pith. machine review for the scientific record. sign in

arxiv: 2604.19386 · v2 · submitted 2026-04-21 · 💻 cs.CV

Recognition: unknown

Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval

Zhiheng Fu , Yupeng Hu , Qianyun Yang , Shiqi Zhang , Zhiwei Chen , Zixu Li
Authors on Pith no claims yet

Pith reviewed 2026-05-10 02:04 UTC · model grok-4.3

classification 💻 cs.CV
keywords composed image retrievalnoisy triplet correspondencerobust learningmultimodal large language modelsknowledge internalizationsemantic ambiguityimage retrieval
0
0 comments X

The pith

Air-Know uses an offline MLLM expert to build a clean anchor set, internalizes its logic in a lightweight proxy, and diverts data into separate streams to break the vicious cycle of noise identification in composed image retrieval.

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

The paper seeks to fix the Noisy Triplet Correspondence problem that arises in Composed Image Retrieval when queries combine an image and text but contain ambiguities such as partial matches. Standard robust learning methods that assume small-loss samples are clean fail here because the ambiguity makes noise detection unreliable, trapping the model in a self-reinforcing loop where the learner and its noise detector pollute each other's representations. Air-Know introduces an Expert-Proxy-Diversion approach: a strong Multimodal Large Language Model first creates a high-precision anchor dataset offline, a lightweight proxy then learns to replicate the expert's decisions, and a dual-stream process finally separates training into one clean alignment path and one feedback path that uses the proxy's confidence scores. A reader would care because successful decoupling would let retrieval systems train reliably on real-world multimodal queries that are often imperfect, without the performance collapse seen in current methods.

Core claim

Air-Know proposes the Expert-Proxy-Diversion decoupling paradigm in which External Prior Arbitration employs Multimodal Large Language Models as an offline expert to construct a high-precision anchor dataset, Expert Knowledge Internalization efficiently transfers the expert's discriminative logic to a lightweight proxy arbiter, and Dual Stream Reconciliation uses the proxy's matching confidence to divert training data into a clean alignment stream and a representation feedback reconciliation stream, thereby preventing the self-dependent vicious cycle and representation pollution that plague existing robust methods under noisy triplet conditions.

What carries the argument

The Expert-Proxy-Diversion decoupling paradigm, which separates offline expert arbitration from online proxy training and data diversion to prevent interdependence between the learner and the noise arbiter.

If this is right

  • Air-Know significantly outperforms existing state-of-the-art robust methods under the Noisy Triplet Correspondence setting on multiple Composed Image Retrieval benchmark datasets.
  • The network remains competitive with strong methods on traditional clean Composed Image Retrieval tasks without noise.
  • The decoupling prevents catastrophic representation pollution by removing the direct dependence between the training learner and the noise identification process.
  • The approach handles semantic ambiguities such as partial matching without relying on the small-loss hypothesis that fails in this domain.

Where Pith is reading between the lines

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

  • The knowledge-internalization step suggests that large multimodal models can serve as one-time teachers whose logic is distilled into smaller, faster models for ongoing use in retrieval pipelines.
  • If the anchor-construction step generalizes beyond image-text pairs, the same expert-proxy structure could apply to other noisy multimodal correspondence tasks such as video-text or audio-text retrieval.
  • Improvements in future Multimodal Large Language Models would raise anchor quality and therefore lift the entire training process without requiring changes to the main retrieval network.

Load-bearing premise

Multimodal Large Language Models can reliably construct a high-precision anchor dataset that correctly identifies reliable triplets even when composed queries contain semantic ambiguities such as partial matching.

What would settle it

Manually auditing a sample of the MLLM-generated anchor dataset for mislabeled triplets on queries with partial matches, or measuring whether swapping the MLLM expert for a weaker model causes Air-Know's performance to fall back to the level of prior robust baselines on NTC benchmarks.

Figures

Figures reproduced from arXiv: 2604.19386 by Qianyun Yang, Shiqi Zhang, Yupeng Hu, Zhiheng Fu, Zhiwei Chen, Zixu Li.

Figure 1
Figure 1. Figure 1: (a) illustrates the semantic ambiguity of noise in NTC. (b) illustrates the vicious cycle of self-dependency caused by unreliable [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed Air-Know consists of three primary modules: (a) External Prior Arbitration leverages an offline multimodal expert [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity to the hyperparameters (a) p and (b) λ. 4.4. Sensitivity Analysis To evaluate the sensitivity of Air-Know to key hyperparam￾eters, we specifically analyzed two core parameters, the MC Dropout rate p and the feedback reconciliation stream loss weight λ, on the FashionIQ and CIRR datasets. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case Study on (a) CIRR and (b) FashionIQ. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of the margin α. We evaluated the impact of it in Equation (13), which serves as a parameter to con￾trol the threshold for penalizing noisy correspondence in the feed￾back reconciliation stream. A lower α imposes stricter filtering on semantically similar samples, while a higher α allows more samples exhibiting uncertainty to pass through, thereby creating distinct trade-offs between n… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of prompt design and ablation study on a real-world NTC case. We present a comparison of the reasoning process [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of NTC recognition results by the EKI module. We present the discrimination results of the EKI module for [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The complete three stage cross-validation prompt architecture. This design enforces a Deconstruct-Reason-Determine process. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The prompt variant in which Step 1 (input deconstruction) is removed. In this setting, we eliminated the instruction requiring [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The prompt variant in which Step 2 (Comparison and Reasoning) is removed. While this design retains the preliminary [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The end-to-end prompt variant in which both Step 1 and Step 2 are removed. This variant strips away all structured intermediate [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison on CIRR. We visualize the retrieval results to demonstrate the performance of the model under different [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison on FashionIQ. We compare Air-Know with TME to demonstrate the retrieval results in scenarios [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
read the original abstract

Composed Image Retrieval (CIR) has attracted significant attention due to its flexible multimodal query method, yet its development is severely constrained by the Noisy Triplet Correspondence (NTC) problem. Most existing robust learning methods rely on the "small loss hypothesis", but the unique semantic ambiguity in NTC, such as "partial matching", invalidates this assumption, leading to unreliable noise identification. This entraps the model in a self dependent vicious cycle where the learner is intertwined with the arbiter, ultimately causing catastrophic "representation pollution". To address this critical challenge, we propose a novel "Expert-Proxy-Diversion" decoupling paradigm, named Air-Know (ArbIteR calibrated Knowledge iNternalizing rObust netWork). Air-Know incorporates three core modules: (1) External Prior Arbitration (EPA), which utilizes Multimodal Large Language Models (MLLMs) as an offline expert to construct a high precision anchor dataset; (2) Expert Knowledge Internalization (EKI), which efficiently guides a lightweight proxy "arbiter" to internalize the expert's discriminative logic; (3) Dual Stream Reconciliation (DSR), which leverages the EKI's matching confidence to divert the training data, achieving a clean alignment stream and a representation feedback reconciliation stream. Extensive experiments on multiple CIR benchmark datasets demonstrate that Air-Know significantly outperforms existing SOTA methods under the NTC setting, while also showing strong competitiveness in traditional CIR.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Air-Know, a robust network for Composed Image Retrieval addressing the Noisy Triplet Correspondence (NTC) problem via an 'Expert-Proxy-Diversion' decoupling paradigm. It consists of three modules: External Prior Arbitration (EPA) using offline Multimodal Large Language Models (MLLMs) to generate a high-precision anchor dataset, Expert Knowledge Internalization (EKI) to distill the expert logic into a lightweight proxy arbiter, and Dual Stream Reconciliation (DSR) to separate training data into a clean alignment stream and a representation feedback stream. The authors claim that this approach significantly outperforms existing SOTA methods under the NTC setting while remaining competitive in traditional CIR on multiple benchmark datasets.

Significance. If the experimental claims hold, the work provides a meaningful advance in robust multimodal retrieval by explicitly decoupling the arbiter from the learner to avoid self-reinforcing noise cycles, a common failure mode when semantic ambiguities invalidate small-loss assumptions. The use of external offline experts for anchor construction and subsequent internalization offers a reusable template for other noisy supervision settings in vision-language tasks.

major comments (2)
  1. [Abstract and §3.1] Abstract and §3.1 (EPA module): The central claim of SOTA outperformance under NTC depends on the EPA module reliably producing high-precision anchors despite the very semantic ambiguities (e.g., partial matching) that the abstract states invalidate the small-loss hypothesis. No quantitative validation, error analysis, or prompting details are supplied showing MLLM robustness on NTC-specific partial-match cases; if MLLM judgments contain systematic errors here, the logic internalized by EKI and the diversion performed by DSR will propagate those errors, recreating representation pollution.
  2. [§4 Experiments] §4 Experiments: The abstract asserts 'extensive experiments' demonstrate significant outperformance, yet supplies no concrete metrics, baselines, NTC construction protocol, ablation results on the three modules, or statistical tests. Without these, the strength of evidence for the load-bearing claim cannot be assessed.
minor comments (2)
  1. [Title and Abstract] The title and abstract acronym ('ArbIteR calibrated Knowledge iNternalizing rObust netWork') is inventive but the capitalization pattern is non-standard and may confuse readers; a conventional expansion would improve clarity.
  2. [Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., recall@K improvement on a named dataset) to support the performance claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below with clarifications and commit to revisions that strengthen the evidence and transparency of the work.

read point-by-point responses

  1. Referee: [Abstract and §3.1] Abstract and §3.1 (EPA module): The central claim of SOTA outperformance under NTC depends on the EPA module reliably producing high-precision anchors despite the very semantic ambiguities (e.g., partial matching) that the abstract states invalidate the small-loss hypothesis. No quantitative validation, error analysis, or prompting details are supplied showing MLLM robustness on NTC-specific partial-match cases; if MLLM judgments contain systematic errors here, the logic internalized by EKI and the diversion performed by DSR will propagate those errors, recreating representation pollution.

    Authors: We agree that explicit validation of the EPA module's reliability on partial-match cases is essential to support the decoupling claims. In the revised manuscript we will expand §3.1 with a new subsection containing: (i) quantitative precision/recall metrics on a held-out set of manually annotated NTC triplets focused on partial matching, (ii) a categorized error analysis of MLLM judgments with representative failure cases, and (iii) the exact prompting templates and temperature settings used for the offline MLLM. These additions will demonstrate that the external expert maintains high fidelity on the targeted ambiguities, thereby justifying safe knowledge internalization by EKI and clean diversion by DSR. revision: yes

  2. Referee: [§4 Experiments] §4 Experiments: The abstract asserts 'extensive experiments' demonstrate significant outperformance, yet supplies no concrete metrics, baselines, NTC construction protocol, ablation results on the three modules, or statistical tests. Without these, the strength of evidence for the load-bearing claim cannot be assessed.

    Authors: We acknowledge that the experimental section requires greater explicitness. Although the manuscript already reports results across multiple benchmarks, the revision will expand §4 to include: the full NTC construction protocol (including how partial-match noise is synthetically introduced while preserving semantic structure), complete tables listing all baselines with numerical metrics, module-wise ablation studies quantifying the contribution of EPA, EKI, and DSR, and statistical significance tests (paired t-tests with p-values). These additions will be placed in the main text and supplementary material to allow full assessment of the claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; external MLLM decouples the process

full rationale

The paper's core derivation introduces an external offline MLLM in the EPA module to generate the anchor dataset before any proxy training or diversion occurs. This explicitly breaks the self-dependent vicious cycle described in the abstract. No load-bearing step reduces by construction to its own inputs: there are no self-definitional relations, no fitted parameters renamed as predictions, and no uniqueness theorems or ansatzes imported via self-citation. The outperformance claims rest on benchmark experiments that are independent of the internal logic. The method is self-contained against external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 4 invented entities

The central claim rests on the unverified reliability of MLLM judgments for creating anchor data and the proxy's ability to internalize discriminative logic; no free parameters or invented physical entities are mentioned.

axioms (1)
  • domain assumption Multimodal Large Language Models can serve as an offline expert to construct a high-precision anchor dataset for composed image queries despite semantic ambiguities
    Invoked in the External Prior Arbitration module description
invented entities (4)
  • Expert-Proxy-Diversion decoupling paradigm no independent evidence
    purpose: To break the self-dependent vicious cycle between learner and arbiter in noisy triplet training
    Core proposed framework named in the abstract
  • External Prior Arbitration (EPA) module no independent evidence
    purpose: Utilize MLLMs to build high-precision anchor dataset
    One of the three core modules
  • Expert Knowledge Internalization (EKI) module no independent evidence
    purpose: Guide lightweight proxy arbiter to internalize expert logic
    One of the three core modules
  • Dual Stream Reconciliation (DSR) module no independent evidence
    purpose: Use EKI confidence to divert training into clean alignment and representation feedback streams
    One of the three core modules

pith-pipeline@v0.9.0 · 5579 in / 1563 out tokens · 58919 ms · 2026-05-10T02:04:35.914339+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 4 Pith papers

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

  1. ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval

    cs.CV 2026-04 unverdicted novelty 7.0

    ConeSep tackles noisy triplet correspondences in composed image retrieval by introducing geometric fidelity quantization to locate noise, negative boundary learning for semantic opposites, and targeted unlearning via ...

  2. OmniTrend: Content-Context Modeling for Scalable Social Popularity Prediction

    cs.CV 2026-04 unverdicted novelty 6.0

    OmniTrend predicts popularity by combining separate content attractiveness and contextual exposure predictors using cross-modal and exogenous signals.

  3. HotComment: A Benchmark for Evaluating Popularity of Online Comments

    cs.AI 2026-04 unverdicted novelty 6.0

    HotComment is a new multimodal benchmark that quantifies online comment popularity via content quality assessment, interaction-based prediction, and agent-simulated user engagement, accompanied by the StyleCmt stylist...

  4. CurEvo: Curriculum-Guided Self-Evolution for Video Understanding

    cs.CV 2026-04 unverdicted novelty 4.0

    CurEvo integrates curriculum guidance into self-evolution to structure autonomous improvement of video understanding models, yielding gains on VideoQA benchmarks.

Reference graph

Works this paper leans on

136 extracted references · 38 canonical work pages · cited by 4 Pith papers · 12 internal anchors

  1. [1]

    Learning with noisy triplet corre- spondence for composed image retrieval

    Shuxian Li, Changhao He, Xiting Liu, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Learning with noisy triplet corre- spondence for composed image retrieval. InCVPR, pages 19628–19637, 2025. 2, 3, 6, 7

    2025

  2. [2]

    Habit: Chrono- synergia robust progressive learning framework for com- posed image retrieval

    Zixu Li, Yupeng Hu, Zhiwei Chen, Shiqi Zhang, Qinlei Huang, Zhiheng Fu, and Yinwei Wei. Habit: Chrono- synergia robust progressive learning framework for com- posed image retrieval. InAAAI, pages 6762–6770, 2026. 6, 7

    2026

  3. [3]

    arXiv preprint arXiv:2603.29291 (2026)

    Guozhi Qiu, Zhiwei Chen, Zixu Li, Qinlei Huang, Zhiheng Fu, Xuemeng Song, and Yupeng Hu. Melt: Improve com- posed image retrieval via the modification frequentation- rarity balance network.arXiv preprint arXiv:2603.29291,

    work page arXiv

  4. [4]

    Intent: Invariance and discrimination-aware noise mitigation for robust com- posed image retrieval

    Zhiwei Chen, Yupeng Hu, Zhiheng Fu, Zixu Li, Jiale Huang, Qinlei Huang, and Yinwei Wei. Intent: Invariance and discrimination-aware noise mitigation for robust com- posed image retrieval. InAAAI, pages 20463–20471, 2026. 2, 6, 7

    2026

  5. [5]

    Chat-driven text generation and interaction for person retrieval

    Zequn Xie, Chuxin Wang, Yeqiang Wang, Sihang Cai, Shulei Wang, and Tao Jin. Chat-driven text generation and interaction for person retrieval. InEMNLP, pages 5259– 5270, 2025

    2025

  6. [6]

    Sentence-level prompts benefit composed image retrieval

    Xinxing Xu, Yong Liu, Salman Khan, Fahad Khan, Wang- meng Zuo, Rick Siow Mong Goh, Chun-Mei Feng, et al. Sentence-level prompts benefit composed image retrieval. InICLR, 2024. 6, 7

    2024

  7. [7]

    arXiv preprint arXiv:2604.01617 (2026)

    Qianyun Yang, Zhiwei Chen, Yupeng Hu, Zixu Li, Zhi- heng Fu, and Liqiang Nie. Stable: Efficient hybrid nearest neighbor search via magnitude-uniformity and cardinality- robustness.arXiv preprint arXiv:2604.01617, 2026

    work page arXiv 2026

  8. [8]

    Queries are not alone: Clustering text embeddings for video search

    Peiyang Liu, Xi Wang, Ziqiang Cui, and Wei Ye. Queries are not alone: Clustering text embeddings for video search. InACM SIGIR, pages 874–883, 2025

    2025

  9. [9]

    Qure: Query-relevant re- trieval through hard negative sampling in composed image retrieval

    Jaehyun Kwak, Ramahdani Muhammad Izaaz Inhar, Se- Young Yun, and Sung-Ju Lee. Qure: Query-relevant re- trieval through hard negative sampling in composed image retrieval. InICML, pages 32063–32077. PMLR, 2025. 6, 7

    2025

  10. [10]

    Retrieval-based unsupervised noisy label detection on text data

    Peiyang Liu, Jinyu Yang, Lin Wang, Sen Wang, Yunlai Hao, and Huihui Bai. Retrieval-based unsupervised noisy label detection on text data. InACM CIKM, pages 4099–4104, 2023

    2023

  11. [11]

    Conquer: Context-aware representation with query enhancement for text-based person search,

    Zequn Xie. Conquer: Context-aware representation with query enhancement for text-based person search.arXiv preprint arXiv:2601.18625, 2026

    work page arXiv 2026

  12. [12]

    Core-mmrag: Cross-source knowledge recon- ciliation for multimodal rag

    Yang Tian, Fan Liu, Jingyuan Zhang, Yupeng Hu, Liqiang Nie, et al. Core-mmrag: Cross-source knowledge recon- ciliation for multimodal rag. InACL, pages 32967–32982, 2025

    2025

  13. [13]

    Hdnet: A hybrid domain network with multi-scale high-frequency information en- hancement for infrared small target detection.IEEE TGRS, 2025

    Mingzhu Xu, Chenglong Yu, Zexuan Li, Haoyu Tang, Yu- peng Hu, and Liqiang Nie. Hdnet: A hybrid domain network with multi-scale high-frequency information en- hancement for infrared small target detection.IEEE TGRS, 2025

    2025

  14. [14]

    Transforming time and space: efficient video super- resolution with hybrid attention and deformable transform- ers.The Visual Computer, pages 1–12, 2025

    Linling Jiang, Xin Wang, Fan Zhang, and Caiming Zhang. Transforming time and space: efficient video super- resolution with hybrid attention and deformable transform- ers.The Visual Computer, pages 1–12, 2025

    2025

  15. [15]

    Llava steering: Vi- sual instruction tuning with 500x fewer parameters through modality linear representation-steering

    Jinhe Bi, Yujun Wang, Haokun Chen, Xun Xiao, Artur Hecker, V olker Tresp, and Yunpu Ma. Llava steering: Vi- sual instruction tuning with 500x fewer parameters through modality linear representation-steering. InACL, pages 15230–15250, 2025

    2025

  16. [16]

    Reassessing layer pruning in LLMs: New insights and methods

    Yao Lu, Hao Cheng, Yujie Fang, Zeyu Wang, Jiaheng Wei, Dongwei Xu, Qi Xuan, and Zhaowei Zhu. Reassessing layer pruning in LLMs: New insights and methods. In ICLR, 2026

    2026

  17. [17]

    Prior knowledge in- tegration via llm encoding and pseudo event regulation for video moment retrieval

    Yiyang Jiang, Wengyu Zhang, Xulu Zhang, Xiao-Yong Wei, Chang Wen Chen, and Qing Li. Prior knowledge in- tegration via llm encoding and pseudo event regulation for video moment retrieval. InACM MM, pages 7249–7258,

  18. [18]

    Optimizing instruc- tion synthesis: Effective exploration of evolutionary space with tree search.arXiv preprint arXiv:2410.10392, 2024

    Chenglin Li, Qianglong Chen, Zhi Li, Feng Tao, Yicheng Li, Hao Chen, Fei Yu, and Yin Zhang. Optimizing instruc- tion synthesis: Effective exploration of evolutionary space with tree search.arXiv preprint arXiv:2410.10392, 2024. 2

    work page arXiv 2024

  19. [19]

    Scientific image synthesis: Benchmark- ing, methodologies, and downstream utility.arXiv preprint arXiv:2601.17027, 2026

    Honglin Lin, Chonghan Qin, Zheng Liu, Qizhi Pei, Yu Li, Zhanping Zhong, Xin Gao, Yanfeng Wang, Conghui He, and Lijun Wu. Scientific image synthesis: Benchmark- ing, methodologies, and downstream utility.arXiv preprint arXiv:2601.17027, 2026

    work page arXiv 2026

  20. [20]

    InACL Findings, pages 8950–8970, 2025

    Yunyao Zhang, Zikai Song, Hang Zhou, Wenfeng Ren, Yi- Ping Phoebe Chen, Junqing Yu, and Wei Yang.ga− s3: Comprehensive social network simulation with group agents. InACL Findings, pages 8950–8970, 2025

    2025

  21. [21]

    Yuanjun Zhang, Fuzel Ahamed Shaik, Suvojit Acharjee, Fahad Khalid, and Mourad Oussalah. Towards reliable mul- timodal disaster severity assessment through preference op- timization and explainable vision-language reasoning.Re- liability Engineering & System Safety, page 112674, 2026

    2026

  22. [22]

    Cotextor: Training-free modular multi- lingual text editing via layered disentanglement and depth- aware fusion

    Zhenyu Yu, MOHD Y AMANI IDNA IDRIS, Pei Wang, and Rizwan Qureshi. Cotextor: Training-free modular multi- lingual text editing via layered disentanglement and depth- aware fusion. InNeurIPS, 2025

    2025

  23. [23]

    Delving deeper: Hierarchi- cal visual perception for robust video-text retrieval,

    Zequn Xie, Boyun Zhang, Yuxiao Lin, and Tao Jin. Delving deeper: Hierarchical visual perception for robust video-text retrieval.arXiv preprint arXiv:2601.12768, 2026

    work page arXiv 2026

  24. [24]

    Topological federated clustering via gravitational po- tential fields under local differential privacy.AAAI, 40(28): 24044–24051, 2026

    Yunbo Long, Jiaquan Zhang, Xi Chen, and Alexandra Brin- trup. Topological federated clustering via gravitational po- tential fields under local differential privacy.AAAI, 40(28): 24044–24051, 2026

    2026

  25. [25]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others

    Jiaye Lin, Yifu Guo, Yuzhen Han, Sen Hu, Ziyi Ni, Licheng Wang, Mingguang Chen, Hongzhang Liu, Ronghao Chen, Yangfan He, et al. Se-agent: Self-evolution trajectory op- timization in multi-step reasoning with llm-based agents. arXiv preprint arXiv:2508.02085, 2025

    work page arXiv 2025

  26. [26]

    Coupled mamba: Enhanced multimodal fusion with coupled state space model.NeurIPS, 37:59808–59832, 2024

    Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model.NeurIPS, 37:59808–59832, 2024

    2024

  27. [27]

    FLARE: Fully Integration of Vision-Language Representations for Deep Cross-Modal Understanding

    Zheng Liu, Mengjie Liu, Jingzhou Chen, Jingwei Xu, Bin Cui, Conghui He, and Wentao Zhang. Fusion: Fully inte- gration of vision-language representations for deep cross- modal understanding.arXiv preprint arXiv:2504.09925, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Autogenic language embedding for coherent point tracking

    Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Autogenic language embedding for coherent point tracking. InACM MM, pages 2021–2030, 2024

    2021

  29. [29]

    Erase: Bypassing collaborative detection of ai counterfeit via comprehensive artifacts elimination.IEEE TDSC, pages 1–18, 2026

    Qianyun Yang, Peizhuo Lv, Yingjiu Li, Shengzhi Zhang, Yuxuan Chen, Zhiwei Chen, Zixu Li, and Yupeng Hu. Erase: Bypassing collaborative detection of ai counterfeit via comprehensive artifacts elimination.IEEE TDSC, pages 1–18, 2026

    2026

  30. [30]

    Open multimodal retrieval- augmented factual image generation.arXiv preprint arXiv:2510.22521, 2025

    Yang Tian, Fan Liu, Jingyuan Zhang, Wei Bi, Yu- peng Hu, and Liqiang Nie. Open multimodal retrieval- augmented factual image generation.arXiv preprint arXiv:2510.22521, 2025

    work page arXiv 2025

  31. [31]

    Prism: Self-pruning intrinsic selection method for training- free multimodal data selection, 2025

    Jinhe Bi, Yifan Wang, Danqi Yan, Aniri, Wenke Huang, Zengjie Jin, Xiaowen Ma, Artur Hecker, Mang Ye, Xun Xiao, Hinrich Schuetze, V olker Tresp, and Yunpu Ma. Prism: Self-pruning intrinsic selection method for training- free multimodal data selection, 2025. 2

    2025

  32. [32]

    Transformer tracking with cyclic shifting window attention

    Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. InCVPR, pages 8791–8800, 2022. 2

    2022

  33. [33]

    Yielding unblemished aesthetics through a unified network for visual imperfec- tions removal in generated images.AAAI, 39(9):9716– 9724, 2025

    Zhenyu Yu and Chee Seng Chan. Yielding unblemished aesthetics through a unified network for visual imperfec- tions removal in generated images.AAAI, 39(9):9716– 9724, 2025

    2025

  34. [34]

    Visual instance-aware prompt tuning

    Xi Xiao, Yunbei Zhang, Xingjian Li, Tianyang Wang, Xiao Wang, Yuxiang Wei, Jihun Hamm, and Min Xu. Visual instance-aware prompt tuning. InACM MM, pages 2880–

  35. [35]

    Association for Computing Machinery, Inc, 2025

    2025

  36. [36]

    Prompt-based adaptation in large-scale vision models: A survey.arXiv preprint arXiv:2510.13219, 2025

    Xi Xiao, Yunbei Zhang, Lin Zhao, Yiyang Liu, Xiaoying Liao, Zheda Mai, Xingjian Li, Xiao Wang, Hao Xu, Jihun Hamm, et al. Prompt-based adaptation in large-scale vision models: A survey.arXiv preprint arXiv:2510.13219, 2025

    work page arXiv 2025

  37. [37]

    Videocogqa: A controllable benchmark for evalu- ating cognitive abilities in video-language models.arXiv preprint arXiv:2411.09105, 2024

    Chenglin Li, Qianglong Chen, Zhi Li, Feng Tao, and Yin Zhang. Videocogqa: A controllable benchmark for evalu- ating cognitive abilities in video-language models.arXiv preprint arXiv:2411.09105, 2024

    work page arXiv 2024

  38. [38]

    Compact transformer tracker with correla- tive masked modeling

    Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correla- tive masked modeling. InAAAI, pages 2321–2329, 2023

    2023

  39. [39]

    Collaborative multi-agent scripts generation for enhancing imperfect-information reasoning in murder mystery games, 2026

    Keyang Zhong, Junlin Xie, Hefeng Wu, Haofeng Li, and Guanbin Li. Collaborative multi-agent scripts generation for enhancing imperfect-information reasoning in murder mystery games, 2026

    2026

  40. [40]

    Hvd: Human vision- driven video representation learning for text-video retrieval,

    Zequn Xie, Xin Liu, Boyun Zhang, Yuxiao Lin, Sihang Cai, and Tao Jin. Hvd: Human vision-driven video rep- resentation learning for text-video retrieval.arXiv preprint arXiv:2601.16155, 2026

    work page arXiv 2026

  41. [41]

    Self-paced learning for images of antinuclear antibodies.IEEE TMI, 2025

    Yiyang Jiang, Guangwu Qian, Jiaxin Wu, Qi Huang, Qing Li, Yongkang Wu, and Xiao-Yong Wei. Self-paced learning for images of antinuclear antibodies.IEEE TMI, 2025

    2025

  42. [42]

    Tempo- ral coherent object flow for multi-object tracking

    Zikai Song, Run Luo, Lintao Ma, Ying Tang, Yi- Ping Phoebe Chen, Junqing Yu, and Wei Yang. Tempo- ral coherent object flow for multi-object tracking. InAAAI, pages 6978–6986, 2025

    2025

  43. [43]

    Ascd: Attention-steerable contrastive decoding for reducing hallu- cination in mllm.arXiv preprint arXiv:2506.14766, 2025

    Yujun Wang, Jinhe Bi, Yunpu Ma, and Soeren Pirk. Ascd: Attention-steerable contrastive decoding for reducing hallu- cination in mllm.arXiv preprint arXiv:2506.14766, 2025

    work page arXiv 2025

  44. [44]

    Cot-kinetics: A theoretical modeling assessing lrm reasoning process

    Jinhe Bi, Danqi Yan, Yifan Wang, Wenke Huang, Haokun Chen, Guancheng Wan, Mang Ye, Xun Xiao, Hinrich Schuetze, V olker Tresp, et al. Cot-kinetics: A theoretical modeling assessing lrm reasoning process.arXiv preprint arXiv:2505.13408, 2025

    work page arXiv 2025

  45. [45]

    Hierarchical consensus hashing for cross-modal re- trieval.IEEE TMM, 26:824–836, 2023

    Yuan Sun, Zhenwen Ren, Peng Hu, Dezhong Peng, and Xu Wang. Hierarchical consensus hashing for cross-modal re- trieval.IEEE TMM, 26:824–836, 2023

    2023

  46. [46]

    Dinov3-powered multi-task foundation model for quantitative remote sensing estimation.AAAI, 40(48):41455–41456, 2026

    Zhenyu Yu, Mohd Yamani Idna Idris, Pei Wang, and Rizwan Qureshi. Dinov3-powered multi-task foundation model for quantitative remote sensing estimation.AAAI, 40(48):41455–41456, 2026

    2026

  47. [47]

    Multi-modal gradual domain osmosis: Stepwise dynamic learning with batch matching for gradual domain adaptation

    Zixi Wang, Yubo Huang, Jingzehua Xu, Jinzhu Wei, Shuai Zhang, and Xin Lai. Multi-modal gradual domain osmosis: Stepwise dynamic learning with batch matching for gradual domain adaptation. InACM MM, page 8959–8967, New York, NY , USA, 2025. Association for Computing Machin- ery. 2

    2025

  48. [48]

    Recon: Enhancing true corre- spondence discrimination through relation consistency for robust noisy correspondence learning

    Quanxing Zha, Xin Liu, Shu-Juan Peng, Yiu-ming Cheung, Xing Xu, and Nannan Wang. Recon: Enhancing true corre- spondence discrimination through relation consistency for robust noisy correspondence learning. InCVPR, pages 29680–29689, 2025. 2

    2025

  49. [49]

    Robust multi-view clustering with noisy correspondence.IEEE Transactions on Knowledge and Data Engineering, 36(12):9150–9162, 2024

    Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence.IEEE Transactions on Knowledge and Data Engineering, 36(12):9150–9162, 2024

    2024

  50. [50]

    Cross-view graph matching guided anchor alignment for incomplete multi-view clustering.Informa- tion Fusion, 100:101941, 2023

    Xingfeng Li, Yinghui Sun, Quansen Sun, Zhenwen Ren, and Yuan Sun. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering.Informa- tion Fusion, 100:101941, 2023

    2023

  51. [51]

    Cross-modal active complementary learning with self-refining correspondence.Advances in neural information processing systems, 36:24829–24840, 2023

    Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence.Advances in neural information processing systems, 36:24829–24840, 2023

    2023

  52. [52]

    Prototype match- ing learning for incomplete multi-view clustering.IEEE Transactions on Image Processing, 34:828–841, 2025

    Honglin Yuan, Yuan Sun, Fei Zhou, Jing Wen, Shihua Yuan, Xiaojian You, and Zhenwen Ren. Prototype match- ing learning for incomplete multi-view clustering.IEEE Transactions on Image Processing, 34:828–841, 2025

    2025

  53. [53]

    Incom- plete multi-view clustering with paired and balanced dy- namic anchor learning.IEEE Transactions on Multimedia, 27:1486–1497, 2024

    Xingfeng Li, Yuangang Pan, Yuan Sun, Quansen Sun, Yinghui Sun, Ivor W Tsang, and Zhenwen Ren. Incom- plete multi-view clustering with paired and balanced dy- namic anchor learning.IEEE Transactions on Multimedia, 27:1486–1497, 2024. 2

    2024

  54. [54]

    Unsupervised label noise model- ing and loss correction

    Eric Arazo, Diego Ortego, Paul Albert, Noel O’Connor, and Kevin McGuinness. Unsupervised label noise model- ing and loss correction. InICML, pages 312–321. PMLR,

  55. [55]

    A closer look at memorization in deep net- works

    Devansh Arpit, Stanisław Jastrz ´kebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep net- works. InICML, pages 233–242. PMLR, 2017. 2

    2017

  56. [56]

    Maskcon: Masked con- trastive learning for coarse-labelled dataset

    Chen Feng and Ioannis Patras. Maskcon: Masked con- trastive learning for coarse-labelled dataset. InCVPR, pages 19913–19922, 2023. 2

    2023

  57. [57]

    Noisebox: Toward more efficient and effective learning with noisy labels.IEEE TCSVT, 34(11):11914–11928,

    Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Noisebox: Toward more efficient and effective learning with noisy labels.IEEE TCSVT, 34(11):11914–11928,

  58. [58]

    Learning with noisy correspon- dence.IJCV, 132(9):3656–3677, 2024

    Zhenyu Huang, Peng Hu, Guocheng Niu, Xinyan Xiao, Jiancheng Lv, and Xi Peng. Learning with noisy correspon- dence.IJCV, 132(9):3656–3677, 2024. 2, 3

    2024

  59. [59]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  60. [60]

    Cogvlm: Visual expert for pretrained language models.NeurIPS, 37:121475–121499, 2024

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models.NeurIPS, 37:121475–121499, 2024

    2024

  61. [61]

    Heterogeneous Adaptive Policy Optimization: Tailoring Optimization to Every Token's Nature

    Zheng Liu, Mengjie Liu, Siwei Wen, Mengzhang Cai, Bin Cui, Conghui He, and Wentao Zhang. From uniform to het- erogeneous: Tailoring policy optimization to every token’s nature.arXiv preprint arXiv:2509.16591, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  62. [62]

    FBS: Modeling Native Parallel Reading inside a Transformer

    Tongxi Wang. Fbs: Modeling native parallel reading inside a transformer.arXiv preprint arXiv:2601.21708, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  63. [63]

    Curriculum-RLAIF: Curriculum Alignment with Reinforcement Learning from AI Feedback

    Mengdi Li, Jiaye Lin, Xufeng Zhao, Wenhao Lu, Peilin Zhao, Stefan Wermter, and Di Wang. Curriculum-rlaif: Curriculum alignment with reinforcement learning from ai feedback.arXiv preprint arXiv:2505.20075, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  64. [64]

    Sf2t: Self-supervised fragment finetuning of video-llms for fine-grained under- standing.arXiv preprint arXiv:2504.07745, 2025

    Yangliu Hu, Zikai Song, Na Feng, Yawei Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Sf2t: Self-supervised fragment finetuning of video-llms for fine-grained under- standing.arXiv preprint arXiv:2504.07745, 2025

    work page arXiv 2025

  65. [65]

    Stable and explainable personality trait evaluation in large language models with internal activations, 2026

    Xiaoxu Ma, Xiangbo Zhang, and Zhenyu Weng. Stable and explainable personality trait evaluation in large language models with internal activations, 2026

    2026

  66. [66]

    MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning

    Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, and Xunliang Cai. Maspo: Unifying gradient utilization, prob- ability mass, and signal reliability for robust and sample- efficient llm reasoning.arXiv preprint arXiv:2602.17550, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  67. [67]

    Expseek: Self-triggered experience seeking for web agents, 2026

    Wenyuan Zhang, Xinghua Zhang, Haiyang Yu, Shuaiyi Nie, Bingli Wu, Juwei Yue, Tingwen Liu, and Yongbin Li. Expseek: Self-triggered experience seeking for web agents, 2026

    2026

  68. [68]

    Synthvlm: High-efficiency and high-quality synthetic data for vision language models

    Zheng Liu, Hao Liang, Xijie Huang, Wentao Xiong, Qin- han Yu, Linzhuang Sun, Chong Chen, Conghui He, Bin Cui, and Wentao Zhang. Synthvlm: High-efficiency and high-quality synthetic data for vision language models. arXiv preprint arXiv:2407.20756, 3, 2024

    work page arXiv 2024

  69. [69]

    Tracking drift: Variation-aware entropy scheduling for non-stationary reinforcement learning.arXiv preprint arXiv:2601.19624, 2026

    Tongxi Wang, Zhuoyang Xia, Xinran Chen, and Shan Liu. Tracking drift: Variation-aware entropy scheduling for non-stationary reinforcement learning.arXiv preprint arXiv:2601.19624, 2026

    work page arXiv 2026

  70. [70]

    Multi-objective unlearning in recommender systems via preference guided pareto exploration.IEEE TSC, 2025

    Yuyuan Li, Yizhao Zhang, Weiming Liu, Xiaohua Feng, Zhongxuan Han, Chaochao Chen, and Chenggang Yan. Multi-objective unlearning in recommender systems via preference guided pareto exploration.IEEE TSC, 2025

    2025

  71. [71]

    Ultrare: Enhancing receraser for recommendation unlearn- ing via error decomposition.NeurIPS, 36:12611–12625,

    Yuyuan Li, Chaochao Chen, Yizhao Zhang, Weiming Liu, Lingjuan Lyu, Xiaolin Zheng, Dan Meng, and Jun Wang. Ultrare: Enhancing receraser for recommendation unlearn- ing via error decomposition.NeurIPS, 36:12611–12625,

  72. [72]

    Pair: Complementarity-guided disentanglement for composed image retrieval

    Zhiheng Fu, Zixu Li, Zhiwei Chen, Chunxiao Wang, Xuemeng Song, Yupeng Hu, and Liqiang Nie. Pair: Complementarity-guided disentanglement for composed image retrieval. InICASSP, pages 1–5. IEEE, 2025. 3

    2025

  73. [73]

    Me- dian: Adaptive intermediate-grained aggregation network for composed image retrieval

    Qinlei Huang, Zhiwei Chen, Zixu Li, Chunxiao Wang, Xuemeng Song, Yupeng Hu, and Liqiang Nie. Me- dian: Adaptive intermediate-grained aggregation network for composed image retrieval. InICASSP, pages 1–5. IEEE,

  74. [74]

    Fash- ion iq: A new dataset towards retrieving images by natural language feedback

    Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fash- ion iq: A new dataset towards retrieving images by natural language feedback. InCVPR, pages 11307–11317, 2021. 3, 6, 2

    2021

  75. [75]

    Image retrieval on real-life images with pre-trained vision-and-language models

    Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. InICCV, pages 2125–2134, 2021. 6, 2

    2021

  76. [76]

    Finecir: Explicit parsing of fine-grained modification semantics for composed image retrieval,

    Zixu Li, Zhiheng Fu, Yupeng Hu, Zhiwei Chen, Haokun Wen, and Liqiang Nie. Finecir: Explicit parsing of fine- grained modification semantics for composed image re- trieval.arXiv preprint arXiv:2503.21309, 2025. 3

    work page arXiv 2025

  77. [77]

    Computing nodes for plane data points by constructing cubic polynomial with constraints

    Hua Wang and Fan Zhang. Computing nodes for plane data points by constructing cubic polynomial with constraints. CAGD, 111:102308, 2024. 3

    2024

  78. [78]

    ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch

    Zheng Liu, Honglin Lin, Chonghan Qin, Xiaoyang Wang, Xin Gao, Yu Li, Mengzhang Cai, Yun Zhu, Zhanping Zhong, Qizhi Pei, et al. Chartverse: Scaling chart reason- ing via reliable programmatic synthesis from scratch.arXiv preprint arXiv:2601.13606, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  79. [79]

    Who stole your data? a method for detecting unauthorized rag theft.arXiv preprint arXiv:2510.07728, 2025

    Peiyang Liu, Ziqiang Cui, Di Liang, and Wei Ye. Who stole your data? a method for detecting unauthorized rag theft.arXiv preprint arXiv:2510.07728, 2025

    work page arXiv 2025

  80. [80]

    Not All Directions Matter: Towards Structured and Task-Aware Low-Rank Model Adaptation

    Xi Xiao, Chenrui Ma, Yunbei Zhang, Chen Liu, Zhuxu- anzi Wang, Yanshu Li, Lin Zhao, Guosheng Hu, Tianyang Wang, and Hao Xu. Not all directions matter: Toward struc- tured and task-aware low-rank adaptation.arXiv preprint arXiv:2603.14228, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

Showing first 80 references.