Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SimulPL: Aligning Human Preferences in Simultaneous Machine Translation

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A preference-learning framework called SimulPL aligns simultaneous machine translation with five human preferences, including latency, and reports higher quality at every latency level.

desk verdict A solid empirical SiMT alignment paper whose central proof has a sign error, so the latency term should be treated as a heuristic rather than a principled objective. read the letter →

arxiv 2502.00634 v2 pith:HTCXYM2E submitted 2025-02-02 cs.CL cs.AI

Donglei Yu , Yang Zhao , Jie Zhu , Yangyifan Xu , Yu Zhou , Chengqing Zong This is my paper · ORCID
classification cs.CLcs.AI
keywords simultaneousmachinetranslationpreferenceoptimizationdirectread/writepolicylatencyhumanalignmentLLM-generateddatamonotonicity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SimulPL is a preference-learning framework for simultaneous machine translation (SiMT), the setting where a translator must produce output while the source is still streaming in. The paper argues that SiMT audiences care about five things: translation quality, monotonic word order, key-point coverage, syntactic simplicity, and low latency, and that existing preference-alignment methods such as DPO ignore the last two because they only score finished responses and never adjust the read/write policy. SimulPL addresses this by using the first four preferences as prompts for a general-purpose LLM to generate human-preferred references, then fine-tuning with a dual objective that learns translation and a confidence-based read/write signal, and finally applying a customized DPO whose objective includes the expected output length as a proxy for latency. Across Chinese→English, German→English, and English→Chinese, the authors report that SimulPL beats supervised-fine-tuning and standard-DPO baselines on BLEU and COMET at all latency levels, with interpreter-based human evaluation showing more preferred outputs in every latency group.

What carries the argument

The load-bearing object is the SimulDPO objective, Equation 9, where the usual DPO log-ratio is summed over token positions and weighted by the model's own read/write confidence $c_t$, with an additional $-\alpha$ per written token. Because the paper writes the output length as $|y|=\sum_t c_t$, the latency term $\alpha\,\mathbb{E}[|y|]$ becomes differentiable with respect to the confidence layer, which is how latency preference is injected into preference optimization and how the read/write policy is updated during alignment. A second piece is the confidence-based inference policy (write when $c_t>0.5$, otherwise read $n$ more source words), which converts the trained confidence into latency control. The third piece is the human-preference prompt set for data construction, encoding the first four preferences (quality, monotonicity, key points, simplicity) so that LLM-generated chosen responses differ from original references in the intended direction.

What would settle it

Recompute the bound in Appendix B.1: from $AL \le -C_1|y| + C_2$ it follows that $-AL \ge C_1|y| - C_2$, so the surrogate objective in Equation 15 is a lower bound of the latency-aware objective rather than an upper bound; demonstrating this sign error falsifies the theoretical grounding for treating $\alpha\mathbb{E}[|y|]$ as a latency proxy. A behavioral check would be an A/B test of SimulDPO with $+\alpha|y|$ against a variant that penalizes over-generation, and seeing whether latency metrics (LAAL/AL) improve independently of output length changes.

Watch

Extended reading notes

Core claim

SimulPL's central claim is that human preference alignment in simultaneous translation must handle not only what is translated but when it is translated, and that both can be optimized in the same preference-learning pass. The paper categorizes SiMT human preferences into five aspects and builds a framework around them: human-preference prompts derived from the first four aspects guide a large language model (GPT-4/4o) to produce preferred translation references; Multi-task Supervised Fine-tuning (MSFT) jointly trains the translation decoder and an additional confidence layer that acts as the read/write policy; and Simultaneous Direct Preference Optimization (SimulDPO) modifies the DPO objective with a latency term $\alpha\,\mathbb{E}[|y|]$, made differentiable by identifying the output length with the sum of the confidence read/write decisions, $\sum_t c_t$. The paper reports that the complete pipeline yields higher SacreBLEU and COMET scores than SFT and standard DPO at every tested latency level on three language pairs, and that professional interpreters prefer its outputs more often in all latency groups.

Load-bearing premise

The framework's latency preference rests on the claim—made in Appendix B.1—that maximizing expected output length is equivalent to minimizing expected Average Lagging; if that inequality runs the wrong way, the length term is just a length bias and the latency story collapses.

Editorial extensions

If this is right

  • SimulPL beats SFT and standard DPO on SacreBLEU and COMET at every tested latency level in Zh→En, De→En, and En→Zh.
  • Interpreter-based human evaluation shows SimulPL outputs are preferred more often in low-, mid-, and high-latency groups.
  • Multi-aspect metrics indicate SimulPL outputs are more monotonic, shorter, and syntactically simpler at matched latency.
  • The latency-aware objective transfers to other preference optimizers: SimulCPO and SimulKTO outperform their base CPO and KTO versions.
  • MSFT's joint training of translation and read/write policy improves over plain SFT initialization for the preference phase.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The confidence-differentiability trick—treating $\sum_t c_t$ as output length—could transfer to other streaming generation tasks, such as real-time speech translation or live captioning, where latency and content quality are jointly optimized.
  • If the latency proxy holds, preference data for SiMT need not be human-annotated; LLM-generated references aligned to prompt-encoded preferences could scale to more language pairs at low cost.
  • Disentangling the latency contribution of $\alpha\mathbb{E}[|y|]$ from its length contribution is a direct next step; controlling output length while varying latency pressure would isolate the mechanism.
  • The framework's five-way preference categorization might generalize to other real-time language tasks, such as simultaneous speech-to-speech interpretation, where monotonicity and key-point fidelity matter as much as speed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. SimulPL proposes a preference-learning framework for simultaneous machine translation (SiMT). It defines five human preferences (translation quality, monotonicity, key points, simplicity, latency), uses the first four to prompt GPT-4/4o to generate human-preferred references (with original references as dispreferred), extracts prefix-level preference pairs, then fine-tunes an LLM-based SiMT policy with a multi-task SFT (MSFT) that adds a confidence head for read/write decisions, followed by SimulDPO, a DPO-style loss with a +alpha|y| latency term and confidence-weighted token log-ratios. Experiments on Zh->En, De->En, En->Zh compare SimulPL against LLM-PFX-SFT and LLM-PFX-SFT+DPO (plus wait-k/SM2 in latency-quality curves), reporting BLEU/COMET gains at matched LAAL, human evaluation wins in three latency bands, and multi-aspect metric improvements. Ablations (MSFT, SimulDPO, pi_ref with full source vs prefix, alpha sensitivity, generalization to SimulCPO/SimulKTO) support the framework components.

Significance. If the empirical claims hold, the paper is a useful step: it is the first systematic attempt to align SiMT outputs with a stated multi-aspect human-preference model, and it ships a full evaluation stack (annotated references, human evaluation by interpreters, latency-banded multi-aspect metrics) plus code/data promises. The framework is modular (SimulDPO generalizes to CPO/KTO), and the reported gains at low latency are nontrivial and plausible in direction. However, the central theoretical justification for the +alpha|y| latency term contains a sign error (Appendix B.1, Eqs. 14-15), and the preference-data construction is partly circular with the evaluation criteria (the same five preferences drive both the data prompts and the metric suite). Neither issue necessarily kills the empirical contribution, but the paper's strongest framing claim—that the length term is derived from the latency objective—is not supportable as written.

major comments (4)
  1. [Appendix B.1, Eqs. (14)-(15)] There is a sign/direction error in the bounding argument. From AL <= -C1|y| + C2 it follows that -AL >= C1|y| - C2, so the objective with +C1 E[|y|] is a lower bound of the latency-aware objective (Eq. 12), not an upper bound as Eq. (15) claims. The sentence 'Based on this upper bound, we can optimize the objective in Equation 12 by optimizing the following one' is therefore mathematically incorrect. Maximizing a lower bound is not a principled way to maximize the original objective; it may still be a reasonable heuristic, and the paper can be revised to present +alpha|y| as an empirical proxy, but as written the proof does not establish the claimed equivalence between latency reduction and output-length maximization.
  2. [Section 4.4, Eq. (9) and the c_t mechanism] The treatment of c_t in the SimulDPO loss needs a clearer justification. The identity |y| = sum_{t=1}^{|y|+1} I(t <= |y|) = sum c_t holds for teacher-forced reference tokens, but in Eq. (9) c_t for the chosen response is a model prediction that also appears inside the log-likelihood ratio via the exponent c_t. The loss then couples the confidence signal and the token probability in a way that is neither a standard DPO objective nor a well-defined maximum-likelihood objective; in particular, the statement that the model 'learns to predict c_t close to 1' depends on the gradient of c_t through log pi_theta, but the surrounding derivation does not analyze this coupling or show that the loss has the claimed fixed point. Please provide a formal derivation of the gradient of Eq. (9) with respect to the confidence-layer parameters, or state clearly that the confidence updates are heuristic.
  3. [Section 4.2 and Table 2 / Figure 5] The preference data and the evaluation metrics operationalize the same five criteria: GPT-4 is prompted to satisfy monotonicity, key-point, and simplicity constraints, and then the same proxy metrics (NIR, SLR, DD) are used to show that SimulPL aligns with those preferences. This is a partial circularity: the chosen/rejected preference pairs encode a particular interpretation of the criteria, and gains on the in-house proxy metrics may reflect the data-generation prompt rather than a robust alignment with human preferences. The human evaluation (Figure 5) mitigates this concern for the headline comparison, but the multi-aspect metrics in Table 2 should be presented as training-data-consistent proxies, and the claim that they 'measure' the preferences should be softened.
  4. [Tables 9-11 and Figure 3] The methodological comparison is incomplete. SimulPL is an LLM-based method and is primarily compared against LLM-PFX-SFT and LLM-PFX-SFT+DPO; Transformer-based wait-k and SM2 are shown in Figures 3-4 but are not discussed in the numerical tables for quality metrics, and the paper does not report results for a Transformer baseline equipped with the same MSFT+SimulDPO treatment. Without such a control, the observed gains could be attributed to the specific LLM base plus prefix-pair SFT configuration rather than to the proposed SimulPL components. Adding a Transformer+SimulPL row, or at least an explicit analysis of this confound, would strengthen the central claim that the framework, not the LLM base, drives the gains.
minor comments (5)
  1. [Section 5.1, 'NIR' definition] The definition of Normalized Inversion Rate uses the inversion number of aligned source positions, but the source positions are first produced by awesome-align, which is a soft aligner; the paper does not state how ties and null alignments are resolved when constructing the sequence A. Please specify the tie-breaking and unaligned-token handling.
  2. [Section 4.4, Eq. (9)] The subscript/superscript notation for r_w^t and r_l^t is inconsistent: Eq. (9) writes r_t^w and r_t^l while the text just below uses r_t^* with * in {w,l}. Also the sentence 'if the SiMT model can accurately predict a token ... then the SiMT model will learn to predict c_t^w close to 1' should specify whether this is a heuristic justification or a proven property of the gradient.
  3. [Table 10, De->En row for SimulPL] The column order in the De->En SimulPL rows differs from the other rows ('BLEU LAAL AL AP DAL COMET' instead of 'BLEU COMET LAAL AL AP DAL'), which is easy to misread. Please align the column ordering.
  4. [Section 5.5, Fig. 7] The alpha ablation only reports SacreBLEU versus LAAL; a COMET or human-preference readout would help distinguish a length/latency tradeoff from a real preference-alignment improvement, especially since +alpha|y| is a length bias.
  5. [Appendix A.2, Table 5] The human evaluation of the annotated references reports 'win ratio', 'lose ratio', and 'tie ratio', but the current column order and the German-English row (lose 80.5%, win 2%) are easy to misread. Please clarify which side is 'win' in the header or caption.

Circularity Check

1 steps flagged · score 4.0 of 10

Multi-aspect evaluation reuses the same preference criteria that generated the training data, making those 'alignment' results partly self-referential; Appendix B.1's sign error is a non-circular correctness flaw.

  1. self definitional [Section 4.2, Appendix A.1/Figure 9, Section 5.1, Table 2.]
    "By leveraging the first four preferences, we construct human preference prompts... 'Only interpret the most important key points... Use simple, common words, and keep the syntax very easy. Make your interpretation as short as possible.' ... 'For Monotonicity Preference, we define Normalized Inversion Rate (NIR)... For Key Point Preference, we use the Sentence Length Ratio (SLR)... For Simplicity Preference, we use Stanza... Dependency Depth (DD)... A smaller DD indicates simpler syntax.'"

    These five preference categories are load-bearing in both directions. They are the content instructions used to generate the chosen references Yw via GPT-4/4o (prompt: keep word order, omit non-key points, use simple syntax, be short), and they are the operational definitions of the multi-aspect evaluation metrics in Section 5.1 (NIR for monotonicity, SLR for key-point concision, DD for simplicity). Training on Yw therefore makes SimulPL's NIR/SLR/DD wins over LLM-PFX-SFT and LLM-PFX-SFT+DPO substantially a measure of how faithfully the model imitates the GPT-4 references rather than an independent confirmation of human preference.

full rationale

The paper's headline claim is not fully circular: BLEU/COMET are computed against test references that were manually revised by qualified simultaneous interpreters (Section 4.2), and Figure 5 reports direct human preference judgments, both of which are external to the GPT-4 data-generation loop. The one clear circular component is the multi-aspect evaluation: the same criteria used to prompt GPT-4/4o into producing Yw are later re-used as the operational metrics (NIR, SLR, DD) for 'preference alignment.' This makes the automatic multi-aspect results partly self-referential, and it is the reason the score is above 2. I also flag a separate, non-circular correctness risk in Appendix B.1: Eq. 14 gives AL ≤ -C1|y| + C2, but Eq. 15 claims the objective with +C1E[|y|] is an upper bound; multiplying by -1 reverses the inequality, so the modified objective is a lower bound, not an upper bound. That invalidates the claimed proof that +αE[|y|] optimizes the latency-aware objective. Because this is a derivation error rather than a reduction of a result to its own inputs, I do not count it as circularity. No self-citation chain or imported uniqueness theorem is load-bearing. Overall: partial circularity (multi-aspect evaluation), plus a serious but non-circular proof flaw in the latency-bound derivation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim relies on a small number of hand-tuned hyperparameters and several domain assumptions. The most consequential are that GPT-4 generated data encodes human preferences, that output length is a valid latency proxy, and that the paper's chosen metrics capture the corresponding preferences. The confidence layer and read/write policy are model components, not invented physical entities.

free parameters (3)
  • α (SimulDPO latency/length weight) = 0.1
    Selected by sweeping 0, 0.1, 1, 10 on the Zh to En task in Figure 7. The claimed proof that this term represents latency is invalid, so α is effectively fit on the development set.
  • β (DPO KL weight) = 0.1
    A standard DPO hyperparameter chosen by hand from prior practice, not derived from SiMT specifics.
  • γ (confidence threshold) = 0.5
    Inference read/write threshold set to 0.5 and tested in Appendix D, Table 6; it is a manual choice affecting the latency-quality trade-off.
assumptions (6)
  • standard math Bradley-Terry preference model and the closed-form DPO reward derivation are valid.
    Used in Section 4.4, Equations 1, 7, and 8, without proof or alternative justification.
  • domain assumption Real-time SiMT audiences prefer accurate, monotonic, key-point-focused, simple, low-latency translations.
    Section 4.1 categorizes human preferences based on cited linguistic and interpreting studies; this is the target the method optimizes.
  • domain assumption GPT-4/4o prompted with the constructed human preference prompts produces translations aligned with human preferences.
    Section 4.2 uses GPT-4 outputs as preferred references, but Appendix A.2 Table 5 shows GPT-generated references score lower than manually revised references in human evaluation, partially contradicting the assumption.
  • domain assumption awesome-align word alignments are accurate enough for prefix-pair extraction.
    Equation 3 relies on word alignment to build source-prefix and target-prefix pairs; no alignment quality analysis is given.
  • ad hoc to paper Output length E[|y|] is a valid proxy for latency reduction.
    Appendix B.1 claims a proof, but the inequality in Equation 15 has the wrong direction, so the connection between the length term and latency is not established and α is tuned instead.
  • ad hoc to paper NIR, SLR, and DD operationalize monotonicity, key-point, and simplicity preferences.
    Section 5.1 defines NIR, SLR = |Y|/|X|, and dependency depth as preference metrics; shorter output and shallower syntax are choices, not established preference measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SimulPL: Aligning Human Preferences in Simultaneous Machine Translation." pith.science (2026). https://pith.science/paper/HTCXYM2E

@misc{pith2026250200634,
  author       = {Pith},
  title        = {Pith review of: SimulPL: Aligning Human Preferences in Simultaneous Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTCXYM2E}},
  note         = {Machine review of arXiv:2502.00634}
}
abstract

Simultaneous Machine Translation (SiMT) generates translations while receiving streaming source inputs. This requires the SiMT model to learn a read/write policy, deciding when to translate and when to wait for more source input. Numerous linguistic studies indicate that audiences in SiMT scenarios have distinct preferences, such as accurate translations, simpler syntax, and no unnecessary latency. Aligning SiMT models with these human preferences is crucial to improve their performances. However, this issue still remains unexplored. Additionally, preference optimization for SiMT task is also challenging. Existing methods focus solely on optimizing the generated responses, ignoring human preferences related to latency and the optimization of read/write policy during the preference optimization phase. To address these challenges, we propose Simultaneous Preference Learning (SimulPL), a preference learning framework tailored for the SiMT task. In the SimulPL framework, we categorize SiMT human preferences into five aspects: \textbf{translation quality preference}, \textbf{monotonicity preference}, \textbf{key point preference}, \textbf{simplicity preference}, and \textbf{latency preference}. By leveraging the first four preferences, we construct human preference prompts to efficiently guide GPT-4/4o in generating preference data for the SiMT task. In the preference optimization phase, SimulPL integrates \textbf{latency preference} into the optimization objective and enables SiMT models to improve the read/write policy, thereby aligning with human preferences more effectively. Experimental results indicate that SimulPL exhibits better alignment with human preferences across all latency levels in Zh$\rightarrow$En, De$\rightarrow$En and En$\rightarrow$Zh SiMT tasks. Our data and code will be available at https://github.com/EurekaForNLP/SimulPL.

Figures

Figures reproduced from arXiv: 2502.00634 by the authors.

Figure 1
Figure 1. Overview of our proposed SimulPL Framework. With the first four preferences, we [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Human evaluation between our annotated target references and origin target references. Our newly annotated references are more preferred. Additionally, several methods, such as DPO (Rafailov et al., 2024), directly conduct preference alignment without a reward model. However, existing preference alignment methods cannot be directly applied to the SiMT task, as their optimization objectives do not account for the lat… view at source ↗
Figure 3
Figure 3. SacreBLEU against LAAL on Zh→En, De→En and En→Zh SiMT tasks. 0 2 4 6 8 10 12 14 LAAL 60.0 62.5 65.0 67.5 70.0 72.5 75.0 77.5 80.0 COMET wait-k SM2 LLM-PFX-SFT LLM-PFX-SFT+DPO SimulPL (a) Zh→En 0 2 4 6 8 10 12 14 LAAL 76 78 80 82 84 86 88 COMET wait-k SM2 LLM-PFX-SFT LLM-PFX-SFT+DPO SimulPL (b) De→En 0 2 4 6 8 10 12 14 LAAL 68 70 72 74 76 78 80 82 84 COMET wait-k SM2 LLM-PFX-SFT LLM-PFX-SFT+DPO SimulPL (c) En→Zh [PI… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: COMET against LAAL on Zh→En, De→En and En→Zh SiMT tasks. complete OMT training sets and then fine-tune on our annotated data. For LLM-based SiMT mod￾els, we begin with SFT on a subset of the OMT training data with the size of 100k, followed by additional SFT on our ann…
Figure 5
Figure 5. Figure 5: Human Evaluation in differenct latency groups on Zh [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation Studies of SimulPL framework on Zh→En SiMT task. 0 2 4 6 8 10 12 14 LAAL 22 23 24 25 26 27 28 29 30 31 32 SacreBLEU α = 0 α = 0.1 α = 1 α = 10 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 9
Figure 9. Figure 9: System: You are a professional simultaneous interpreter, simulating the scenario of interpreting a speaker's Chinese speech into English in real time. User: As the speaker speaks, you need to: 1.Interpret the speaker's Chinese sentence into fluent English for the audie…
Figure 10
Figure 10. Figure 10: Our prompt template in the SimulPL framework and other LLM-based SiMT models. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 13
Figure 13. Figure 13: The compari￾son between different prefer￾ence optimization methods on De→En SiMT task. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice

    cs.CL 2025-07 conditional novelty 6.0 of 10

    An end-to-end simultaneous speech-to-speech translation model with voice cloning, trained with a two-stage reinforcement learning reward scheme, reports high accuracy and low latency on the authors' RealSI benchmark.

Reference graph

Works this paper leans on

35 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv

  2. [5]

    Besides, the score distribution shows that most of the GPT-generated translations scored 4 or higher

    The average scores of GPT-generated translations are close to those are manually revised. Besides, the score distribution shows that most of the GPT-generated translations scored 4 or higher. These results suggest that the GPT-generated data aligns well with SiMT human preferences. Table 5: Human evaluation on GPT-generated translations and manually revis...

    work page 2024

  3. [6]

    Improving Simultaneous Translation by Incorporating Pseudo-References with Fewer Reorderings

    Junkun Chen, Renjie Zheng, Atsuhito Kita, Mingbo Ma, and Liang Huang. Improving simul- taneous translation by incorporating pseudo-references with fewer reorderings. arXiv preprint arXiv:2010.11247,

    work page Pith review arXiv 2010

  4. [8]

    Efficient wait-k models for simultaneous machine translation

    Maha Elbayad, Laurent Besacier, and Jakob Verbeek. Efficient wait-k models for simultaneous machine translation. arXiv preprint arXiv:2005.08595,

    arXiv 2005

  5. [9]

    Kto: Model alignment as prospect theoretic optimization

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306,

    arXiv

  6. [10]

    Don’t until the final verb wait: Reinforcement learning for simultaneous machine translation

    11 Published as a conference paper at ICLR 2025 Alvin Grissom II, He He, Jordan Boyd-Graber, John Morgan, and Hal Daum ´e III. Don’t until the final verb wait: Reinforcement learning for simultaneous machine translation. In Proceedings of the 2014 Conference on empirical methods in natural language processing (EMNLP) , pp. 1342– 1352,

    work page 2025

  7. [13]

    [INST] <<SYS>> You are a professional translator

    Through these experiments, we further validate our findings: Both SimulDPO and MSFT improve model performance, with a more pronounced effect at low latency levels; although the effect is less pronounced in the De →En task compared to the Zh →En task, SimulPL’s performance is still influenced by α; SimulPL also generalizes well to other preference optimiza...

    work page 2025

  8. [14]

    Pref- erence transformer: Modeling human preferences using transformers for rl

    Changyeon Kim, Jongjin Park, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. Pref- erence transformer: Modeling human preferences using transformers for rl. arXiv preprint arXiv:2303.00957,

    arXiv

Show all 35 references
  1. [15]

    Findings of the 2022 conference on machine transla- tion (WMT22)

    Tom Kocmi, Rachel Bawden, Ondˇrej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Nov´ak, Martin...

    2022

  2. [19]

    Expectations of users of conference interpretation

    12 Published as a conference paper at ICLR 2025 Peter Moser. Expectations of users of conference interpretation. Interpreting, 1(2):145–178,

    2025

  3. [21]

    Disentangling length from quality in direct preference optimization

    Ryan Park, Rafael Rafailov, Stefano Ermon, and Chelsea Finn. Disentangling length from quality in direct preference optimization. arXiv preprint arXiv:2403.19159,

    arXiv

  4. [22]

    A call for clarity in reporting bleu scores

    Matt Post. A call for clarity in reporting bleu scores. WMT 2018, pp. 186,

    2018

  5. [23]

    Llama 2: Open founda- tion and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko- lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open founda- tion and fine-tuned chat models. arXiv preprint arXiv:2307.09288,

    arXiv

  6. [24]

    Simultaneous machine translation with large language models

    Minghan Wang, Jinming Zhao, Thuy-Trang Vu, Fatemeh Shiri, Ehsan Shareghi, and Gholam- reza Haffari. Simultaneous machine translation with large language models. arXiv preprint arXiv:2309.06706, 2023a. Minghan Wang, Thuy-Trang Vu, Ehsan Shareghi, and Gholamreza Haffari. Convers...

  7. [25]

    Better simultaneous translation with monotonic knowledge distillation

    Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pp. 2334–2349, 20...

    arXiv

  8. [26]

    Self-modifying state modeling for simultaneous machine translation

    13 Published as a conference paper at ICLR 2025 Donglei Yu, Xiaomian Kang, Yuchen Liu, Yu Zhou, and Chengqing Zong. Self-modifying state modeling for simultaneous machine translation. In Lun-Wei Ku, Andre Martins, and Vivek Sriku- mar (eds.), Proceedings of the 62nd Annual Mee...

    2025

  9. [27]

    Rrhf: Rank responses to align language models with human feedback without tears

    Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrhf: Rank responses to align language models with human feedback without tears. arXiv preprint arXiv:2304.05302,

    arXiv

  10. [28]

    Information-transport-based policy for simultaneous translation

    Shaolei Zhang and Yang Feng. Information-transport-based policy for simultaneous translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pp. 992–1013, Abu Dhabi, United Arab Emirates, December

    2022

  11. [29]

    URL https://aclanthology.org/2022.emnlp-main.65

    Association for Computational Linguistics. URL https://aclanthology.org/2022.emnlp-main.65. Shaolei Zhang and Yang Feng. Hidden markov transformer for simultaneous machine translation. arXiv preprint arXiv:2303.00257,

    2022 arXiv

  12. [31]

    User: As the speaker speaks, you need to: 1.Interpret the speaker's Chinese sentence into fluent English for the audience

    System: You are a professional simultaneous interpreter, simulating the scenario of interpreting a speaker's Chinese speech into English in real time. User: As the speaker speaks, you need to: 1.Interpret the speaker's Chinese sentence into fluent English for the audience. 2.Y...

    2022

  13. [33]

    We use the instruction-following format to guide the LLM in completing the SiMT task

    In the training process, we share the LoRA in the MSFT and SimulDPO phases. We use the instruction-following format to guide the LLM in completing the SiMT task. Our used prompt 18 Published as a conference paper at ICLR 2025 template is shown in Figure

    2025

  14. [34]

    (2020b) to apply greedy search during inference

    For a fair comparison between Transformer-based and LLM-based SiMT models, we following Ma et al. (2020b) to apply greedy search during inference. F N UMERICAL RESULTS Tables 9, 10, and 11 respectively present the numerical results of different SiMT models on the Zh→En, De→En,...

    2020

  15. [1952]

    Livemind: Low-latency large language models with simultaneous inference

    Chuangtao Chen, Grace Li Zhang, Xunzhao Yin, Cheng Zhuo, Ulf Schlichtmann, and Bing Li. Livemind: Low-latency large language models with simultaneous inference. arXiv preprint arXiv:2406.14319,

    arXiv

  16. [1985]

    A generative framework for simultaneous machine translation

    Yishu Miao, Phil Blunsom, and Lucia Specia. A generative framework for simultaneous machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 6697–6706,

    2021

  17. [1996]

    fairseq: A fast, extensible toolkit for sequence modeling

    Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of NAACL-HLT 2019: Demonstrations,

    2019

  18. [2001]

    Cross attention augmented transducer networks for simultaneous translation

    Dan Liu, Mengge Du, Xiaoxi Li, Ya Li, and Enhong Chen. Cross attention augmented transducer networks for simultaneous translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 39–55,

    2021

  19. [2010]

    Taking Zh→En SiMT task as an example, our complete human preferences prompts are shown in Figure

    14 Published as a conference paper at ICLR 2025 A D ATASET CONSTRUCTION AND ANALYSIS A.1 H UMAN PREFERENCE PROMPTS Based on our categorized SiMT human preferences, we construct human preference prompts, which account for translation quality preference, monotonicity preference,...

    2025

  20. [2013]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 ,

    arXiv

  21. [2016]

    Improving machine translation with human feedback: An exploration of quality estimation as a reward model

    Zhiwei He, Xing Wang, Wenxiang Jiao, Zhuosheng Zhang, Rui Wang, Shuming Shi, and Zhaopeng Tu. Improving machine translation with human feedback: An exploration of quality estimation as a reward model. arXiv preprint arXiv:2401.12873,

    arXiv

  22. [2017]

    Simultaneous machine translation with tailored ref- erence

    Shoutao Guo, Shaolei Zhang, and Yang Feng. Simultaneous machine translation with tailored ref- erence. arXiv preprint arXiv:2310.13588,

    arXiv

  23. [2020]

    Can neural machine translation do simultaneous translation? arXiv Preprint, CoRR, arXiv: abs/1606.02012,

    K Cho. Can neural machine translation do simultaneous translation? arXiv Preprint, CoRR, arXiv: abs/1606.02012,

    arXiv

  24. [2021]

    Quality of inter- preting from users’ perspectives

    Mansour Amini, Noraini Ibrahim-Gonz ´alez, Leelany Ayob, and Davoud Amini. Quality of inter- preting from users’ perspectives. International Journal of Language and Education , 2(1):2013,

    2013

  25. [2022]

    URL https://aclanthology.org/2022.wmt-1.1

    Association for Computational Linguistics. URL https://aclanthology.org/2022.wmt-1.1. Ingrid Kurz. Conference interpreting: Quality in the ears of the user. Meta, 46(2):394–409,

    2022

  26. [2023]

    Interpretese vs

    He He, Jordan Boyd-Graber, and Hal Daum ´e III. Interpretese vs. translationese: The uniqueness of human strategies in simultaneous interpretation. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2016

  27. [2024]

    Translation-based supervision for policy generation in simultaneous neural machine translation

    Ashkan Alinejad, Hassan S Shavarani, and Anoop Sarkar. Translation-based supervision for policy generation in simultaneous neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pp. 1734–1744,

    2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.