Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Standard segment-level MT benchmarks miss a long-context collapse in quantized EuroLLM models, while Hy-MT2 stays near baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 07:46 UTC pith:6TJKZP52

load-bearing objection A useful empirical paper pairing quantization and chunking for MT serving, with a large collapse result that is real for one GPTQ/SmoothQuant recipe but not yet established as intrinsic to EuroLLM. the 3 major comments →

arxiv 2607.29397 v1 pith:6TJKZP52 submitted 2026-07-31 cs.CL cs.PF

Studying quantization trade-offs for efficient inference deployment in machine translation

Jim Zhao , Sohir Maskey , Koen Oostermeijer , Douglas Orr , Teryn Jones This is my paper
classification cs.CL cs.PF
keywords quantizationmachine translationdocument-level evaluationchunkinginference efficiencylong-contextpost-training quantizationlatency-throughput trade-off
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to establish that standard sentence-level machine-translation benchmarks can mislead deployment decisions for quantized models. On long-document inputs, quantized EuroLLM-9B and 22B lose roughly half their translation quality at around 800 tokens, while quantized Hy-MT2 models stay close to the unquantized baseline—and the degradation is invisible in segment-level scores. The paper also demonstrates that pairing low-bit quantization with a 200–400 token document-chunking strategy improves the latency-throughput trade-off for medium-sized documents on a single datacenter GPU. If true, production evaluation should move from isolated-sentence tests to document-level tests.

Core claim

The core finding is empirical: in a document-level evaluation built from a segment-level MT benchmark, the translation quality of quantized EuroLLM-9B and 22B collapses as target prompt length grows, while Hy-MT2 is largely unaffected. For German→English at a target threshold of 800 tokens, EuroLLM-9B chrF++ (a character/word n-gram metric for full documents) drops from 64.93 (BF16) to 33.70 (W8A8, 8-bit weights and activations), 24.62 (W4A16, 4-bit weights only), and 27.03 (W4A8, 4-bit weights with 8-bit activations); the same checkpoints lose only about 4–7% in xCOMET, a segment-level neural metric. The paper interprets this as a model-family-specific interaction between quantization and l

What carries the argument

The central machinery is a document-chunking evaluation protocol: parallel documents are greedily concatenated into chunks that meet a target prompt-length threshold without splitting aligned segments, each chunk is translated independently, and outputs are concatenated and scored per document with chrF++. That protocol exposes context-length-dependent failures that segment-level metrics miss. On the efficiency side, a closed-loop online benchmark with synthetic 1,200-token documents, chunk sizes 50–400 tokens, and up to 500 concurrent users shows where the three quantization formats—W8A8 (8-bit weights and activations), W4A8 (4-bit weights, 8-bit activations), W4A16 (4-bit weights only)—shi

Load-bearing premise

The headline quality drops rest on at most 20 documents per translation direction, each scored once, and on a single quantization recipe; if those documents or that recipe are not representative, the collapse could be an artifact of the evaluation rather than a property of the models.

What would settle it

Re-run the document-level protocol on hundreds of held-out parallel documents per direction, and re-quantize the same models with an alternative calibration set or quantization-aware training. If EuroLLM chrF++ at 800-token chunks returns to within a few points of BF16 under any reasonable alternative recipe, the claimed intrinsic sensitivity is refuted; if a 40–60% drop persists across recipes and document samples, it stands.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For models at 9B parameters and above, W8A8 (on A100) and W4A8 (on H100) can improve throughput at similar latency; below about 2B, quantization gains are marginal or negative.
  • Chunking medium documents into 200–400 token pieces is Pareto-optimal in most serving settings, so chunk length and quantization format should be selected jointly rather than separately.
  • Segment-level MT evaluation can substantially understate quantized-model degradation, so deployment-time quality needs document-level measurement.
  • Different model families can respond very differently to the same quantization recipe: Hy-MT2 tracks BF16 closely, while EuroLLM-9B/22B degrade drastically with longer context.
  • Long-chunk failure modes include refusals, source copying, wrong-language summaries, degenerate repetition, and assistant-style responses instead of translations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper tests only one calibration recipe; EuroLLM's collapse could be a calibration-mismatch artifact rather than an intrinsic property, so an alternative recipe or quantization-aware training is a direct test.
  • With 11–20 documents per direction and one chrF++ run per document, the 48–62% numbers could be inflated by a handful of pathological documents; a larger corpus with repeated sampling would sharpen the estimate.
  • The same document-level protocol could be used to audit quantization in other long-context generation tasks, such as summarization or retrieval-augmented question answering, where segment-level metrics may hide similar failures.
  • Hardware conclusions are partly confounded by kernel support (W4A8 was not benchmarked on A100); a kernel-controlled comparison would separate format effects from kernel effects.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies quantization trade-offs for machine translation deployment on a single A100/H100 GPU, comparing five EuroLLM and Hy-MT2 models (1.7B–22B) under W8A8, W4A8, and W4A16 quantization. The authors report controlled vLLM-based offline and closed-loop online benchmarks showing that W8A8/W4A8 plus a 200–400 token document-chunking strategy improves the latency-throughput Pareto curve for medium-length documents, with small (<2B) models seeing little benefit. They additionally construct a document-level evaluation from WMT24++ across four translation directions and claim that standard segment-level benchmarks (e.g., xCOMET on WMT24++) underestimate long-context degradation: Hy-MT2 quantized checkpoints track the BF16 baseline, while EuroLLM-9B/22B exhibit drastic chrF++ drops at target prompt lengths of 800 tokens (e.g., EuroLLM-9B W8A8 chrF++ falls from 64.93 to 33.70, a 48.1% drop, despite only a 5.3% xCOMET drop on segment-level WMT24++).

Significance. If the central claim holds, the paper makes a valuable contribution: it demonstrates that deployment-oriented evaluation of quantized MT models requires document-level long-context assessment, not just segment-level quality metrics, and it provides a concrete efficiency benchmark under realistic serving conditions. The roofline model in Appendix A.1 is a parameter-free upper bound based on hardware specifications and gives useful context for interpreting the speedups. The qualitative failure modes in Appendix A.6 (refusals, source copying, repetitions) add credibility to the existence of severe degradation. However, the strength of the family-level conclusion is currently limited by the use of a single quantization recipe and by the small document sample, as detailed below.

major comments (3)
  1. [§3.3 and §6] The headline claim that 'EuroLLM models are sensitive to quantization and degrade rapidly with context length' (contribution 3, Introduction) is established only for one PTQ pipeline: GPTQ with SmoothQuant α=0.8 for W8A8 and α=0.4 for W4A8, calibrated on 1024 OPUS samples translated by Seed-X. The same hyperparameters are applied to both model families without an ablation or per-family calibration check. It is therefore possible that the observed EuroLLM collapse is a calibration mismatch artifact rather than an intrinsic property of the model family. Section 6 explicitly concedes that 'more sophisticated PTQ methods or QAT do not show degradation' is possible. This does not invalidate the existence of a segment-level benchmark blind spot for this particular deployment configuration, but it does not support the generalized family-level conclusion. Please either soften the claims to be pi
  2. [§4.3, Table 1, §A.3] The headline quantitative drops (48–62% for EuroLLM-9B at T=800) are computed from only 19 documents for DE→EN and 11–20 documents for the other directions, with a single chrF++ run per document. Table 1 reports point deltas without confidence intervals, although Figure 3 shows bootstrapped CIs. With N≈19, a few pathological documents of the kind shown in Figure 11 could drive the aggregate drop. Please report per-document score distributions or confidence intervals for the deltas, and ideally extend the document set; it would also help to state how many of the 19 documents individually show the collapse rather than only the mean.
  3. [§3.6, Figures 1–2] The efficiency comparisons (offline wall-time/throughput and closed-loop p99 document latency) are reported without any indication of the number of repetitions or run-to-run variance. Some of the conclusions, such as W8A8 vs W4A8 Pareto ordering on the H100 or the 'throughput reduces at large batch size' observation for the 22B model, depend on differences that could be within measurement noise. If the benchmarks were repeated, please report means with error bars and the number of runs; if not, add repeated runs to support the efficiency claims at the level of precision with which they are stated.
minor comments (5)
  1. [Table 1 and §3.5.1] The table header labels the first metric as 'WMT24++' without explicitly saying it is the segment-level xCOMET result; clarify that this is the standard isolated-segment protocol, distinct from the document-level chrF++ columns.
  2. [Figure captions 3, 8–10] The axis label reads 'T arget prompt-length threshold' with an unintended space; fix the typo. Also, the figure captions say 'averaged across a subset' but could state which N corresponds to each panel for clarity.
  3. [Table 2 (Appendix A.1)] The architectural table is mangled in rendering (e.g., '563216/810241.657B'), making the hidden dimension, head counts, and parameter counts hard to read. Please format the table properly.
  4. [Figure 2 caption] The caption text '1 500 500' and '500 500' appears garbled; clarify which marker positions correspond to which concurrency levels.
  5. [§3.5.2 and §A.3] The document inclusion rule (exclude documents below 0.8T and then intersect valid documents across all models/formats/thresholds) may introduce a selection bias toward longer, harder documents. Please state explicitly what fraction of the original WMT24++ documents survive the intersection, since the headline results are conditional on this subset.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claims are direct empirical measurements against external benchmarks and a parameter-free roofline model.

full rationale

The paper's main claims are empirical: quantized EuroLLM degrades on long document-level translation while Hy-MT2 does not, and chunking with W4A8/W8A8 improves latency-throughput Pareto curves. These are evaluated against external reference points (WMT24++, chrF++, xCOMET, vLLM, A100/H100) rather than being derived from definitions or fitted parameters. The roofline model in Appendix A.1 is an idealized upper bound computed from hardware specifications (Ppeak, HBM bandwidth, model dimensions) and explicitly labeled as an upper bound; it is not fitted to the reported speedups and the paper even notes where observations depart from it. The translation-quality collapse is measured directly on held-out documents, not predicted from a fitted model. The only apparent self-citation, Maskey et al. (2026), appears in Section 3.2 as part of a list justifying exclusion of sub-4-bit formats; this is peripheral and does not load-bear on the central quantization-quality or efficiency findings. The limitations stated in Section 6 — that only GPTQ+SmoothQuant was tested, that other PTQ or QAT methods might not degrade, and that chrF++ was computed over only 11–20 documents — are validity and generality caveats, not circularity. No step in the paper's derivation reduces by construction to its inputs, and no fitted parameter is renamed as a prediction. Therefore the appropriate finding is no significant circularity, score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central empirical claims rest on measured workloads and translated outputs rather than on a fitted model. The main 'paid-for' inputs are hand-chosen quantization hyperparameters (SmoothQuant α, GPTQ calibration subset) and small document-evaluation subsets. The roofline model is an acknowledged idealized upper bound and does not generate the headline quality results.

free parameters (3)
  • SmoothQuant smoothing strength α = 0.8 for W8A8, 0.4 for W4A8
    Chosen by hand (Section 3.3) with no reported sensitivity analysis; directly controls activation outlier suppression and thus both quantization quality and speed.
  • GPTQ calibration subset = 1024 samples subsampled from OPUS; target translations generated by Seed-X
    The exact language/pair composition is unspecified; calibration distribution can strongly affect PTQ quality and is a candidate explanation for EuroLLM's collapse.
  • Document inclusion threshold factor = 0.8 × target prompt threshold T
    Ad hoc exclusion rule in Section 3.5.2; documents shorter than 0.8T are excluded, and common-subset intersection across all formats/thresholds shapes the evaluation set.
axioms (4)
  • domain assumption Random token IDs are representative of real MT prompts and generations for latency/throughput benchmarks.
    Section 3.6: 'We benchmark the models using randomly generated token IDs for both inference evaluations.' This simplifies scheduling behavior but ignores real tokenization, vocabulary effects, and output length distributions.
  • domain assumption Chunk-wise independent decoding with concatenated outputs captures document-level translation quality.
    Section 3.5.2: 'Each chunk is passed individually to the model and the generated outputs are concatenated in their original order.' The evaluation measures long-chunk translation, not context-aware decoding across chunk boundaries.
  • domain assumption A single GPTQ+SmoothQuant recipe is a fair comparator across EuroLLM and Hy-MT2.
    The paper applies one PTQ recipe to both model families; if the calibration data or hyperparameters are mismatched to EuroLLM, the observed robustness difference could be an artifact. Section 6 notes only GPTQ was considered.
  • standard math Roofline model assumes GEMM-only traffic and peak HBM bandwidth.
    Appendix A.1 is an explicitly idealized upper bound; the paper acknowledges it ignores dequantization overhead, kernel utilization, and KV-cache effects. It is used for insight rather than as the headline empirical claim.

pith-pipeline@v1.3.0-daily-deepseek · 17642 in / 11313 out tokens · 130595 ms · 2026-08-03T07:46:50.120651+00:00 · methodology

0 comments
read the original abstract

Deploying large language models in realistic server environments poses challenges, as the system needs to provide high-quality responses with low latency. Quantization is a common approach to reduce the memory footprint and improve inference efficiency, yet its impact on latency and throughput is rarely evaluated under controlled, orchestration-level workloads. In this work we study the quantization trade-offs of two translation model families, EuroLLM \citep{martins2025eurollm} and Hy-MT2 \citep{zheng2026hy} across five models ranging from 1.7B to 22B for efficient deployment on a single A100 or H100 GPU. We demonstrate that combining a document-chunking strategy with W4A8 or W8A8 quantization improves the latency-throughput Pareto-curve under a wide range of workloads. Furthermore, since standard machine translation (MT) benchmarks rely on isolated sentences and fail to capture long-context dynamics, we introduce a document-level evaluation from WMT24++ to assess how text chunking strategies affect translation quality under quantization. Our results reveal that standard segment-level evaluation can fail to predict the interaction between quantization and long-context document translation. While Hy-MT2 remains robust under quantization, EuroLLM shows strong sensitivity and translation quality collapses rapidly for all considered quantization formats. Overall, our experiments show that the trade-off between inference efficiency and translation quality depends not only on the quantization format, but also on the choice of text chunking strategy.

Figures

Figures reproduced from arXiv: 2607.29397 by Douglas Orr, Jim Zhao, Koen Oostermeijer, Sohir Maskey, Teryn Jones.

Figure 1
Figure 1. Figure 1: Offline benchmark measuring average output throughput in tok/s against average wall time for a sequence [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Closed-loop online benchmark for user concurrency from 1 to 500 users measuring throughput in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Chunk length vs. chrF++ Score for Hy-MT2 (top row) and EuroLLM (bottom row) models across different [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Memory-bandwidth upper bound on the speedup from BF16 to INT4 weights for the three EuroLLM [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Offline benchmark measuring average output throughput in tok/s against average wall time for a sequence [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Offline benchmark measuring average output throughput in tok/s against average wall time for a sequence [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Offline benchmark measuring average output throughput in tok/s against average wall time for a sequence [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Chunk length vs. chrF++ Score for Hy-MT2 (top row) and EuroLLM (bottom row) models across different [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Chunk length vs. chrF++ Score for Hy-MT2 (top row) and EuroLLM (bottom row) models across different [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Chunk length vs. chrF++ Score for Hy-MT2 (top row) and EuroLLM (bottom row) models across [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative failure modes of quantized EuroLLM-9B at target threshold [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 18 linked inside Pith

  1. [1]

    Findings of the WMT 2024 Shared Task of the Open Language Data Initiative

    Burchell, Laurie and Maillard, Jean and Anastasopoulos, Antonios and Federmann, Christian and Koehn, Philipp and Wang, Skyler. Findings of the WMT 2024 Shared Task of the Open Language Data Initiative. Proceedings of the Ninth Conference on Machine Translation. 2024. doi:10.18653/v1/2024.wmt-1.4

    doi:10.18653/v1/2024.wmt-1.4 2024

  2. [2]

    Findings of the WMT 2025 Shared Task of the Open Language Data Initiative

    Dale, David and Burchell, Laurie and Maillard, Jean and Abdulmumin, Idris and Anastasopoulos, Antonios and Caswell, Isaac and Koehn, Philipp. Findings of the WMT 2025 Shared Task of the Open Language Data Initiative. Proceedings of the Tenth Conference on Machine Translation. 2025. doi:10.18653/v1/2025.wmt-1.26

    doi:10.18653/v1/2025.wmt-1.26 2025

  3. [3]

    NLLB Team and Costa-juss \`a , Marta R. and Cross, James and C elebi, Onur and Elbayad, Maha and Heafield, Kenneth and Heffernan, Kevin and Kalbassi, Elahe and Lam, Janice and Licht, Daniel and Maillard, Jean and Sun, Anna and Wang, Skyler and Wenzek, Guillaume and Youngblood, Al and Akula, Bapi and Barrault, Loic and Gonzalez, Gabriel Mejia and Hansanti,...

    2024

  4. [4]

    Goyal, Naman and Gao, Cynthia and Chaudhary, Vishrav and Chen, Peng-Jen and Wenzek, Guillaume and Ju, Da and Krishnan, Sanjana and Ranzato, Marc’Aurelio and Guzmán, Francisco and Fan, Angela , journal =. The

  5. [5]

    arXiv preprint arXiv:2509.25149 , year=

    Pretraining large language models with nvfp4 , author=. arXiv preprint arXiv:2509.25149 , year=

    arXiv

  6. [6]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  7. [7]

    18th USENIX symposium on operating systems design and implementation (OSDI 24) , pages=

    Taming \ Throughput-Latency \ tradeoff in \ LLM \ inference with \ Sarathi-Serve \ , author=. 18th USENIX symposium on operating systems design and implementation (OSDI 24) , pages=

  8. [8]

    arXiv preprint arXiv:1308.3432 , year=

    Estimating or propagating gradients through stochastic neurons for conditional computation , author=. arXiv preprint arXiv:1308.3432 , year=

    Pith/arXiv arXiv

  9. [9]

    Proceedings of the 14th International Conference on Spoken Language Translation , pages=

    Overview of the IWSLT 2017 evaluation campaign , author=. Proceedings of the 14th International Conference on Spoken Language Translation , pages=

    2017

  10. [10]

    arXiv preprint arXiv:2507.13618 , year=

    Seed-x: Building strong multilingual translation llm with 7b parameters , author=. arXiv preprint arXiv:2507.13618 , year=

    Pith/arXiv arXiv

  11. [11]

    IEEE Micro , volume=

    Nvidia hopper h100 gpu: Scaling performance , author=. IEEE Micro , volume=. 2023 , publisher=

    2023

  12. [12]

    Proceedings of the 28th international conference on evaluation and assessment in software engineering , pages=

    A performance study of llm-generated code on leetcode , author=. Proceedings of the 28th international conference on evaluation and assessment in software engineering , pages=

  13. [13]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    WMT24++: Expanding the language coverage of WMT24 to 55 languages & dialects , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  14. [14]

    arXiv preprint arXiv:2406.17415 , year=

    Layer-wise quantization: A pragmatic and effective method for quantizing llms beyond integer bit-levels , author=. arXiv preprint arXiv:2406.17415 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2210.17323 , year=

    Gptq: Accurate post-training quantization for generative pre-trained transformers , author=. arXiv preprint arXiv:2210.17323 , year=

    Pith/arXiv arXiv

  16. [16]

    Transactions of the Association for Computational Linguistics , volume=

    xcomet: Transparent machine translation evaluation through fine-grained error detection , author=. Transactions of the Association for Computational Linguistics , volume=. 2024 , publisher=

    2024

  17. [17]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Kvquant: Towards 10 million context length llm inference with kv cache quantization , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    2026 IEEE International Parallel and Distributed Processing Symposium (IPDPS) , pages=

    Microbenchmarking nvidia’s blackwell architecture: An in-depth architectural analysis , author=. 2026 IEEE International Parallel and Distributed Processing Symposium (IPDPS) , pages=. 2026 , organization=

    2026

  20. [20]

    Proceedings of the Eighth Conference on Machine Translation , pages=

    MetricX-23: The Google submission to the WMT 2023 metrics shared task , author=. Proceedings of the Eighth Conference on Machine Translation , pages=

    2023

  21. [21]

    Large Language Models Effectively Leverage Document-level Context for Literary Translation, but Critical Errors Persist

    Karpinska, Marzena and Iyyer, Mohit. Large Language Models Effectively Leverage Document-level Context for Literary Translation, but Critical Errors Persist. Proceedings of the Eighth Conference on Machine Translation. 2023. doi:10.18653/v1/2023.wmt-1.41

    doi:10.18653/v1/2023.wmt-1.41 2023

  22. [22]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    Pith/arXiv arXiv

  23. [23]

    Proceedings of the 29th symposium on operating systems principles , pages=

    Efficient memory management for large language model serving with pagedattention , author=. Proceedings of the 29th symposium on operating systems principles , pages=

  24. [24]

    Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation

    L. Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1512

    doi:10.18653/v1/d18-1512 2018

  25. [25]

    GetMobile: Mobile Computing and Communications , volume=

    Awq: Activation-aware weight quantization for on-device llm compression and acceleration , author=. GetMobile: Mobile Computing and Communications , volume=. 2025 , publisher=

    2025

  26. [26]

    arXiv preprint arXiv:2402.18158 , year=

    Evaluating quantized large language models , author=. arXiv preprint arXiv:2402.18158 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2411.15692 , year=

    Drugagent: Automating ai-aided drug discovery programming through llm multi-agent collaboration , author=. arXiv preprint arXiv:2411.15692 , year=

    Pith/arXiv arXiv

  28. [28]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

  29. [29]

    Nature medicine , volume=

    A generalist medical language model for disease diagnosis assistance , author=. Nature medicine , volume=. 2025 , publisher=

    2025

  30. [30]

    Amin and Bawden, Rachel and Zhang, Michael and Martins, Andr \'e F

    Lopes, Ant \'o nio and Farajian, M. Amin and Bawden, Rachel and Zhang, Michael and Martins, Andr \'e F. T. Document-level Neural MT : A Systematic Comparison. Proceedings of the 22nd Annual Conference of the European Association for Machine Translation. 2020

    2020

  31. [31]

    arXiv preprint arXiv:2402.17764 , year=

    The era of 1-bit llms: All large language models are in 1.58 bits , author=. arXiv preprint arXiv:2402.17764 , year=

    Pith/arXiv arXiv

  32. [32]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    How does quantization affect multilingual LLMs? , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  33. [33]

    arXiv preprint arXiv:2508.20893 , year=

    The Uneven Impact of Post-Training Quantization in Machine Translation , author=. arXiv preprint arXiv:2508.20893 , year=

    Pith/arXiv arXiv

  34. [34]

    Procedia Computer Science , volume=

    Eurollm: Multilingual language models for europe , author=. Procedia Computer Science , volume=. 2025 , publisher=

    2025

  35. [35]

    ACM Computing Surveys (CSUR) , volume=

    A survey on document-level neural machine translation: Methods and evaluation , author=. ACM Computing Surveys (CSUR) , volume=. 2021 , publisher=

    2021

  36. [36]

    arXiv preprint arXiv:2602.15563 , year=

    1-Bit Wonder: Improving QAT Performance in the Low-Bit Regime through K-Means Quantization , author=. arXiv preprint arXiv:2602.15563 , year=

    arXiv

  37. [37]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Data-free quantization through weight equalization and bias correction , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  38. [38]

    2023 , howpublished =

    2023

  39. [39]

    arXiv preprint arXiv:2203.02155 , year=

    Training language models to follow instructions with human feedback , author=. arXiv preprint arXiv:2203.02155 , year=

    Pith/arXiv arXiv

  40. [40]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  41. [41]

    Proceedings of the second conference on machine translation , pages=

    chrF++: words helping character n-grams , author=. Proceedings of the second conference on machine translation , pages=

  42. [42]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Is semantic chunking worth the computational cost? , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  43. [43]

    arXiv preprint arXiv:2310.10537 , year=

    Microscaling data formats for deep learning , author=. arXiv preprint arXiv:2310.10537 , year=

    Pith/arXiv arXiv

  44. [44]

    Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026)-Vol

    Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3 , author=. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026)-Vol. 1 , pages=

    2026

  45. [45]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  46. [46]

    Proceedings of the 19th Annual Conference of the European Association for Machine Translation: Projects/Products , year=

    OPUS--parallel corpora for everyone , author=. Proceedings of the 19th Annual Conference of the European Association for Machine Translation: Projects/Products , year=

  47. [47]

    arXiv preprint arXiv:2302.13971 , year=

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    Pith/arXiv arXiv

  48. [48]

    The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives

    Voita, Elena and Sennrich, Rico and Titov, Ivan. The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2...

    doi:10.18653/v1/d19-1448 2019

  49. [49]

    Document-Level Machine Translation with Large Language Models

    Wang, Longyue and Lyu, Chenyang and Ji, Tianbo and Zhang, Zhirui and Yu, Dian and Shi, Shuming and Tu, Zhaopeng. Document-Level Machine Translation with Large Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.1036

    doi:10.18653/v1/2023.emnlp-main.1036 2023

  50. [50]

    arXiv preprint arXiv:2411.13476 , year=

    When precision meets position: Bfloat16 breaks down rope in long-context training , author=. arXiv preprint arXiv:2411.13476 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2310.11453 , year=

    Bitnet: Scaling 1-bit transformers for large language models , author=. arXiv preprint arXiv:2310.11453 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2510.13998 , year=

    BitNet Distillation , author=. arXiv preprint arXiv:2510.13998 , year=

    arXiv

  53. [53]

    International conference on machine learning , pages=

    Smoothquant: Accurate and efficient post-training quantization for large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  54. [54]

    arXiv preprint arXiv:2601.20088 , year=

    Quantization-Aware Distillation for NVFP4 Inference Accuracy Recovery , author=. arXiv preprint arXiv:2601.20088 , year=

    arXiv

  55. [55]

    16th USENIX symposium on operating systems design and implementation (OSDI 22) , pages=

    Orca: A distributed serving system for \ Transformer-Based \ generative models , author=. 16th USENIX symposium on operating systems design and implementation (OSDI 22) , pages=

  56. [56]

    2025 International Joint Conference on Neural Networks (IJCNN) , pages=

    Lsaq: Layer-specific adaptive quantization for large language model deployment , author=. 2025 International Joint Conference on Neural Networks (IJCNN) , pages=. 2025 , organization=

    2025

  57. [57]

    2024 IEEE International Conference on Multimedia and Expo (ICME) , pages=

    Integer or floating point? new outlooks for low-bit quantization on large language models , author=. 2024 IEEE International Conference on Multimedia and Expo (ICME) , pages=. 2024 , organization=

    2024

  58. [58]

    arXiv preprint arXiv:2406.09904 , year=

    Qqq: Quality quattuor-bit quantization for large language models , author=. arXiv preprint arXiv:2406.09904 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2605.22064 , year=

    Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild , author=. arXiv preprint arXiv:2605.22064 , year=

    Pith/arXiv arXiv