Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Rendering retrieved documents as images and compressing each according to query relevance gives RAG higher accuracy with one-eighth the tokens, and the compression ratios themselves rank passages.

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

2026-08-05 00:18 UTC pith:G5LAUIJF

load-bearing objection Interesting idea (query-aware optical compression with GRPO, ratio-as-reranker), but the reported numbers don't add up and the manuscript text is corrupted; it needs a full redo before it can be taken seriously. the 3 major comments →

arxiv 2608.00765 v1 pith:G5LAUIJF submitted 2026-08-01 cs.CL

RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation

Jiayang Yu , Jialun Zhong , Lei Zou This is my paper
classification cs.CL
keywords Retrieval-Augmented Generationoptical character renderingquery-aware compressionvisual tokensGRPOmultimodal question answeringrerankingmedical QA
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.

RAGOCR claims that the way to make retrieval-augmented generation cheaper and more accurate is to stop treating retrieved passages as text tokens and start treating them as images with query-dependent resolutions. A lightweight vision-language compressor, trained with Group Relative Policy Optimization (a reinforcement-learning method that compares a sampled group of outputs), looks at the question and all rendered documents together and assigns each page a compression factor; pages relevant to the answer keep high resolution, while peripheral pages are downscaled. On five medical QA benchmarks over the MedOmniKB corpus, the paper reports 61.47% average accuracy with about 2,284 RAG tokens per question, versus 44.63% accuracy and 17,914 tokens for naive RAG—a double-digit accuracy gain at roughly one-eighth the input. The same compression factors also act as relevance scores, letting the compressor double as a reranker that the paper says beats a dedicated 4B text reranker. If this holds, query-aware optical compression is not a sacrifice made to save tokens; it is a way to feed a model more of what matters within a fixed context.

Core claim

Central claim: rendering retrieved documents as images and shrinking each by a query-dependent factor compresses RAG context while improving accuracy. A small vision-language compressor, trained with GRPO, outputs one compression factor per document; a reward checks whether a frozen generator answers correctly from the compressed images, so the policy learns to keep resolution where the answer lives. On five medical QA benchmarks over MedOmniKB, the paper reports 61.47% average accuracy with 2,284 RAG tokens per case versus 44.63% with 17,914 tokens for naive RAG. The same factors rank documents: inverse compression beats a dedicated 4B reranker, and combining them improves further. On slide

What carries the argument

The central object is the query-aware optical compressor: a lightweight vision-language model that sees the question and every rendered document image in one forward pass and emits a compression factor per image. GRPO (Group Relative Policy Optimization) trains it with a reward that is positive only when the frozen generator gives the correct answer, plus a compression bonus that is sparse and active only on correct answers. Each factor rescales the image, reducing visual tokens by the factor; the inverse of the factor doubles as a relevance score, so the same model compresses and reranks.

Load-bearing premise

The load-bearing premise is that the compressor's resolution choices, learned from answer-correctness rewards with a sparse compression bonus, transfer across datasets and generators instead of overfitting to the pattern of correct answers in the training splits.

What would settle it

Use the trained compressor with a different fixed vision-language generator at inference time, no retraining, on the same five benchmarks; if the accuracy advantage over naive RAG disappears, the gains come from the generator's image-reading ability rather than from the query-aware allocation itself.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Retrieval depth stops being the main cost driver: a RAG system can take in many more documents than before because each one is cheap in visual tokens, and the compressor decides what deserves detail.
  • A compressor and a reranker can be merged into one model; the paper shows the compression ratios rank passages better than a pointwise reranker and combine additively with it.
  • Uniform compression hurts accuracy, so resolution allocation is the active ingredient: the paper reports 55.02% accuracy for static compression versus 61.47% for the query-aware version.
  • The approach extends past plain text to slide decks and other image-rich documents, improving exact match and F1 over full-resolution visual input.
  • Because rendering is deterministic and independent of the query, it can be precomputed, so the only per-query overhead is one lightweight compressor pass before the generator.

Where Pith is reading between the lines

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

  • If the listwise ranking signal generalizes, a single model could serve as both compressor and reranker for any text-first RAG pipeline, since the exact same forward pass provides both functions.
  • The same resolution-allocation policy could in principle be applied to documents that have no clean text tokenization—tables, charts, scans, or mixed-media pages—where text-level compression has no direct analogue.
  • A direct test of the reward design's stability would be to retrain the compressor on a different corpus or with a smaller generator and check whether the per-document ratios remain interpretable as relevance scores.

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 / 4 minor

Summary. The paper proposes RAGOCR, a framework that renders retrieved documents as images and trains a lightweight vision-language compressor via GRPO to assign query-dependent compression ratios, so that relevant passages are preserved at higher resolution and peripheral passages are aggressively downscaled. The compressed images are fed to a frozen VLM generator. The authors claim that RAGOCR surpasses naive RAG by over 15% in accuracy while using roughly one-eighth of the input tokens, that it consistently outperforms hard and soft compression baselines, that it generalizes to multimodal slide decks, and that the learned compression ratios can serve as an effective reranking signal. The submission includes experiments on five medical QA benchmarks using the MedOmniKB retrieval corpus and on SlideVQA.

Significance. If the reported results were reliable, the core idea would be valuable: using visual rendering as a high-density compression channel for RAG, conditioning the compression on the query, and obtaining a relevance ranking as a by-product. The paper does not release code or data, however, and the central quantitative claims are not internally consistent: the paper's own tables and text report mutually incompatible averages. Because the headline contributions rest entirely on these numbers, the significance of the work cannot be assessed from the submitted evidence. The conceptual contribution is interesting, but the evaluation as presented is not trustworthy.

major comments (3)
  1. [§4.3, Table 1, §4.5, Table 3, Table 4] The reported accuracy averages are internally inconsistent. Table 1 lists RAGOCR per-benchmark accuracies (71.36, 45.75, 69.21, 68.87, 58.00) whose arithmetic mean is 62.64, while the table prints 62.51 and the text repeatedly cites 61.47 (§4.3, and also in §4.6 for the w=800 configuration). In §4.5, 'R&R' is said to achieve 63.60%, but Table 3's row (73.7, 51.3, 67.6, 70.8, 61.6) averages 65.00; 'As Reranker' is said to achieve 59.15% but its row averages 59.06; 'compress' is said to achieve 60.10% but its row averages 58.74. Table 4's text values (59.85, 61.47, 62.74) also do not match the arithmetic means of the corresponding rows (60.24, 62.64, 64.08). Since these numbers are the basis for the 'over 15% over naive RAG' and 'consistent outperformance' claims, the evaluation cannot be considered reliable as submitted.
  2. [Abstract, §4.3, Table 1] The abstract and conclusion state that RAGOCR 'consistently outperforms' hard and soft compression baselines. Table 1 directly contradicts this: OSCAR-24B exceeds RAGOCR on BioASQ (80.31 vs 71.36) and MedMCQA (51.00 vs 45.75). The §4.3 discussion narrows the claim to 'methods that share the same generator backbone,' but that qualification is absent from the abstract, the introduction, and the conclusion. The paper should either remove the unqualified claim or present a comparison that is controlled for generator capacity and training setup.
  3. [§3.4, Eq. (13); §3.5; §4.5] The claim that the compressor's inverse compression ratios 'naturally' serve as relevance scores and outperform dedicated rerankers is not established by the training objective. In Eq. (13), the compression reward is nonzero only when r_task = 1, and it is an aggregate reward over all documents; there is no per-passage relevance supervision. The policy can increase the objective by preserving exactly the passages needed for a correct answer on the training questions, without learning a general notion of passage relevance. To support the emergent-reranker claim, the paper needs a direct analysis or ablation, for example correlation of inverse ratios with human relevance judgments, comparison against random compression-ratio assignment, or a per-passage supervised baseline. As written, the evidence is consistent with the compressor having overfit to answer patterns rather than learned a ge
minor comments (4)
  1. [§4.2] 'These two methods remain top-k=5 documents' should read 'retain the top-5 documents'.
  2. [Tables 3 and 4] The tables report 'weighted average' accuracy, but no weights are specified anywhere in the text. Without the weights, the reader cannot verify the reported averages, especially given the arithmetic inconsistencies noted above.
  3. [§4.4] The SlideVQA training details are incomplete: no training-set size, number of GRPO steps, or hyperparameters are given. The claim in Figure 3 that RAGOCR 'precisely preserved the resolution of this image almost unchanged' is not quantified.
  4. [Front matter] The manuscript's full text contains a long medical retrieval passage and the phrase 'LOST IN MIDDLELLM' before the title. If this text is not intended content, it should be removed before any resubmission.

Circularity Check

1 steps flagged

Compressor-as-reranker claim is partly by construction (reward directly optimizes answer correctness), but the main compression-efficiency result is self-contained.

specific steps
  1. fitted input called prediction [§3.4 Eq. (13) and §3.5 'Compressor as reranker'; evaluated in §4.5]
    "r_comp = (1/N Σ log c_i, if r_task=1, 0 otherwise). This design prevents the model from trivially maximizing compression at the expense of task performance. ... Since the compressor learns to assign lower compression ratios to more query-relevant passages, the inverse of the compression factor s_i = 1/c_i naturally serves as a relevance score."

    The only link between the compressor's per-image ratios and passage relevance is the training reward: r_task=1 iff the fixed generator answers correctly from the compressed images, and the compression bonus is nonzero only when r_task=1. Thus the policy is optimized, by construction, to give more visual budget to passages whose preservation makes the correct answer more likely. Presenting 1/c_i as an 'emergent' relevance signal that 'surprisingly' outperforms a reranker is therefore a fitted input called a prediction: the ranking is a direct product of the answer-correctness reward, not an independent discovery. The main compression accuracy/efficiency comparison is empirical and not circular, which is why the overall score is moderate.

full rationale

The core RAGOCR pipeline is self-contained: a GRPO-trained compressor is rewarded for preserving enough detail for a fixed generator to answer correctly, and the resulting accuracy/token trade-off is measured against external baselines on held-out benchmarks. That central claim is not circular. However, the paper's secondary but prominent claim that compression ratios 'inherently encode passage relevance' and can serve as a reranker is substantially by construction: Eq. (13) activates the compression reward only when r_task=1, i.e., only when the compressed images already enabled a correct answer, so the learned ratios are explicitly fitted to answer correctness. Calling this an 'emergent' or 'surprisingly effective' reranking signal overstates what the training objective already encodes. I did not find load-bearing self-citation, imported uniqueness theorems, or ansatz-smuggling via citation. Separately, the reported numbers contain internal inconsistencies (Table 1 row average 62.51 vs §4.3's 61.47; OSCAR-24B beats RAGOCR on BioASQ and MedMCQA, contradicting the abstract's 'consistently outperforms'), but those are correctness/consistency issues, not circularity, and I have not counted them in the circularity score.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical or conceptual entities. The load-bearing items are the hand-chosen rendering and reward hyperparameters, plus the domain assumptions that visual compression preserves answer-relevant information and that the fixed generator is a reliable judge during RL training.

free parameters (5)
  • Maximum canvas width w = 800 (or 1600 in ablation; Table 1 uses 800x918 per Section 4.6)
    Chosen by hand; Section 4.6 shows accuracy varies from 59.85% to 62.74% across w=400, 800, 1600. The paper does not specify which width is used for the main results.
  • Font size f = 12
    Fixed at 12 in the ablation and main experiments; affects line height and number of visual tokens.
  • Compression reward weight lambda = Not disclosed
    Introduced in Equation 11; the value is not reported, so the balance between task reward and compression reward is not reproducible.
  • Format reward weight alpha = Not disclosed
    Also from Equation 11; no value or ablation is given.
  • Sparse injection schedule for compression reward = Not specified
    Section 3.4 says compression reward is 'injected only at sparse intervals', but no schedule is defined, making the training procedure underspecified.
axioms (4)
  • domain assumption Visual tokens carry higher information density than text tokens.
    Section 1 relies on DeepSeek-OCR and Glyph results; the entire approach assumes rendering text to images preserves enough information for QA.
  • domain assumption The fixed generator Qwen3-VL-8B-Thinking can answer correctly from compressed images.
    Used throughout Section 4; the reward computation depends on this generator's output, and no analysis is given for its failure modes.
  • domain assumption The MedOmniKB retrieval corpus returns documents whose relevance is well captured by the learned compression ratios.
    The evaluation is built on this corpus; the paper does not characterize retrieval quality or its effect on the results.
  • standard math GRPO with the stated reward design is a valid and stable optimizer for the compressor.
    The paper uses GRPO as a standard RL algorithm; the specific reward shaping is not proven stable across datasets.

pith-pipeline@v1.3.0-alltime-deepseek · 26815 in / 9922 out tokens · 103703 ms · 2026-08-05T00:18:21.569683+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation." pith.science (2026). https://pith.science/paper/G5LAUIJF

@misc{pith2026260800765,
  author       = {Pith},
  title        = {Pith review of: RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G5LAUIJF}},
  note         = {Machine review of arXiv:2608.00765}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Retrieval-Augmented Generation (RAG) has become essential for knowledge-intensive question answering, yet scaling RAG pipelines remains challenging due to the prohibitive computational cost of processing lengthy retrieved contexts. Existing compression approaches face a fundamental trade-off: hard compression methods operate online in a query-aware fashion but achieve only modest compression rates and typically require fine-tuning the generative model, while soft compression methods attain higher ratios but rely on costly offline encoding that is entirely agnostic to the input query. To bridge this gap, we introduce RAGOCR, a novel framework that compresses retrieved documents into compact visual representations conditioned on the input query. To further balance compression rate and information fidelity, we introduce a query-aware dynamic resolution mechanism that adaptively allocates visual granularity based on each document's estimated relevance and complexity: highly relevant passages are rendered at higher resolution to preserve fine-grained details, while peripheral documents are aggressively compressed at lower resolution. Experiments on five QA benchmarks using the MedOmniKB retrieval corpus demonstrate that RAGOCR surpasses naive RAG by over 15\% in accuracy while requiring only one-eighth the number of input tokens, and consistently outperforms both hard and soft compression baselines across varying retrieval depths.

Figures

Figures reproduced from arXiv: 2608.00765 by Jialun Zhong, Jiayang Yu, Lei Zou.

Figure 1
Figure 1. Figure 1: Comparison of three RAG paradigms. (a) RAG con [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the RAGOCR framework, which operates effectively across knowledge sources of different modalities. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An example of data from SlideVQA, the key in [PITH_FULL_IMAGE:figures/full_fig_p007_3.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

48 extracted references · 16 canonical work pages · 2 internal anchors

  1. [1]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond.arXiv preprint arXiv:2308.12966(2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-VL Technical Rep...

    Pith/arXiv arXiv 2025

  3. [3]

    Zhe Chen, Yusheng Liao, Shuyang Jiang, Pingjie Wang, Yiqiu Guo, Yanfeng Wang, and Yu Wang. 2025. Towards omni-rag: Comprehensive retrieval-augmented generation for large language models in medical applications. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 15285–15309

    work page 2025

  4. [4]

    Jiale Cheng, Yusen Liu, Xinyu Zhang, Yulin Fei, Wenyi Hong, Ruiliang Lyu, Weihan Wang, Zhe Su, Xiaotao Gu, Xiao Liu, et al. 2025. Glyph: Scaling context windows via visual-text compression.arXiv preprint arXiv:2510.17800(2025)

    arXiv 2025

  5. [5]

    Xin Cheng, Xun Wang, Xingxing Zhang, Tao Ge, Si-Qing Chen, Furu Wei, Huishuai Zhang, and Dongyan Zhao. 2024. xrag: Extreme context compres- sion for retrieval-augmented generation with one token.Advances in Neural Information Processing Systems37 (2024), 109487–109516

    2024

  6. [6]

    Nadezhda Chirkova, Thibault Formal, Vassilina Nikoulina, and Stéphane Clin- chant. 2025. Provence: efficient and robust context pruning for retrieval- augmented generation.arXiv preprint arXiv:2501.16214(2025)

    Pith/arXiv arXiv 2025

  7. [7]

    Yufeng Du, Minyang Tian, Srikanth Ronanki, Subendhu Rongali, Sravan Bodapati, Aram Galstyan, Azton Wells, Roy Schwartz, Eliu A Huerta, and Hao Peng. 2025. Context length alone hurts LLM performance despite perfect retrieval.arXiv preprint arXiv:2510.05381(2025)

    arXiv 2025

  8. [8]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. InInternational conference on machine learning. PMLR, 3929–3938

    2020

  9. [9]

    Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2021. Aligning AI With Shared Human Values.Proceedings of the International Conference on Learning Representations (ICLR)(2021)

    work page 2021

  10. [10]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Under- standing.Proceedings of the International Conference on Learning Representations (ICLR)(2021)

    2021

  11. [11]

    Jerry Huang, Siddarth Madala, Risham Sidhu, Cheng Niu, Hao Peng, Julia Hocken- maier, and Tong Zhang. 2025. Rag-rl: Advancing retrieval-augmented generation via rl and curriculum learning.arXiv preprint arXiv:2503.12759(2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Taeho Hwang, Sukmin Cho, Soyeong Jeong, Hoyun Song, SeungYoon Han, and Jong C Park. 2025. Exit: Context-aware extractive compression for enhancing retrieval-augmented generation. InFindings of the Association for Computational Linguistics: ACL 2025. 4895–4924

    work page 2025

  13. [13]

    Huiqiang Jiang, Qianhui Wu, , Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Sriku...

    work page 2024

  14. [14]

    Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. LLM- Lingua: Compressing Prompts for Accelerated Inference of Large Language Mod- els. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 1...

    doi:10.18653/v1/2023.emnlp- 2023

  15. [15]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1658–1677

    2024

  16. [16]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516(2025)

    Pith/arXiv arXiv 2025

  17. [17]

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams.Applied Sciences11, 14 (2021), 6421

    2021

  18. [18]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu

  19. [19]

    Dongwon Jung, Qin Liu, Tenghao Huang, Ben Zhou, and Muhao Chen. 2024. Familiarity-aware evidence compression for retrieval-augmented generation. arXiv preprint arXiv:2409.12468(2024)

    work page arXiv 2024

  20. [20]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  21. [21]

    Mingxin Li, Yanzhao Zhang, Dingkun Long, Chen Keqin, Sibo Song, Shuai Bai, Zhibo Yang, Pengjun Xie, An Yang, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2026. Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Frame- work for State-of-the-Art Multimodal Retrieval and Ranking.arXiv preprint arXiv:2601.04720(2026)

    Pith/arXiv arXiv 2026

  22. [22]

    Abdi, Yuqing Yang, and Lili Qiu

    Yucheng Li, Surin Ahn, Huiqiang Jiang, Amir H. Abdi, Yuqing Yang, and Lili Qiu

    work page

  23. [23]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts.Transactions of the association for computational linguistics12 (2024), 157–173

    2024

  24. [24]

    Maxime Louis, Hervé Déjean, and Stéphane Clinchant. 2025. Pisco: Pretty simple compression for retrieval-augmented generation. InFindings of the Association for Computational Linguistics: ACL 2025. 15506–15521

    work page 2025

  25. [25]

    Maxime Louis, Thibault Formal, Hervé Dejean, and Stéphane Clinchant. 2025. Oscar: Online soft compression and reranking.arXiv preprint arXiv:2504.07109 (2025)

    arXiv 2025

  26. [26]

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. 2022. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. InConference on health, inference, and learning. PMLR, 248– 260

    2022

  27. [27]

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, et al. 2024. Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. In Findings of the Association for Computational Linguistics: ACL 2024. 963–981

    2024

  28. [28]

    Vicky Zhao, Lili Qiu, and Dongmei Zhang

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Ruhle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. 2024. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression. InFindings of the Association for Computational Linguistics ACL 2024, Lun-Wei Ku, An...

    work page 2024

  29. [29]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. Replug: Retrieval-augmented black-box language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers). 8371–8384

    work page 2024

  31. [31]

    Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question an- swering on multiple images. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 13636–13645

    2023

  32. [32]

    Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https: //arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  33. [33]

    George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioannis Partalas, Matthias Zschunke, Michael R Alvers, Dirk Weissenborn, Anastasia Krithara, Ser- gios Petridis, Dimitris Polychronopoulos, et al. 2015. An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition. BMC bioinformatics16, 1 (2015), 138

    work page 2015

  34. [34]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution.arXiv preprint arXiv:2409.12191(2024)

    Pith/arXiv arXiv 2024

  35. [35]

    Zhengren Wang, Jiayang Yu, Dongsheng Ma, Zhe Chen, Yu Wang, Zhiyu Li, Feiyu Xiong, Yanfeng Wang, Linpeng Tang, Wentao Zhang, et al. 2025. Rare: Retrieval-augmented reasoning modeling.arXiv preprint arXiv:2503.23513(2025)

    Pith/arXiv arXiv 2025

  36. [36]

    Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. Deepseek-ocr: Contexts optical compression.arXiv preprint arXiv:2510.18234(2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Haoran Wei, Yaofeng Sun, and Yukun Li. 2026. DeepSeek-OCR 2: Visual Causal Flow.arXiv preprint arXiv:2601.20552(2026)

    arXiv 2026

  38. [38]

    Ling Xing, Alex Jinpeng Wang, Rui Yan, Xiangbo Shu, and Jinhui Tang. 2025. Vision-centric token compression in large language model.arXiv preprint arXiv:2502.00791(2025)

    arXiv 2025

  39. [39]

    Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2023. Recomp: Improving retrieval- augmented lms with compression and selective augmentation.arXiv preprint arXiv:2310.04408(2023)

    Pith/arXiv arXiv 2023

  40. [40]

    RenYe Yan, Jikang Cheng, Yaozhong Gan, Shikun Sun, You Wu, Yunfan Yang, Liang Ling, Jinlong Lin, Yeshuang Zhu, Jie Zhou, et al. 2025. Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment. InProceedings of the IEEE/CVF International Conference on Computer Vision. 1924–1934

    work page 2025

  41. [41]

    Renye Yan, Jikang Cheng, Shikun Sun, Yi Sun, Wei Peng, Yaozhong Gan, You Wu, Ling Liang, Junliang Xing, and Yimao Cai. [n. d.]. MultiTune: Phase-Aware Multi-Objective Optimization for Diffusion Models. ([n. d.])

    work page

  42. [42]

    Renye Yan, Jikang Cheng, Shikun Sun, Yi Sun, You Wu, Wei Peng, Zongwei Wang, Ling Liang, Junliang Xing, and Yimao Cai. 2026. Do Less, Achieve More: Do We Need Every-Step Optimization for RL Fine-tuning of Diffusion Models?arXiv preprint arXiv:2605.15855(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [43]

    Renye Yan, Jikang Cheng, You Wu, Ling Liang, Wei Peng, Athanasios V Vasilakos, Qingyu Zhao, Yu Zhang, Ehsan Adeli, Kilian M Pohl, et al . 2026. Pixel-Space Diffusion Transformers.arXiv preprint arXiv:2607.17585(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Le Zhang, Bo Wang, Xipeng Qiu, Siva Reddy, and Aishwarya Agrawal. 2025. Rearank: Reasoning re-ranking agent via reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2458–2471

    work page 2025

  45. [45]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou

  46. [48]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

    Pith/arXiv arXiv 2025

  47. [2019]

    In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)

    PubMedQA: A Dataset for Biomedical Research Question Answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2567–2577

    2019

  48. [2025]

    InSecond Conference on Language Modeling

    SecurityLingua: Efficient Defense of LLM Jailbreak Attacks via Security- Aware Prompt Compression. InSecond Conference on Language Modeling. https: //openreview.net/forum?id=tybbSo6wba

    work page