acceptodds
Under review as a conference paper at ICLR 2027

Ontology-Guided Semantic Mixing and Adaptive Low-Rank Experts for Long-Tailed Recognition

Abstract

Learning from long-tailed data remains an obstacle for recognition systems, because rare categories carry the greatest decision value but provide the weakest training signal. However, correcting class imbalance alone does not ensure that augmented samples preserve domain-specific semantics. Mixing-based augmentation guided by generic vision-language embeddings selects partners by linguistic co-occurrence rather than domain semantics, injecting semantic noise where data are scarcest. We propose a knowledge-guided learning framework that couples an explicit symbolic knowledge source with parameter-efficient adaptation. Ontology-Guided Semantic Mixing (OGSM) transforms frozen CLIP text embeddings into ontology-consistent representations through triplet metric learning over expertcurated relations, producing a similarity matrix that restricts augmentation to semantically admissible class pairs. Bag of LoRAs (BAGL) then supplies the discriminative capacity the enriched distribution demands. The number of low-rank experts is derived in closed form from the class count and the imbalance factor, which removes per-dataset hyperparameter search, and per-expert logit tempering is optimized by a secondary objective that jointly minimizes mean task loss and inter-expert loss variance. Because the knowledge component is precomputed and architecture-agnostic, it transfers across tasks without modification. On four long-tailed land-cover benchmarks the framework attains 91.1% average accuracy with consistent tail-class gains over the strongest baseline, and OGSM alone improves tail mIoU on iSAID segmentation by 14.5 points over CutMix.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-Tailed Recognition, Semantic Mixing, LoRA Experts, Ontology, Remote Sensing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2ttjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ontology-guided semantic mixing and an ensemble of LoRA experts for long-tailed remote sensing recognition. An ontology-trained text adapter guides augmentation, while dataset-dependent expert allocation and adaptive logit adjustment support classification. Evaluation includes four classification datasets and iSAID segmentation.

Critical strengths

Incorporating domain knowledge into augmentation addresses a relevant limitation of text-based similarity in specialized recognition.

Component ablations and evaluation across classification and segmentation provide useful evidence of the method, particularly for tail classes.

Critical weaknesses

Key methodological claims are insufficiently justified. The ontology and triplet construction are underspecified, and Eq. 4 assigns positive sampling probability to incompatible classes unless an additional mask is used. Likewise, Eq. 9 is presented as an optimal allocation without a supporting optimization problem or derivation.

Numerical inconsistencies weaken confidence in the evaluation. In Table 2, the listed OGSM class IoUs average approximately 67.63%, rather than the reported 72.1%, changing the claimed overall segmentation improvement.

Questions for authors

What ontology, relation definitions, and triplet construction procedure were used? Does Eq. 4 include an unreported compatibility mask?

How do you reconcile Table 2’s per-class IoUs with the reported overall mIoU?

Why does Eq. 9 give the best number of experts? Does your method still perform better when the baselines use the same number of experts and the same number of trainable parameters?

Reviewer_ov4MOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a two-stage framework for long-tailed remote sensing recognition consisting of Ontology-Guided Semantic Mixing (OGSM) for CutMix partner selection and Bag of LoRAs (BAGL) for multi-expert adaptation on a frozen CLIP ViT-B/16 backbone. However, a detailed technical evaluation is superseded by severe research integrity and reproducibility issues, specifically multiple hallucinated/corrupted bibliographic references and the complete omission of the formal ontology O\mathcal{O} that supervises the proposed method.

Critical strengths

None

Critical weaknesses
  • Multiple Hallucinated / Corrupted Citations and Policy Violations (Grounds for Desk Rejection): Per the ICLR 2027 Reviewer Guidelines and AI Policy, LLM-generated falsehoods and hallucinated citations are grounds for desk rejection. A bibliographic audit of this submission reveals multiple corrupted or fabricated references:
    1. Chen et al. (2024) ("Jiawei Chen, Jingchao Wang, and Qiulong Yuan. Dynamic learnable logit adjustment for long-tailed visual recognition. IEEE TCSVT, 34(10):9803–9815, 2024"): The actual authors of this exact IEEE TCSVT 2024 article are Enhao Zhang, Chuanxing Geng, and Songcan Chen (https://ieeexplore.ieee.org/document/10487932/). The author list in the submission is completely fabricated.
    2. Tao et al. (2023) ("Jian Tao, Bingchen Zhao, and Hao Yang. Local and global logit adjustments for long-tailed learning. In ICCV, pp. 11773–11783, 2023"): The actual authors of this ICCV 2023 paper are Yingfan Tao, Jingna Sun, Hao Yang, Li Chen, Xu Wang, Wenming Yang, Daniel Du, and Min Zheng (https://openaccess.thecvf.com/content/ICCV2023/html/Tao_Local_and_Global_Logit_Adjustments_for_Long-Tailed_Learning_ICCV_2023_paper.html).
    3. Yang et al. (2025) ("Minghao Yang, Guang Li, Guofeng Cui, Yuzhen Niu, and Haowei Huang. Adaptive shared experts with lora-based mixture of experts for multi-task learning. arXiv:2510.00570, 2025"): The actual authors of arXiv:2510.00570 (https://arxiv.org/abs/2510.00570) are Minghao Yang, Ren Togo, Guang Li, Takahiro Ogawa, and Miki Haseyama.
    4. Unverifiable primary baseline and missing mandatory statements: The main SOTA competitor cited throughout the paper ("Yifan Wang et al., Dmrs: Long-tailed remote sensing recognition via semantic-aware mixing and diversity experts, Int. J. Appl. Earth Obs. Geoinf., 141:104623, 2025") cannot be located in indexing databases under that volume/article ID, the RemoteCLIP (https://arxiv.org/abs/2306.11029) citation omits its author list (liu (2024)), and the manuscript omits the mandatory ICLR 2027 AI Use Statement.
  • Omission of the Core Ontology O\mathcal{O} and Internal Contradictions: Even setting aside the citation issues, the paper never specifies or provides the "formal remote sensing ontology O\mathcal{O}" used to train Stage 1 across the five datasets, rendering the method irreproducible. Furthermore, Section 4.1 claims evaluation on 117,000 natural images of RSD46-WHU while Appendix B (Table 4) reports training on only 3,847 synthetically subsampled images, and Equation (9) confuses natural and base-10 logarithms.
Questions for authors
  1. Can the authors explain how the fabricated author lists for Chen et al. (2024), Tao et al. (2023), and Yang et al. (2025) were generated, and provide a verifiable DOI/URL for the primary baseline DMRS (Wang et al., 2025)?
Reviewer_WianOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a two-stage knowledge-guided framework for long-tailed remote sensing recognition, consisting of Ontology-Guided Semantic Mixing (OGSM) and Bag of LoRAs (BAGL). OGSM adapts CLIP's general embedding to the remote sensing domain through triplet metric learning supervised by an expert-curated remote sensing ontology, and subsequently performs CutMix-based data augmentation. BAGL introduces multiple LoRA experts to enhance classification performance on long-tailed datasets. It adaptively determines the number and tempering of experts to mitigate class imbalance, promote complementary expert learning, and improve tail-class recognition. The method is evaluated on four long-tailed remote sensing classification datasets and one semantic segmentation dataset. The reported results show improvements in tail-class accuracy and tail mIoU over several baselines.

Critical strengths
  1. The paper is well-motivated to integrate domain ontology into vision-language-guided data augmentation. OGSM addresses this gap between general CLIP embedding and domain knowledge by introducing an independently curated ontology and using triplet metric learning to adapt the embedding space. The use of a residual MLP adapter is lightweight and general. It preserves the general-purpose knowledge encoded by CLIP while incorporating domain-specific constraints.

  2. The experimental evaluation is comprehensive and demonstrates strong empirical results. Extensive experiments are conducted across multiple remote sensing classification benchmarks, achieving strong classification performance and demonstrating the effectiveness of the proposed method in addressing long-tailed distributions. Moreover, the evaluation extends beyond classification to semantic segmentation, providing additional evidence of the potential generalizability of OGSM across different tasks. The reported improvement in tail mIoU is particularly encouraging.

Critical weaknesses
  1. lthough the overall effectiveness of OGSM is supported by the reported experiments, the specific contribution of ontology-supervised triplet learning remains unclear. Since the authors already have an expert-curated remote sensing ontology, why not directly construct the semantic similarity matrix based on inter-class distances in the ontology graph, without training an additional MLP adapter? The paper lacks a comparison between this simpler ontology-based approach and the proposed method, which preserves CLIP's continuous semantic representations and aligns them with the ontology through triplet learning. Such a comparison is necessary to demonstrate whether the additional learning process provides meaningful benefits over directly using ontology-derived similarities.

  2. Adaptive expert specialization lacks mechanistic evidence. The paper attributes part of the performance improvement to preventing expert collapse. However, the presented accuracy-based ablations do not directly quantify expert diversity, representation similarity, or prediction disagreement. Additional analyses of individual expert performance, and pairwise expert diversity would help substantiate the claimed mechanism.

Questions for authors
  1. OGSM relies heavily on an expert-curated remote sensing ontology to generate triplet supervision and construct the domain-aligned semantic similarity matrix. Could the authors clarify how the ontology was constructed? The criteria used to define positive and negative class relationships and other more details should be clarified.

  2. In Eq. (12), an inter-expert loss variance regularization to encourage balanced expert performance. However, reducing the variance of expert losses does not necessarily lead to better ensemble accuracy. Whether it may also constrain well-performing experts without sufficiently improving weaker ones? And whether it may potentially reduce expert diversity due to limited preservation of complementary specializations?

  3. In Eq. (13), the factors are initialized with evenly spaced values in [−0.9, 2.5]. Could the authors explain the motivation behind the initialization range ([-0.9,2.5])?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.