Pith. sign in

REVIEW 4 major objections 5 minor 32 references

A CLIP-based framework reaches state-of-the-art on lifelong person re-ID with mixed clothing states.

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 ↗

CMLReID uses dynamic text prompts and dual-path prototypes to improve lifelong person re-identification in hybrid clothing states, reporting gains of about 5 mAP over four baselines.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A sensible CLIP-based lifelong ReID method, but the SOTA claim is untested against the closest baseline and the empirical support is unverifiable due to broken references and missing code/data. the 4 major comments →

arxiv 2509.11247 v1 pith:REDFPRE5 submitted 2025-09-14 cs.CV

Contextualized Multimodal Lifelong Person Re-Identification in Hybrid Clothing States

Robert Long , Rongxin Jiang , Mingrui Yan This is my paper
classification cs.CV
keywords person re-identificationlifelong learningcloth-changing Re-IDCLIPprompt learningcatastrophic forgettingLReID-Hybrid
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper introduces LReID-Hybrid, a lifelong person re-identification setting in which a model must keep learning across a stream of datasets while handling both same-cloth and cloth-changing conditions. It proposes CMLReID, a CLIP-based framework with two components: CASP dynamically generates text prompts conditioned on the inferred clothing state, and AKFP maintains separate slow-updated text prototypes for SC and CC and projects visual features into the appropriate space. The authors report that CMLReID outperforms four baselines on four ReID datasets in all six learning orders, with ablations showing each component contributes. The significance is a step toward surveillance systems that keep working as people change clothes over time.

Core claim

The paper claims that CMLReID consistently achieves leading or highly competitive performance across all metrics and reaches state-of-the-art performance in the LReID-Hybrid setting. On the seen-domain benchmark (Order 1), it reports a total average mAP of 46.7% and Rank-1 of 60.5%, surpassing the strongest baseline LwF (45.2% mAP, 59.3% R-1) and other methods like AKA, SFT, and CLIP-ReID. The gains are attributed to two mutually collaborative modules: Context-Aware Semantic Prompt (CASP), which generates adaptive text prompts based on an estimated clothing state to align multi-grained visual cues with semantic text space, and Adaptive Knowledge Fusion and Projection (AKFP), which maintains

What carries the argument

The central objects are CASP (Context-Aware Semantic Prompt) and AKFP (Adaptive Knowledge Fusion and Projection). CASP uses a context encoder to estimate whether an input image is same-cloth or cloth-changing, then combines base semantic prompts with dynamically generated modulation prompts to produce a contextualized text embedding. AKFP maintains two slowly updated text prototypes, one for SC and one for CC, and projects visual features into the appropriate prototype space using a clothing-state classifier and a dual-path projection; a Clothing-State-Aware Projection Loss aligns the projected visual features with the correct prototype. Together they create an alternating training loop that

Load-bearing premise

The paper's central superiority claim relies on the assumption that the four baselines it compares against (AKA, SFT, LwF, CLIP-ReID) are the relevant state of the art and are implemented fairly, yet it omits the method that originally introduced the LReID-Hybrid task, and no code or data are provided to confirm the baseline implementations.

What would settle it

Re-run the LReID-Hybrid protocol (Order 1) including the previously proposed image-text-image method as a baseline and release the code for CMLReID; if that baseline matches or exceeds the reported 46.7% total mAP, the state-of-the-art claim collapses. Additionally, ablate the context encoder by feeding it reversed clothing-state labels; if total mAP does not drop substantially, the context-aware mechanism is not doing the claimed work.

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

If this is right

  • If the reported gains hold, lifelong ReID systems can be trained on mixed clothing-state streams without separating same-cloth and cloth-changing data into different pipelines.
  • The alternating CASP-AKFP loop offers a way to reuse CLIP's pretrained text-image alignment for continual learning, potentially reducing catastrophic forgetting while adapting to new domains.
  • The method's robustness across six learning orders suggests that order-dependent performance variance can be mitigated with dual-path prototypes and context-aware prompts.
  • Human evaluation results indicate practical usability in cloth-changing surveillance, with a reported improvement in top-5 retrieval accuracy from 68.5% to 75.3% on CC scenarios.

Where Pith is reading between the lines

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

  • Beyond the paper: if accurate clothing-state estimation is what drives the gains, then the method's performance may degrade in genuinely ambiguous or mixed-state images; a test that artificially corrupts the state classifier's output could reveal this dependency.
  • The dual-path slow learner idea could be transferred to other fine-grained continual classification tasks where the key discrimination axis is a binary or categorical attribute (e.g., age, pose, or style).
  • One untested extension is to combine CMLReID with unsupervised or self-supervised learning, since the current setting assumes identity labels; that would broaden applicability to real-world streams without annotations.
  • Because the paper reports generalization to unseen SC and CC datasets only qualitatively, a quantitative breakdown of those results would sharpen the claim of transferable, clothing-invariant features.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper introduces LReID-Hybrid, a lifelong person re-identification setting in which same-cloth (SC) and cloth-changing (CC) datasets are learned sequentially, and proposes CMLReID, a CLIP-based method with two components: Context-Aware Semantic Prompt (CASP) learning and Adaptive Knowledge Fusion and Projection (AKFP). Training alternates between prompt learning and image-encoder updates, and a clothing-state-aware projection loss aligns visual features with slowly updated text prototypes. Experiments on Market-1501, LTCC, MSMT17, and PRCC under six learning orders claim state-of-the-art results, with ablations, sensitivity analyses, and a small human evaluation.

Significance. If the empirical claims could be verified, the task definition and framework would be a useful contribution to an active area. The method is described in sufficient detail that the design is plausible, and the ablations are structured to isolate the main components. However, the empirical case is currently weak: the most relevant baseline is missing, several dataset references are incorrect, and no code, data, or supplementary material is provided. The paper's central claim of state-of-the-art performance is therefore not established by the evidence presented.

major comments (4)
  1. [§4.2 / Table 1] The state-of-the-art claim is not tested against Teata [11], the method that introduced the LReID-Hybrid task and is described in Section 2.1 as using an image-text-image closed loop with structured semantic prompts — the same design family as CMLReID. Excluding this task-defining baseline leaves the central claim unverified. Please add a direct comparison, or justify quantitatively why it is omitted.
  2. [§4.1] The dataset citations are incorrect. Ref [4] is not Market-1501, Ref [5] is not MSMT17, Ref [7] is not PRCC, Ref [8] is not CUHK01, and Ref [9] does not match Celeb-ReID. These are central experimental resources; the mismatch makes the setup untraceable. Correct all dataset references and specify dataset versions, splits, and protocols.
  3. [§4.8] Generalization results are deferred to 'the supplementary material', but no supplementary file is included, and no code is provided. Since generalization to unseen SC and CC domains is one of the stated contributions, the actual numerical results must be in the main text or an accessible supplement, with enough detail for independent reproduction.
  4. [§4.3 / Table 2] The order-robustness experiment reports only the SFT baseline. The claim of 'consistently achieves leading or highly competitive performance' across all metrics and learning orders is not supported for other baselines. Report results for AKA, LwF, and CLIP-ReID on all six orders as well, and provide multiple runs or statistical tests; single-run differences of 0.2–0.5 mAP may be within noise.
minor comments (5)
  1. [Abstract] The sentence 'Mismatched representations and forgetting from one task to the next are significant issues, we address this with' is a run-on; consider splitting into two sentences.
  2. [Eq. (6)] The notation 'Mean(e_T^{(t)} | ground truth state is s)' is ambiguous; specify that the mean is over the batch elements whose ground-truth clothing state is s.
  3. [§3.2] The paper resizes images to 256×128, but CLIP ViT-B/16 natively expects 224×224 input. Please clarify how positional embeddings are adapted, or state that this was handled within the implementation.
  4. [§4.9] The human evaluation uses only 20 query images per setting and reports no inter-annotator agreement or selection details. Add agreement metrics and a clear query-selection protocol.
  5. [References] Multiple references are arXiv preprints without venue/DOI. Provide full publication data where available, and re-verify every reference because of the dataset citation errors noted in the major comments.

Circularity Check

0 steps flagged

No significant circularity; the central method is an end-to-end training objective with at most a minor non-load-bearing self-citation.

full rationale

This is an empirical machine-learning paper, not a derivation chain in which a prediction is shown equal to its inputs by construction. The claimed state-of-the-art result (Table 1) is a comparison of retrieval metrics after training, not a quantity fitted from the evaluation data. The text prototypes in Eq. (6) are updated from current-batch contextualized text embeddings, and Eq. (9) aligns projected visual features to those prototypes; this mutual dependence is a standard alternating two-stage training procedure (Section 3.2), not a logical circularity. Hyperparameters lambda and beta_s are tuned on validation (Tables 6-7), but final test-set numbers are not forced by that tuning. The only apparent self-citation is [23] (MoPE, likely by co-author Rongxin Jiang), used in Related Work as an example of prompt-expert methods; it is not load-bearing for any claim. Separate reproducibility/correctness concerns--omission of the task-originator Teata [11] from Table 1, wrong dataset references ([4], [5], [7]), and unseen-domain numbers deferred to an absent supplementary (Section 4.8)--undermine the SOTA claim but are not circularity. The paper's assertion that it 'introduce[s] the LReID-Hybrid task' conflicts with Section 2.1 crediting [11] with introducing it; that is an attribution/novelty inconsistency, not a circular reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on the availability of clothing-state labels, the suitability of CLIP, and stable alternating optimization. The hyperparameters lambda and beta_s are tuned on the validation set, adding a small set of fitted constants that the reported performance depends on.

free parameters (2)
  • projection loss weight lambda = 0.5
    Tuned via sensitivity analysis (Table 6); the default value gives the best total average mAP.
  • slow learner momentum beta_s = 0.001
    Tuned via sensitivity analysis (Table 7); the default value gives the best total average mAP.
axioms (3)
  • domain assumption CLIP's pretrained visual and text encoders provide a useful shared embedding space for person re-identification.
    The entire framework is built on CLIP ViT-B/16 and its text encoder; if this pretrained model is not suitable for ReID, the method fails.
  • domain assumption Ground-truth clothing-state labels (SC vs CC) are available for every training image.
    The Clothing-State-Aware Projection Loss in Eq. (9) uses the ground truth state s for each image; the context encoder is also supervised to predict this label.
  • domain assumption The two-stage alternating optimization (CASP then AKFP) converges and does not introduce catastrophic interference between the stages.
    The method relies on alternating phases, but no convergence analysis or stability guarantees are provided.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Contextualized Multimodal Lifelong Person Re-Identification in Hybrid Clothing States." pith.science (2026). https://pith.science/paper/REDFPRE5

@misc{pith2026250911247,
  author       = {Pith},
  title        = {Pith review of: Contextualized Multimodal Lifelong Person Re-Identification in Hybrid Clothing States},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/REDFPRE5}},
  note         = {Machine review of arXiv:2509.11247}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Person Re-Identification (ReID) has several challenges in real-world surveillance systems due to clothing changes (CCReID) and the need for maintaining continual learning (LReID). Previous existing methods either develop models specifically for one application, which is mostly a same-cloth (SC) setting or treat CCReID as its own separate sub-problem. In this work, we will introduce the LReID-Hybrid task with the goal of developing a model to achieve both SC and CC while learning in a continual setting. Mismatched representations and forgetting from one task to the next are significant issues, we address this with CMLReID, a CLIP-based framework composed of two novel tasks: (1) Context-Aware Semantic Prompt (CASP) that generates adaptive prompts, and also incorporates context to align richly multi-grained visual cues with semantic text space; and (2) Adaptive Knowledge Fusion and Projection (AKFP) which produces robust SC/CC prototypes through the use of a dual-path learner that aligns features with our Clothing-State-Aware Projection Loss. Experiments performed on a wide range of datasets and illustrate that CMLReID outperforms all state-of-the-art methods with strong robustness and generalization despite clothing variations and a sophisticated process of sequential learning.

Figures

Figures reproduced from arXiv: 2509.11247 by Mingrui Yan, Robert Long, Rongxin Jiang.

Figure 1
Figure 1. Figure 1: Motivation diagram contrasting the limitations of existing ReID methods with the proposed LReID-Hybrid task and CMLReID solution. treat cloth-changing situations as entirely separate tasks. We argue that a truly versatile and robust LReID system must be capable of simultaneously handling both same-cloth and cloth-changing conditions within a continuous learning pro￾cess. This leads us to propose the LReID-… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed CMLReID framework, illustrating the CASP and AKFP modules with their interactions for lifelong person re-identification under hybrid clothing states. A dedicated context encoder Ectx then processes fv to produce a compact con￾text vector c. This encoder, typically implemented as a lightweight multi-layer perceptron (MLP) or a small transformer block, incorporates a self… view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy of Context Encoder Ectx in Identifying Clothing States (on validation sets) SC Image (e.g., Market-1501) CC Image (e.g., LTCC) Ambiguous Image Input Image Type 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Similarity 0.89 0.58 0.75 0.62 0.87 0.73 Similarity to SC Concept Similarity to CC Concept [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cosine Similarity of CASP Embeddings with State-Specific Concepts 4.6 Analysis of Adaptive Knowledge Fusion and Projection (AKFP) Strategy The Adaptive Knowledge Fusion and Projection (AKFP) strategy is crucial for lifelong knowledge adaptation and handling hybrid clothing states. We analyze two key aspects: the behavior of dynamic projection weights and the quality of the resulting projected feature space… 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

32 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    CoRR (2021)

    Lin, X., Ren, P., Yeh, C., Yao, L., Song, A., Chang, X.: Unsupervised person re-identification: A systematic survey of challenges and solutions. CoRR (2021)

    2021

  2. [2]

    In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018

    He, L., Liang, J., Li, H., Sun, Z.: Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 7073–7082. Computer Vision Foundation / IEEE Computer Society (2018). https://doi.org/10.1109/C...

    arXiv 2018

  3. [3]

    arXiv preprint arXiv:2505.24517v1 (2025)

    Li, Y., Zhao, J., Chang, H., Hou, R., Shan, S., Chen, X.: un 2clip: Improv- ing clip’s visual detail capturing ability via inverting unclip. arXiv preprint arXiv:2505.24517v1 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:1501.07445v2 (2015)

    Fernando Cordero, L.P.O.: Strong asymptotic arbitrage in the large fractional bi- nary market. arXiv preprint arXiv:1501.07445v2 (2015)

    Pith/arXiv arXiv 2015

  5. [5]

    arXiv preprint arXiv:2412.18874v2 (2024)

    Moghaddam, A.S., Anvari, F., Haghighi, M.M., Fakhari, M., Mohammadi, M.R.: A culturally-aware benchmark for person re-identification in modest attire. arXiv preprint arXiv:2412.18874v2 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    In: Advances in Computer Graphics - 39th Computer Graphics International Confer- ence, CGI 2022, Virtual Event, September 12-16, 2022, Proceedings

    Wang, Z., Jiang, X., Xu, K., Sun, T.: A transformer-based cloth-irrelevant patches feature extracting method for long-term cloth-changing person re-identification. In: Advances in Computer Graphics - 39th Computer Graphics International Confer- ence, CGI 2022, Virtual Event, September 12-16, 2022, Proceedings. pp. 278–289. Springer (2022). https://doi.org...

    work page doi:10.1007/978-3-031-23473-6 2022

  7. [7]

    arXiv preprint arXiv:1903.05829v1 (2019)

    Kumar, R., Chattopadhyay, S., Mani, B.K., Angom, D.: Electric dipole polariz- ability of group-iiia ions using prcc: Large correlation effects from nonlinear terms. arXiv preprint arXiv:1903.05829v1 (2019)

    Pith/arXiv arXiv 1903

  8. [8]

    In: 22nd IEEE/ACM International Conference on Mining Software Repositories, MSR@ICSE 2025, Ottawa, ON, Canada, April 28-29, 2025

    Yashavant, C.S., Chavda, M., Kumar, S., Karkare, A., Karmakar, A.: SCRUBD: smart contracts reentrancy and unhandled exceptions vulnerability dataset. In: 22nd IEEE/ACM International Conference on Mining Software Repositories, MSR@ICSE 2025, Ottawa, ON, Canada, April 28-29, 2025. pp. 149–153. IEEE (2025). https://doi.org/10.1109/MSR66628.2025.00033

    arXiv 2025

  9. [9]

    In: MM ’21: ACM Multimedia Conference, Virtual Event, China, October 20 - 24, 2021

    Jiang, Y., Chen, W., Sun, X., Shi, X., Wang, F., Li, H.: Exploring the quality of GAN generated images for person re-identification. In: MM ’21: ACM Multimedia Conference, Virtual Event, China, October 20 - 24, 2021. pp. 4146–4155. ACM (2021). https://doi.org/10.1145/3474085.3475547

    arXiv 2021

  10. [10]

    CoRR (2025)

    Ha, R., Jiang, S., Li, B., Pan, B., Zhu, Y., Zhang, J., Zhu, X., Gong, S., Wang, J.: Multi-modal multi-platform person re-identification: Benchmark and method. CoRR (2025). https://doi.org/10.48550/ARXIV.2503.17096

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.17096 2025

  11. [11]

    CoRR (2024)

    Wang, Q., Qian, X., Li, B., Fu, Y., Xue, X.: Image-text-image knowledge transfer- ring for lifelong person re-identification with hybrid clothing states. CoRR (2024). https://doi.org/10.48550/ARXIV.2405.16600

    work page doi:10.48550/arxiv.2405.16600 2024

  12. [12]

    In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021

    Pu, N., Chen, W., Liu, Y., Bakker, E.M., Lew, M.S.: Lifelong person re-identification via adaptive knowledge accumulation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7901–7910. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00781

    arXiv 2021

  13. [13]

    arXiv preprint arXiv:2507.13107v1 (2025)

    Guo, X., Cai, Y., Liu, Z., Wang, Z., Pan, L., Li, H.: R2moe: Redundancy-removal mixture of experts for lifelong concept learning. arXiv preprint arXiv:2507.13107v1 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Unsupervised Lifelong Person Re-identification via Contrastive Rehearsal

    Chen, H., Lagadec, B., Brémond, F.: Unsupervised lifelong per- son re-identification via contrastive rehearsal. CoRR (2022). https://doi.org/10.48550/ARXIV.2203.06468 CMLReID 19

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2203.06468 2022

  15. [15]

    Neurocomput- ing p

    Lee, H., Park, J., Oh, J., Eom, C.: Domain generalization for person re- identification: A survey towards domain-agnostic person matching. Neurocomput- ing p. 130763 (2025). https://doi.org/10.1016/J.NEUCOM.2025.130763

    arXiv 2025

  16. [16]

    In: 2024 IEEE International Conference on Multimedia and Expo (ICME)

    Wang, X., Zhang, K.: Adaptive style transfer learning for generalizable person re- identification. In: 2024 IEEE International Conference on Multimedia and Expo (ICME). pp. 1–6. IEEE (2024)

    2024

  17. [17]

    CoRR (2013)

    Satta, R.: Appearance descriptors for person re-identification: a comprehensive review. CoRR (2013)

    2013

  18. [18]

    In: Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024

    Zhou, Y., Li, X., Wang, Q., Shen, J.: Visual in-context learning for large vision- language models. In: Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024. pp. 15890– 15902. Association for Computational Linguistics (2024)

    2024

  19. [19]

    CoRR (2025)

    Truong, T., Bobda, C., Agarwal, N., Luu, K.: MANGO: multimodal attention-based normalizing flow approach to fusion learning. CoRR (2025). https://doi.org/10.48550/ARXIV.2508.10133

    work page doi:10.48550/arxiv.2508.10133 2025

  20. [20]

    CoRR (2024)

    Bleeker, M.J.R., Hendriksen, M., Yates, A., de Rijke, M.: Demonstrating and reducing shortcuts in vision-language representation learning. CoRR (2024). https://doi.org/10.48550/ARXIV.2402.17510

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.17510 2024

  21. [21]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Zhou, Y., Shen, J., Cheng, Y.: Weak to strong generalization for large language models with multi-capabilities. In: The Thirteenth International Conference on Learning Representations (2025)

    2025

  22. [22]

    In: Proceedings of the 62nd Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers), ACL 2024, Bangkok, Thailand, August 11-16, 2024

    Guo, Z., Jin, T., Zhao, Z.: Multimodal prompt learning with missing modalities for sentiment analysis and emotion recognition. In: Proceedings of the 62nd Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers), ACL 2024, Bangkok, Thailand, August 11-16, 2024. pp. 1726–1736. Association for Computational Linguistics (2024). https://d...

    doi:10.18653/v1/2024.acl- 2024

  23. [23]

    arXiv preprint arXiv:2403.10568v3 (2024)

    Jiang, R., Liu, L., Chen, C.: Mope: Mixture of prompt experts for parameter- efficient and scalable multimodal fusion. arXiv preprint arXiv:2403.10568v3 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    arXiv preprint arXiv:2506.12830 (2025)

    Wang, C., Zhou, Y., Wang, Q., Wang, Z., Zhang, K.: Complexbench-edit: Bench- marking complex instruction-driven image editing via compositional dependencies. arXiv preprint arXiv:2506.12830 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    CoRR (2021)

    Qian, H., Dou, Z., Zhu, Y., Ma, Y., Wen, J.: Learning implicit user profiles for personalized retrieval-based chatbot. CoRR (2021)

    2021

  26. [26]

    arXiv preprint arXiv:2309.15494v1 (2023)

    Wang, Y., Zeng, D., Wada, S., Kurihara, S.: Videoadviser: Video knowledge distil- lation for multimodal transfer learning. arXiv preprint arXiv:2309.15494v1 (2023)

    Pith/arXiv arXiv 2023

  27. [27]

    Semantic Item Graph Enhancement for Multimodal Recommendation

    Zhang, X., Zhou, X., Zeng, Z., Niyato, D., Shen, Z.: Semantic item graph enhancement for multimodal recommendation. CoRR (2025). https://doi.org/10.48550/ARXIV.2508.06154

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.06154 2025

  28. [28]

    IEEE Sensors Journal24(9), 14596– 14607 (2024)

    Lin, Z., Zhang, Q., Tian, Z., Yu, P., Lan, J.: Dpl-slam: enhancing dynamic point- line slam through dense semantic methods. IEEE Sensors Journal24(9), 14596– 14607 (2024)

    2024

  29. [29]

    Sensors24(19), 6258 (2024)

    Lin,Z.,Tian,Z.,Zhang,Q.,Zhuang,H.,Lan,J.:Enhancedvisualslamforcollision- free driving with lightweight autonomous cars. Sensors24(19), 6258 (2024)

    2024

  30. [30]

    arXiv preprint arXiv:2504.15320 (2025)

    Li, Q., Tian, Z., Wang, X., Yang, J., Lin, Z.: Efficient and safe planner for automated driving on ramps considering unsatisfication. arXiv preprint arXiv:2504.15320 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    arXiv preprint arXiv:2412.20367 (2024) 20 Long et al

    Wang, J., Zhang, Z., He, Y., Song, Y., Shi, T., Li, Y., Xu, H., Wu, K., Qian, G., Chen, Q., et al.: Enhancing code llms with reinforcement learning in code generation. arXiv preprint arXiv:2412.20367 (2024) 20 Long et al

    Pith/arXiv arXiv 2024

  32. [32]

    arXiv preprint arXiv:2503.23512 (2025)

    Yi, Q., He, Y., Wang, J., Song, X., Qian, S., Yuan, X., Zhang, M., Sun, L., Li, K., Lu, K., et al.: Score: Story coherence and retrieval enhancement for ai narratives. arXiv preprint arXiv:2503.23512 (2025)

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.