Prompt Compression in Diffusion Large Language Models: Evaluating LLMLingua-2 on LLaDA

Abigayle Brown; Jiakang Xu; Jiyoo Noh; Jonathan Chan; Kaung Myat Kyaw; Sterling Huang; Wantong Huo

arxiv: 2605.17932 · v1 · pith:4ONEMQWCnew · submitted 2026-05-18 · 💻 cs.CL · cs.AI

Prompt Compression in Diffusion Large Language Models: Evaluating LLMLingua-2 on LLaDA

Sterling Huang , Abigayle Brown , Jiyoo Noh , Jiakang Xu , Wantong Huo , Kaung Myat Kyaw , Jonathan Chan This is my paper

Pith reviewed 2026-05-20 11:13 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords prompt compressiondiffusion large language modelsLLaDALLMLingua-2semantic similarityinformation omissionmathematical reasoningsummarization

0 comments

The pith

Prompt compression methods for autoregressive models do not transfer uniformly to diffusion large language models.

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

This paper tests whether prompt compression techniques built for standard autoregressive language models can be used directly on diffusion large language models. It applies LLMLingua-2 at an approximate 2x compression ratio to the 8B-parameter LLaDA model across math reasoning, summarization, and general prompt tasks on three datasets. Results show that high semantic similarity between original and compressed prompts does not ensure stable model outputs, with mathematical reasoning degrading more than summarization. A reader would care because diffusion models generate text through a different process than autoregressive ones, so borrowing efficiency tools may not save compute without hurting accuracy. The work points to information omission as the main driver of failures rather than broad meaning shifts.

Core claim

The paper claims that prompt compression methods designed for autoregressive models do not transfer uniformly to diffusion large language models. On LLaDA with LLMLingua-2 at roughly 2x compression, summarization tasks proved comparatively robust while mathematical reasoning on GSM8K showed substantial degradation despite strong semantic similarity scores from BERTScore and other metrics. Reconstruction experiments indicated that semantically similar prompts can still omit reasoning-critical details required for stable denoising. BERTScore recall consistently fell below precision, suggesting compression issues stem mainly from information omission rather than semantic drift. These patterns,

What carries the argument

Evaluation of LLMLingua-2 compression on the diffusion model LLaDA, using output comparisons via exact-match accuracy, BLEU, ROUGE, and BERTScore for original, compressed, and reconstructed prompts across GSM8K, DUC2004, and ShareGPT.

If this is right

Summarization tasks remain more stable under compression than mathematical reasoning in diffusion models.
Compression failures arise mainly from omitted information rather than overall semantic changes.
Prompts that retain high semantic similarity can still lack details needed for reliable diffusion denoising.
New compression approaches tailored to the diffusion process are needed instead of direct adaptation from autoregressive methods.

Where Pith is reading between the lines

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

Application developers using diffusion LLMs should validate compression effects on their specific reasoning or generation tasks rather than relying on general semantic metrics.
The focus on information omission suggests value in creating new evaluation metrics that check preservation of step-by-step reasoning chains.
Similar sensitivity to prompt changes may appear in other non-autoregressive model families, warranting parallel tests.

Load-bearing premise

The study assumes that results from the selected datasets, a fixed approximate 2x compression ratio, and 250 prompts per task can support general claims about non-uniform transferability to diffusion large language models.

What would settle it

Testing LLMLingua-2 or similar compressors on additional diffusion LLMs at varying compression ratios and finding consistent performance across tasks would support the claim of non-uniform transfer; uniform success across models would challenge it.

Figures

Figures reproduced from arXiv: 2605.17932 by Abigayle Brown, Jiakang Xu, Jiyoo Noh, Jonathan Chan, Kaung Myat Kyaw, Sterling Huang, Wantong Huo.

read the original abstract

Prompt compression reduces inference cost and context length in large language models, but prior evaluations focus primarily on autoregressive architectures. This study investigates whether prompt compression transfers effectively to diffusion large language models (DLLMs) using LLMLingua-2, specifically the 8B-parameter DLLM LLaDA. We evaluate compression performance on GSM8K, DUC2004, and ShareGPT using 250 prompts per dataset at an approximate 2$\times$ compression ratio, across mathematical reasoning, prompt reconstruction, and summarization tasks. Outputs generated from original prompts, compressed prompts, reconstructed prompts, and reconstructed-prompt reasoning were compared using exact-match accuracy, BLEU, ROUGE, and BERTScore. Results show that semantic preservation does not necessarily imply stable downstream behavior in diffusion models. Summarization tasks remained comparatively robust under compression, while mathematical reasoning degraded substantially despite high semantic similarity scores. Reconstruction experiments further showed that semantically similar prompts may still omit reasoning-critical information required for stable denoising. Across tasks, BERTScore recall was consistently lower than precision, suggesting that compression failures are primarily driven by information omission rather than semantic drift. These findings indicate that prompt compression methods designed for autoregressive models do not transfer uniformly to diffusion large language models and motivate the development of diffusion-aware compression strategies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLMLingua-2 on LLaDA preserves semantics but tanks reasoning accuracy, yet the non-uniform transfer claim rests on inference without an AR baseline.

read the letter

The main takeaway is that LLMLingua-2 compression on the 8B diffusion model LLaDA keeps BERTScore high but cuts exact-match accuracy on GSM8K math problems substantially, while summarization on DUC2004 stays more stable. Reconstruction tests also show that semantically close prompts can still drop reasoning-critical details that the diffusion denoising process needs. This is the first reported evaluation of an established autoregressive compressor on a diffusion LLM, and the task split plus the precision-recall gap in BERTScore are the clearest signals in the results. The experiments use 250 prompts per dataset at a fixed roughly 2x ratio and compare original, compressed, and reconstructed versions with standard metrics, which gives a practical picture of where things break. The reconstruction angle is particularly useful because it ties the performance drop to information omission rather than vague semantic drift. The soft spots are straightforward. There is no side-by-side run on a comparable autoregressive model under identical prompts, compression, and metrics, so the claim that these methods do not transfer uniformly remains an inference instead of a direct contrast. The write-up gives no variance numbers, no statistical tests, and no exploration of other compression ratios or model sizes. Those gaps make it harder to generalize beyond this single 8B checkpoint and these three tasks. This paper is aimed at people working on efficiency for diffusion-based LLMs or anyone who needs to know where off-the-shelf compression tools fall short on reasoning. It is the kind of targeted benchmark that can steer follow-up work even if the current evidence is preliminary. I would send it for peer review. The core observation is worth referee scrutiny, but the authors should add an autoregressive control and basic statistical reporting before it goes further.

Referee Report

2 major / 2 minor

Summary. The paper evaluates LLMLingua-2 prompt compression on the 8B diffusion LLM LLaDA across GSM8K (math reasoning), DUC2004 (summarization), and ShareGPT (general prompts) using 250 prompts per task at ~2× compression. It compares outputs from original, compressed, reconstructed, and reconstructed-reasoning prompts via exact-match accuracy, BLEU, ROUGE, and BERTScore, concluding that semantic similarity (high BERTScore) does not ensure stable downstream performance in DLLMs, with math accuracy dropping sharply while summarization is more robust, and that failures arise mainly from information omission rather than drift.

Significance. If the empirical patterns hold, the work supplies useful early evidence that AR-optimized compression techniques can produce unstable behavior under diffusion denoising even when semantic metrics look acceptable, thereby motivating targeted research on diffusion-aware compression. The multi-task design and reconstruction experiments are strengths that go beyond single-metric evaluations.

major comments (2)

[Experimental setup / Results] The central claim that AR-designed methods 'do not transfer uniformly' to DLLMs rests on performance differences observed only on LLaDA. No parallel control experiment on a comparable autoregressive model (e.g., Llama-3-8B) with identical prompts, compression ratio, and metrics is reported; without this contrast it is not possible to isolate diffusion-specific effects from properties of LLMLingua-2 or the chosen tasks. This directly affects the load-bearing inference in the abstract and results discussion.
[Methods] The manuscript provides no information on statistical significance testing, run-to-run variance, or exact controls for prompt-length effects after compression. These omissions make it difficult to judge whether the reported accuracy drops (especially on GSM8K) are reliable or could be explained by length or sampling artifacts.

minor comments (2)

[Datasets] Clarify how the 250 prompts per dataset were sampled and whether any length or difficulty stratification was applied; this would strengthen claims of representativeness.
[Results] The abstract states 'BERTScore recall was consistently lower than precision'—include the per-task numerical values and standard deviations in the main results table for transparency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Experimental setup / Results] The central claim that AR-designed methods 'do not transfer uniformly' to DLLMs rests on performance differences observed only on LLaDA. No parallel control experiment on a comparable autoregressive model (e.g., Llama-3-8B) with identical prompts, compression ratio, and metrics is reported; without this contrast it is not possible to isolate diffusion-specific effects from properties of LLMLingua-2 or the chosen tasks. This directly affects the load-bearing inference in the abstract and results discussion.

Authors: We agree that a parallel control experiment on an autoregressive model would help isolate diffusion-specific effects and strengthen the claim of non-uniform transfer. Our primary focus was on characterizing behavior within the diffusion LLM, where we observe that high BERTScore does not ensure stable task performance (particularly on GSM8K). To directly address this concern, we will add a control comparison using Llama-3-8B under identical prompts, compression ratio, and metrics in the revised manuscript. revision: yes
Referee: [Methods] The manuscript provides no information on statistical significance testing, run-to-run variance, or exact controls for prompt-length effects after compression. These omissions make it difficult to judge whether the reported accuracy drops (especially on GSM8K) are reliable or could be explained by length or sampling artifacts.

Authors: We thank the referee for highlighting these methodological gaps. In the revision we will add statistical significance testing (bootstrap resampling with 95% confidence intervals) for key performance differences, report run-to-run variance across multiple sampling seeds, and include an explicit analysis of prompt-length effects post-compression to rule out length or sampling artifacts as explanations for the observed drops. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmarking with no derivations or self-referential predictions

full rationale

This is a standard empirical evaluation paper that measures LLMLingua-2 compression performance on the LLaDA diffusion model across three tasks using fixed metrics and prompt counts. No equations, fitted parameters, uniqueness theorems, or derivation chains appear in the abstract or described methodology. All claims rest on direct output comparisons (exact-match, BLEU, ROUGE, BERTScore) rather than any reduction of a 'prediction' to its own inputs by construction. The absence of an AR-model control experiment is a potential experimental-design limitation but does not constitute circularity under the defined criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

As an empirical evaluation, the claim rests on the assumption that standard NLP metrics capture the relevant differences in diffusion model behavior and that the selected tasks and compression level generalize. No free parameters or invented entities are introduced.

axioms (1)

domain assumption BERTScore, BLEU, and ROUGE reliably distinguish information omission from semantic drift in compressed prompts for downstream task performance.
Paper uses these metrics to conclude that failures are driven by omission rather than drift.

pith-pipeline@v0.9.0 · 5784 in / 1317 out tokens · 41831 ms · 2026-05-20T11:13:35.072851+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Results show that semantic preservation does not necessarily imply stable downstream behavior in diffusion models... BERTScore recall was consistently lower than precision, suggesting that compression failures are primarily driven by information omission rather than semantic drift.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 6 internal anchors

[1]

Llmlingua: Com- pressing prompts for accelerated inference of large language models,

H. Jiang, Q. Wu, C.-Y . Lin, Y . Yang, and L. Qiu, “Llmlingua: Com- pressing prompts for accelerated inference of large language models,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 13 358–13 376

work page 2023
[2]

Longllmlingua: Accelerating and enhancing llms in long context sce- narios via prompt compression,

H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y . Lin, Y . Yang, and L. Qiu, “Longllmlingua: Accelerating and enhancing llms in long context sce- narios via prompt compression,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 1658–1677

work page 2024
[3]

Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,

Z. Pan, Q. Wu, H. Jiang, M. Xia, X. Luo, J. Zhang, Q. Lin, V . R ¨uhle, Y . Yang, C.-Y . Linet al., “Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 963– 981

work page 2024
[4]

An empirical study on prompt compression for large language models,

Z. Zhang, J. Li, Y . Lan, X. Wang, and H. Wang, “An empirical study on prompt compression for large language models,”arXiv preprint arXiv:2505.00019, 2025

work page arXiv 2025
[5]

The Llama 3 Herd of Models

A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[6]

Qwen3 Technical Report

A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

DeepSeek-V3 Technical Report

A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[8]

Efficient unsupervised sentence compression by fine-tuning transformers with reinforcement learning,

D. Ghalandari, C. Hokamp, and G. Ifrim, “Efficient unsupervised sentence compression by fine-tuning transformers with reinforcement learning,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 1267– 1280

work page 2022
[9]

Kimi K2: Open Agentic Intelligence

K. Team, Y . Bai, Y . Bao, Y . Charles, C. Chen, G. Chen, H. Chen, H. Chen, J. Chen, N. Chenet al., “Kimi k2: Open agentic intelligence,” arXiv preprint arXiv:2507.20534, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[10]

Keep it simple: Unsupervised simplification of multi-paragraph text,

P. Laban, T. Schnabel, P. Bennett, and M. A. Hearst, “Keep it simple: Unsupervised simplification of multi-paragraph text,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021, pp. 6365–6378

work page 2021
[11]

Large language diffusion models,

S. Nie, F. Zhu, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y . Lin, J.-R. Wen, and C. Li, “Large language diffusion models,”Advances in Neural Information Processing Systems, vol. 38, pp. 50 608–50 646, 2026

work page 2026
[12]

BERTScore: Evaluating Text Generation with BERT

T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi, “Bertscore: Evaluating text generation with bert,”arXiv preprint arXiv:1904.09675, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1904
[13]

Training Verifiers to Solve Math Word Problems

K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakanoet al., “Training verifiers to solve math word problems,”arXiv preprint arXiv:2110.14168, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[14]

The effects of human variation in DUC summarization evaluation,

D. Harman and P. Over, “The effects of human variation in DUC summarization evaluation,” inProceedings of the ACL 2004 Workshop on Text Summarization Branches Out, 2004, pp. 10–17

work page 2004
[15]

Sharegpt4v: Improving large multi-modal models with better captions,

L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 370– 387

work page 2024
[16]

Unsu- pervised cross-lingual representation learning at scale,

A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzm´an, E. Grave, M. Ott, L. Zettlemoyer, and V . Stoyanov, “Unsu- pervised cross-lingual representation learning at scale,” inProceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 8440–8451

work page 2020
[17]

Bleu: a method for automatic evaluation of machine translation,

K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” inProceedings of the 40th annual meeting of the Association for Computational Linguistics, 2002, pp. 311–318

work page 2002
[18]

Rouge: A package for automatic evaluation of summaries,

C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” inText summarization branches out, 2004, pp. 74–81

work page 2004
[19]

Compressing context to enhance inference efficiency of large language models,

Y . Li, B. Dong, F. Guerin, and C. Lin, “Compressing context to enhance inference efficiency of large language models,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 6342–6353

work page 2023

[1] [1]

Llmlingua: Com- pressing prompts for accelerated inference of large language models,

H. Jiang, Q. Wu, C.-Y . Lin, Y . Yang, and L. Qiu, “Llmlingua: Com- pressing prompts for accelerated inference of large language models,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 13 358–13 376

work page 2023

[2] [2]

Longllmlingua: Accelerating and enhancing llms in long context sce- narios via prompt compression,

H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y . Lin, Y . Yang, and L. Qiu, “Longllmlingua: Accelerating and enhancing llms in long context sce- narios via prompt compression,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 1658–1677

work page 2024

[3] [3]

Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,

Z. Pan, Q. Wu, H. Jiang, M. Xia, X. Luo, J. Zhang, Q. Lin, V . R ¨uhle, Y . Yang, C.-Y . Linet al., “Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 963– 981

work page 2024

[4] [4]

An empirical study on prompt compression for large language models,

Z. Zhang, J. Li, Y . Lan, X. Wang, and H. Wang, “An empirical study on prompt compression for large language models,”arXiv preprint arXiv:2505.00019, 2025

work page arXiv 2025

[5] [5]

The Llama 3 Herd of Models

A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[6] [6]

Qwen3 Technical Report

A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[7] [7]

DeepSeek-V3 Technical Report

A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[8] [8]

Efficient unsupervised sentence compression by fine-tuning transformers with reinforcement learning,

D. Ghalandari, C. Hokamp, and G. Ifrim, “Efficient unsupervised sentence compression by fine-tuning transformers with reinforcement learning,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 1267– 1280

work page 2022

[9] [9]

Kimi K2: Open Agentic Intelligence

K. Team, Y . Bai, Y . Bao, Y . Charles, C. Chen, G. Chen, H. Chen, H. Chen, J. Chen, N. Chenet al., “Kimi k2: Open agentic intelligence,” arXiv preprint arXiv:2507.20534, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[10] [10]

Keep it simple: Unsupervised simplification of multi-paragraph text,

P. Laban, T. Schnabel, P. Bennett, and M. A. Hearst, “Keep it simple: Unsupervised simplification of multi-paragraph text,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021, pp. 6365–6378

work page 2021

[11] [11]

Large language diffusion models,

S. Nie, F. Zhu, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y . Lin, J.-R. Wen, and C. Li, “Large language diffusion models,”Advances in Neural Information Processing Systems, vol. 38, pp. 50 608–50 646, 2026

work page 2026

[12] [12]

BERTScore: Evaluating Text Generation with BERT

T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi, “Bertscore: Evaluating text generation with bert,”arXiv preprint arXiv:1904.09675, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1904

[13] [13]

Training Verifiers to Solve Math Word Problems

K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakanoet al., “Training verifiers to solve math word problems,”arXiv preprint arXiv:2110.14168, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021

[14] [14]

The effects of human variation in DUC summarization evaluation,

D. Harman and P. Over, “The effects of human variation in DUC summarization evaluation,” inProceedings of the ACL 2004 Workshop on Text Summarization Branches Out, 2004, pp. 10–17

work page 2004

[15] [15]

Sharegpt4v: Improving large multi-modal models with better captions,

L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 370– 387

work page 2024

[16] [16]

Unsu- pervised cross-lingual representation learning at scale,

A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzm´an, E. Grave, M. Ott, L. Zettlemoyer, and V . Stoyanov, “Unsu- pervised cross-lingual representation learning at scale,” inProceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 8440–8451

work page 2020

[17] [17]

Bleu: a method for automatic evaluation of machine translation,

K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” inProceedings of the 40th annual meeting of the Association for Computational Linguistics, 2002, pp. 311–318

work page 2002

[18] [18]

Rouge: A package for automatic evaluation of summaries,

C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” inText summarization branches out, 2004, pp. 74–81

work page 2004

[19] [19]

Compressing context to enhance inference efficiency of large language models,

Y . Li, B. Dong, F. Guerin, and C. Lin, “Compressing context to enhance inference efficiency of large language models,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 6342–6353

work page 2023