidence (soft tokens) and a calibrated global alignment score to an LLM reasoner whose Yes/No logit difference is the final score. The paper reports this structure gives balanced high accuracy on both content corruptions (mismatch, shuffle, dub) and temporal corruptions (reverse, shift, freeze, swap), reaching mean pooled AUC 0.920 with a 9B reasoner on a seven-language benchmark, and that temporal sensitivity is what separates ATR from every baseline: video-language models hold 0.510–0.520 AUC on temporal axes while ATR reaches 0.931–0.970. The authors attribute this to the monotonic scoring structure encoding order directly, with the LLM contributing content discrimination — the ablation without the reasoner drops mean AUC to 0.856 mainly on content axes, and without the calibrated scalar to 0.672 mainly on temporal axes, confirming the split.

The premise it rests on. The temporal-sensitivity claim rests on the four synthetic video corruptions (reverse, circular shift, freeze, swap) being representative of real dubbing timing errors — the paper trains on these same corruptions and measures on them, so if production mis-dubs are subtler, such as a phrase starting late or a phonetic near-miss rather than a reversed or frozen clip, the measured AUC may not carry over.

If the paper is right (Pith's inferences, not the authors')

A reference-free judge can score candidate lines from silent video and text before dubbed audio is synthesized, fitting review-time dubbing pipelines and test-time-scaling verifier loops that sample several candidate lines.

Because text is represented phonetically rather than in a language-specific vocabulary, the same alignment mechanism serves all seven benchmark languages and three unseen ones without retraining.

Temporal errors that defeat fine-tuned video-language models — reversed, shifted, or frozen frames — are caught because the monotonic scoring structure encodes order directly, not because the LLM infers timing from pixels.

On real dubbing lines the judge ranks the professional line above meaning-preserving paraphrases (52% relative Top-1 gain) and recovers script-to-clip correspondence (17.7% relative gain), supporting use as a dub-line verifier.

Because the paper's central claim splits temporal grounding (scorer) from semantic judgment (LLM), a natural testable extension would swap the LLM for a cheaper content classifier and check how much of the 0.920 AUC survives; if nearly all of it does, the LLM's role is exactly content discrimination as claimed, and the judge could be deployed at lower cost.

The same scorer–reasoner interface could apply to other silent-video verification problems — for example, checking whether a translated subtitle or caption agrees with visible articulation in user-generated video — since nothing in the design depends on dubbing-specific data.

The two-scalar recalibration needed to reach 0.875 AUC on the unseen MuAViC languages indicates the raw alignment score is domain-sensitive, so a production judge would likely require continuous recalibration or a distribution-robust score; the paper acknowledges this domain-shift issue but does not provide a mechanism beyond refitting two constants.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Dubbing quality control requires a reference-free judge that can determine whether a candidate text line matches a speaker's visible articulation in both content and timing, using only silent video and text because dubbed audio may not yet exist. Existing visual speech recognizers and video-language models are poorly suited to this setting: even when fine-tuned to recover spoken content from lip motion, they remain largely insensitive to temporal errors. We introduce Align Then Reason (ATR), a multilingual lip-sync judge that first establishes a monotonic alignment between frame-level lip representations and the phonetic units of the candidate line, then reasons over this alignment to make the final judgment. An alignment scorer provides the LLM with both local evidence for each phonetic unit and a calibrated global alignment score, enabling it to reason jointly about content and timing. On a seven-language benchmark, our method improves mean AUC over the corresponding Qwen3.5 SFT baselines by 59.4%, 50.2%, and 50.8% with 2B, 4B, and 9B reasoners, respectively. The gains generalize across LLM families, reaching mean AUC improvements of 45.9% and 46.6% over the best baseline for LLaMA-3.1-8B and Mistral-7B, respectively. They also transfer across datasets to three unseen MuAViC languages. Furthermore, we evaluate on two downstream tasks built from real dubbing lines. On dub-line reranking, ATR-9B outperforms the best lip-reading baseline by 52.0%, while on script-to-clip assignment, ATR-9B improves over the best lip-reading baseline by 17.7%.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-02

The authors' main claim, as the classifier picked it out 0.77

“On a seven-language benchmark, our method improves mean AUC over the corresponding Qwen3.5 SFT baselines by 59.4%, 50.2%, and 50.8% with 2B, 4B, and 9B reasoners, respectively.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.09 · human participants 0.06 · controlled comparison 0.97 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Dubbing quality control requires a reference-free judge that can determine whether a candidate text line matches a speaker's visible articulation in both content and timing, using only silent video and text because dubbed audio may not yet exist. background 0.99 0.08 0.05 asserted 0.98
Existing visual speech recognizers and video-language models are poorly suited to this setting: even when fine-tuned to recover spoken content from lip motion, they remain largely insensitive to temporal errors. gap 0.81 0.07 0.02 asserted 1.00
We introduce $\textit{Align Then Reason}$ (ATR), a multilingual lip-sync judge that first establishes a monotonic alignment between frame-level lip representations and the phonetic units of the candidate line, then reasons over this alignment to make the final judgment. contribution 0.57 0.92 0.17 asserted 1.00
An alignment scorer provides the LLM with both local evidence for each phonetic unit and a calibrated global alignment score, enabling it to reason jointly about content and timing. method 0.99 0.80 0.08 asserted 1.00
On a seven-language benchmark, our method improves mean AUC over the corresponding Qwen3.5 SFT baselines by 59.4%, 50.2%, and 50.8% with 2B, 4B, and 9B reasoners, respectively. result 1.00 0.96 0.25 asserted 1.00
The gains generalize across LLM families, reaching mean AUC improvements of 45.9% and 46.6% over the best baseline for LLaMA-3.1-8B and Mistral-7B, respectively. result 0.99 0.95 0.13 asserted 1.00
They also transfer across datasets to three unseen MuAViC languages. result 0.99 0.92 0.06 asserted 1.00
Furthermore, we evaluate on two downstream tasks built from real dubbing lines. contribution 0.71 0.94 0.03 asserted 0.98
On dub-line reranking, ATR-9B outperforms the best lip-reading baseline by 52.0%, while on script-to-clip assignment, ATR-9B improves over the best lip-reading baseline by 17.7%. result 1.00 0.96 0.16 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-02: 1 request, 7,225 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (1)
Figure 1 of the paper
Figure 1. Overview of Align Then Reason (ATR). Stage 1 (Align): the silent mouth video is encoded into frame embeddings h1:T and the candidate line into phonetic-unit embeddings u1:N . Their similarity Mt,n is scored by CTC whose labels are the positions of the units in the line, so only a monotonic path through M (highlighted) scores well, and reordering, shifting, or freezing the frames breaks it. Trained with content and temporal negatives, the scorer emits a calibrated global score a(V, c) and per-unit soft tokens e soft 1:N . Stage 2 (Reason): the soft tokens are prepended to a prompt…

Figures are the authors', reproduced from the paper.

References (51)
  1. [1] Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks , author=. Proceedings of the 23rd international conference on Machine learning , pages= on Pith
  2. [2] Learning audio-visual speech representation by masked multimodal cluster prediction , author=. arXiv preprint arXiv:2201.02184 , year= on PithPith's reviewarXiv
  3. [3] Revisiting the “video” in video-language understanding , author=. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2022 , organization= on Pith
  4. [4] Revealing single frame bias for video-and-language learning , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages= on Pith
  5. [5] Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation , author=. arXiv preprint arXiv:2303.00628 , year= on PithPith's reviewarXiv
  6. [6] Auto-AVSR: Audio-Visual Speech Recognition with Automatic Labels , author =. 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year = on Pith
  7. [7] Ma, Pingchuan and Haliassos, Alexandros and Fernandez-Lopez, Adriana and Chen, Honglie and Petridis, Stavros and Pantic, Maja , booktitle=. Auto-AVSR: Audio-Visual Speech Recognition with Automatic Labels , year= on Pith
  8. [8] Yeo, Jeonghun and Han, Seunghee and Kim, Minsu and Ro, Yong Man , booktitle =. Where Visual Speech Meets Language:. 2024 , month = nov, address = on Pith
  9. [9] From Speech-to-Speech Translation to Automatic Dubbing , author =. Proceedings of the 17th International Conference on Spoken Language Translation , year = on Pith
  10. [10] Tam, Derek and Lakew, Surafel M. and Virkar, Yogesh and Mathur, Prashant and Federico, Marcello , booktitle =. 2022 , pages =. doi:10.21437/Interspeech.2022-11136 , issn = on Pithdoi:10.21437/interspeech.2022-11136
  11. [11] Nguyen, Linh The and Pham, Thinh and Nguyen, Dat Quoc , booktitle =. 2023 , pages =. doi:10.21437/Interspeech.2023-444 , issn = on Pithdoi:10.21437/interspeech.2023-444
  12. [12] Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle =. Judging on Pith
  13. [13] The Power of Scale for Parameter-Efficient Prompt Tuning , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , year = on Pith
  14. [14] On Calibration of Modern Neural Networks , author =. Proceedings of the 34th International Conference on Machine Learning , year = on Pith
  15. [15] From speech-to-speech translation to automatic dubbing , author=. Proceedings of the 17th International Conference on Spoken Language Translation , pages= on Pith
  16. [16] Dubbing in practice: A large scale study of human localization with insights for automatic dubbing , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher= on Pith
  17. [17] Synchformer: Efficient synchronization from sparse cues , author=. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2024 , organization= on Pith
  18. [18] Audio--visual synchronization and lip movement analysis for real-time deepfake detection , author=. International Journal of Computational Intelligence Systems , volume=. 2025 , publisher= on Pith
  19. [19] Cooperative learning of audio and video models from self-supervised synchronization , author=. Advances in Neural Information Processing Systems , volume= on Pith
  20. [20] Large language models are strong audio-visual speech recognition learners , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization= on Pith
show 31 more
  1. [21] Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and others , journal=. Qwen3- on Pith
  2. [22] Liu, Yuanxin and Li, Shicheng and Liu, Yi and Wang, Yuxiang and Ren, Shuhuai and Li, Lei and Chen, Sishuo and Sun, Xu and Hou, Lu , booktitle= on Pith
  3. [23] Temporal Reasoning Transfer from Text to Video , author=. International Conference on Learning Representations , year= on Pith
  4. [24] Improve Temporal Reasoning in Multimodal Large Language Models via Video Contrastive Decoding , author=. Advances in Neural Information Processing Systems , volume= on Pith
  5. [25] Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle= on Pith
  6. [26] Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year= on Pith
  7. [27] Deep Visual Forced Alignment: Learning to Align Transcription with Talking Face Video , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year= on Pith
  8. [28] Multimodal Few-Shot Learning with Frozen Language Models , author=. Advances in Neural Information Processing Systems , volume= on Pith
  9. [29] A Lip Sync Expert Is All You Need for Speech to Lip Generation In the Wild , author=. Proceedings of the 28th ACM International Conference on Multimedia , pages= on Pith
  10. [30] Mistral 7B , author=. arXiv preprint arXiv:2310.06825 , year= on PithPith's reviewarXiv
  11. [31] The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year= on PithPith's reviewarXiv
  12. [32] 6G non-terrestrial networks enabled low-altitude economy: Opportunities and challenges , author=. arXiv preprint arXiv:2311.09047 , year= on PithPith's reviewarXiv
  13. [33] Audiovisual translation: dubbing , author=. 2020 , publisher= on Pith
  14. [34] Claude Opus 5 , year = on Pith
  15. [35] GPT-5.6: Frontier intelligence that scales with your ambition , howpublished = on Pith
  16. [36] Out of time: automated lip sync in the wild , author=. Asian conference on computer vision , pages=. 2016 , organization= on Pith
  17. [37] DeltaRubric: Generative Multimodal Reward Modeling via Joint Planning and Verification , author=. arXiv preprint arXiv:2605.09269 , year= on PithPith's reviewarXiv
  18. [38] Self-rewarding vision-language model via reasoning decomposition , author=. arXiv preprint arXiv:2508.19652 , year= on PithPith's reviewarXiv
  19. [39] Reinforcing Multimodal Reasoning Against Visual Degradation , author=. arXiv preprint arXiv:2605.09262 , year= on PithPith's reviewarXiv
  20. [40] Dual-Uncertainty Guided Policy Learning for Multimodal Reasoning , author=. arXiv preprint arXiv:2510.01444 , year= on PitharXiv
  21. [41] Adaptive Conformal Guidance for Learning under Uncertainty , author=. International Conference on Learning Representations , volume= on Pith
  22. [42] Cde: Curiosity-driven exploration for efficient reinforcement learning in large language models , author=. International Conference on Learning Representations , volume= on Pith
  23. [43] Parallel-r1: Towards parallel thinking via reinforcement learning , author=. International Conference on Learning Representations , volume= on Pith
  24. [44] Save the good prefix: Precise error penalization via process-supervised rl to enhance llm reasoning , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages= on Pith
  25. [45] Stable and efficient single-rollout rl for multimodal reasoning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages= on Pith
  26. [46] Parallel-probe: Towards efficient parallel thinking via 2d probing , author=. arXiv preprint arXiv:2602.03845 , year= on PitharXiv
  27. [47] Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling , author=. arXiv preprint arXiv:2606.03102 , year= on PithPith's reviewarXiv
  28. [48] LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling , author=. arXiv preprint arXiv:2605.08083 , year= on PithPith's reviewarXiv
  29. [49] Mmcd: Multi-modal collaborative decision-making for connected autonomy with knowledge distillation , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization= on Pith
  30. [50] Caml: Collaborative auxiliary modality learning for multi-agent systems , author=. Advances in Neural Information Processing Systems , volume= on Pith
  31. [51] Dream-RSI: Recursive Self-Improvement through Evolving Worlds , author=. arXiv preprint arXiv:2609.14858 , year= on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment