acceptodds
Withdrawn from ICLR 2027

ACD-Retrieval: Towards Multimodal Knowledge Transfer via Aligned Cross-domain Demonstration Retrieval

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal large language models (MLLMs) have achieved remarkable progress in complex reasoning tasks, yet existing reasoning enhancement methods still rely on high-quality in-domain annotations or expert demonstrations. We observe that, even when in-domain demonstrations are unavailable, demonstrations from other domains can provide transferable reasoning knowledge for multimodal ICL. However, existing multimodal retrieval methods mainly rely on simple embedding similarity, which may not effectively reflect how well a cross-domain demonstration transfers to the target reasoning task, potentially leading to unstable retrieval. To address this challenge, we propose **ACD-Retrieval**, a unified framework for multimodal cross-domain ICL. ACD-Retrieval performs domain-wise centering and SVD-based subspace projection to obtain more reliable cross-domain similarity estimation, while a self-selection mechanism suppresses negative transfer during inference. Extensive experiments demonstrate that our method consistently achieves positive transfer across nine MLLMs, while exhibiting strong generalization across diverse multimodal cross-domain settings.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Multimodal large language models, cross-domain transfer learning, in-context learning, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_52HSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ACD-Retrieval for multimodal cross-domain in-context learning when target-domain demonstrations are unavailable. The method centers embeddings separately within each domain, retrieves demonstrations using cosine similarity in source-domain PCA subspaces, and asks the same model to select between zero-shot and demonstration-augmented responses. Experiments treat mathematical disciplines in MATH-V as distinct domains and report improvements in overall accuracy across multiple model families and sizes.

Critical strengths
  1. The complete framework achieves consistent improvements in overall accuracy across multiple model families and sizes.
  2. The paper includes preliminary component ablations and results for all evaluated domains, helping readers assess component effects and domain-specific performance.
Critical weaknesses
  1. Limited technical novelty and insufficient positioning relative to established methods. The SVD decomposition, cumulative-energy-based dimensionality selection, and principal-direction projection used in the proposed Adaptive Subspace Projection are mathematically equivalent to standard PCA. The specific adaptation consists primarily of domain-wise centering and projecting target queries into source-domain principal subspaces. However, the paper does not mention PCA or clearly distinguish these standard operations from its additional design choices.

    Closely related work has already investigated cross-domain reasoning demonstration retrieval [1] and multimodal demonstration transfer without task-specific support data [2], but neither is cited or discussed. The contribution currently appears to be primarily a combination of existing techniques evaluated in the MATH-V cross-discipline setting. Its independent methodological contribution and claimed originality are not sufficiently established.

  2. The description of a shared aligned subspace does not fully match the actual construction. The method builds a separate principal subspace for each source domain, without explicitly optimizing source–target subspace alignment or constructing a common projection space across all domains. It remains insufficiently explained whether cosine scores computed in different subspaces are directly comparable and how candidates from different source domains are ranked and merged.

  3. Important details of the experimental protocol are unclear. The paper does not clearly specify which samples are used to estimate target-domain means, how demonstration pools and evaluation data are partitioned, whether demonstrations contain reasoning traces, or which data are used for hyperparameter selection. These details affect the interpretation of the setting, the reported results, and the method's practical applicability.

[1] Yan et al. Towards Effective In-context Cross-domain Knowledge Transfer via Domain-invariant-neurons-based Retrieval. 2026.

[2] Fang et al. Why is a Bird's Caption a Good Demonstration? Towards Effective Multimodal In-Context Learning without Dedicated Data. ACM Multimedia, 2025.

[3] Weng et al. Optimizing Knowledge Integration in Retrieval-Augmented Generation with Self-Selection. 2025.

Questions for authors

Please clarify the paper's independent contribution beyond combining existing techniques:

  1. How does the proposed subspace construction differ mathematically from standard PCA, and why is this connection not explicitly acknowledged?
  2. Relative to prior work on cross-domain demonstration retrieval, multimodal demonstration transfer, and answer self-selection [1–3], what new mechanism or empirical finding does this paper establish?
  3. What substantive challenges does the MATH-V cross-discipline setting introduce beyond providing a new evaluation instance of cross-domain ICL?
Reviewer_oEtUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multimodal cross-domain in-context learning for reasoning. The authors propose ACD-Retrieval, a training-free retrieval framework consisting of domain-wise centering, SVD-based subspace projection, and a self-selection step that chooses between zero-shot and ICL outputs. Experiments are conducted on MATH-V by treating different math sub-domains as different domains, and results are reported across multiple MLLMs. The paper claims that the proposed retrieval and selection strategy improves cross-domain transfer compared with zero-shot and simple retrieval baselines.

Critical strengths
  1. The method is clearly presented and easy to follow. The overall framework is simple and well-structured, and the roles of centering, subspace projection, and self-selection are clearly described.

  2. The paper addresses an interesting problem setting: cross-domain multimodal ICL. Studying whether demonstrations from other domains can help multimodal reasoning is a worthwhile direction, especially when in-domain demonstrations are unavailable.

Critical weaknesses
  1. I do not think the experiments are truly cross-domain, since they are all conducted within the same benchmark (MATH-V) and the same broad mathematical reasoning setting. Different sub-domains within one dataset are not sufficient, in my opinion, to support strong claims about general cross-domain knowledge transfer;
  2. The baselines are too limited, and comparing mainly against RICES is not sufficient given the large literature on ICL retrieval and reranking;
  3. The self-selection step functions similarly to self-consistency / answer selection, and a stronger baseline such as RICES + self-selection may already perform well, so the current comparison is not fair; Table 2 also suggests that without self-selection, the method is sometimes worse than zero-shot.
Questions for authors

see Weaknesses above.

Reviewer_fYQZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ACD-Retrieval studies multimodal in-context learning when demonstrations come from different mathematical domains. It centers embeddings separately by domain, projects queries and demonstrations into source-domain SVD subspaces, and retrieves examples by cosine similarity. The model then selects between zero-shot and demonstration-conditioned answers. Experiments use the sixteen disciplines in MATH-V and nine models from three families, reporting improved overall accuracy for the complete pipeline.

Critical strengths

1.The setting is useful when target-domain demonstrations are unavailable. The retrieval transformation is simple and requires no parameter training, making the method relatively easy to apply with an existing multimodal encoder. 2.The evaluation covers nine model configurations, and the full method improves overall accuracy over zero-shot inference for all of them. Table 2 also separates aligned retrieval from the final self-selection step, showing that both affect performance.

Critical weaknesses

1.The strongest gains are not cleanly attributable to aligned retrieval. Self-selection adds a zero-shot answer and a further judging call, but Table 2 does not show ordinary embedding retrieval with the same selection procedure. On Qwen-7B, aligned retrieval alone scores 24.31 versus 25.66 for zero-shot, while the full pipeline reaches 27.57. A matched selection control is important to establish the retrieval contribution. 2.The evaluation protocol needs clarification. Appendix A.3 selects the SVD threshold and demonstration counts using Overall accuracy without identifying a separate validation set. The demonstration-pool split and the samples used to estimate the target-domain mean are also unclear. These details affect the interpretation of cross-domain generalization.

Questions for authors

1.Were RICES or ordinary embedding retrieval evaluated with the same zero-shot comparison and self-selection calls? If so, what gains remain from alignment? 2.Which splits supply demonstrations, target-domain means, and hyperparameter selection? Were final evaluation labels excluded from selecting the threshold and demonstration count?

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

Related papers

Discussion (0)

Sign in to comment.