Beyond Surface Imitation: Contrastive Modeling for Reasoning Path Alignment in Multimodal In-Context Learning

# Beyond Surface Imitation: Contrastive Modeling for Reasoning Path Alignment in Multimodal In-Context Learning
Source: [https://arxiv.org/html/2609.10177](https://arxiv.org/html/2609.10177)
Conference:Proceedings of the 34th ACM International Conference on Multimedia; November 10–14, 2026; Rio de Janeiro, BrazilProceedings of the 34th ACM International Conference on Multimedia \(MM ’26\), November 10–14, 2026, Rio de Janeiro, BrazilISBN:979\-8\-4007\-2213\-4/2026/11DOI:[10\.1145/3767308\.3836595](https://doi.org/10.1145/3767308.3836595)CCS:Computing methodologies Artificial intelligenceMingbo Yang[https://orcid.org/0009-0000-2952-5044](https://orcid.org/0009-0000-2952-5044)Note:Mingbo Yang and Wenqiang Wang contributed equally to this work\.Affiliation:Sun Yat\-Sen University,Shenzhen,Guangdong,Chinaemail:[yangmb3@mail2\.sysu\.edu\.cn](mailto:[email protected])Wenqiang Wang[https://orcid.org/0009-0008-5034-1379](https://orcid.org/0009-0008-5034-1379)Affiliation:Sun Yat\-Sen University,Shenzhen,Guangdong,Chinaemail:[wangwq69@mail2\.sysu\.edu\.cn](mailto:[email protected]),Zhaolu Kang[https://orcid.org/0009-0000-1163-1615](https://orcid.org/0009-0000-1163-1615)Affiliation:Peking University,Beijing,Beijing,Chinaemail:[zlkang25@stu\.pku\.edu\.cn](mailto:[email protected]),Peng Chen[https://orcid.org/0009-0001-0327-4053](https://orcid.org/0009-0001-0327-4053)Affiliation:Sun Yat\-Sen University,Shenzhen,Guangdong,Chinaemail:[chenpeng052@gmail\.com](mailto:[email protected]),Yannan Chen[https://orcid.org/0009-0007-1896-3568](https://orcid.org/0009-0007-1896-3568)Affiliation:Sun Yat\-Sen University,Shenzhen,Guangdong,ChinaAffiliation:Pengcheng Laboratory,Shenzhen,Guangdong,Chinaemail:[chenyn288@mail2\.sysu\.edu\.cn](mailto:[email protected]),Sunshang Wang[https://orcid.org/0009-0004-9237-2887](https://orcid.org/0009-0004-9237-2887)Affiliation:Tianjin University of Science and Technology,Tianjin,Tianjin,Chinaemail:[sunshangwww@mail\.tust\.edu\.cn](mailto:[email protected])andYan Xiao[https://orcid.org/0000-0002-2563-083X](https://orcid.org/0000-0002-2563-083X)Note:Corresponding author\.Affiliation:Sun Yat\-Sen University,Shenzhen,Guangdong,Chinaemail:[xiaoyan\.hhu@gmail\.com](mailto:[email protected])

© cc

###### Abstract\.

In\-context learning \(ICL\) is widely used in multimodal large language models \(MLLMs\) and achieves strong performance across a wide range of multimodal tasks\. However, existing multimodal ICL methods often rely on surface level imitation of in\-context demonstrations, making it difficult for MLLMs to align their responses with the reasoning path required by the given multimodal input\. This limitation becomes more pronounced in complex multimodal tasks, thereby restricting further improvements in MLLM performance\. To address this issue, we propose a new multimodal ICL framework that combines contrastive demonstration modeling with the self\-refinement capability of MLLMs\. Specifically, our framework reformulates each demonstration by explicitly contrasting a suboptimal response with a better response under the same input, together with a reasoning path that reveals how the response should be refined\. This contrastive formulation makes the reasoning path toward the desired response more explicit and guides the MLLM beyond superficial imitation\. Furthermore, because effective refinement depends on the current response, we introduce a response\-conditioned retrieval mechanism to select demonstrations whose reasoning paths are more relevant to the current response\. In addition, we use a lightweight alignment controller to predict response quality and determine whether further refinement is needed\. Experiments on three types of multimodal tasks show that the proposed framework consistently improves MLLM performance, with particularly notable gains on visual question answering \(VQA\)\.

###### Keywords:

Multimodal Large Language Models; Multimodal In\-Context Learning

††cc\-license:by## 1\.Introduction

![Comparison between standard multimodal ICL and COMIL, illustrating how contrastive demonstrations and refinement guidance move the model from surface imitation toward evidence-grounded reasoning path alignment.](https://arxiv.org/html/2609.10177v1/motivation.png)Figure 1\.From surface level imitation to reasoning path alignment\. Standard multimodal ICL mainly aligns the final output and may imitate surface patterns from visually similar demonstrations\. In contrast, COMIL uses contrastive demonstrations to make explicit which visual evidence is relevant and how the current response should be revised, thereby guiding the MLLM toward the desired reasoning path\.Comparison between standard multimodal ICL and COMIL, illustrating how contrastive demonstrations and refinement guidance move the model from surface imitation toward evidence\-grounded reasoning path alignment\.Multimodal In\-Context Learning \(multimodal ICL\) has demonstrated remarkable capabilities in adapting Multimodal Large Language Models \(MLLMs\) to novel tasks\([Li et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib1);[Zhan et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib2);[Liao et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib3);[Li et al\., 2025b](https://arxiv.org/html/2609.10177#bib.bib4)\)\. By conditioning MLLMs on demonstrations, multimodal ICL enables task adaptation at inference time without parameter updates\([Zheng et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib36)\)\. This makes multimodal ICL an attractive paradigm for multimodal tasks, as it flexibly leverages the pretrained knowledge of MLLMs\([Wang et al\., 2024a](https://arxiv.org/html/2609.10177#bib.bib37);[Yang et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib38)\)\.

Despite these advantages, current multimodal ICL paradigms often encourage surface level imitation rather than reasoning grounded in multimodal evidence\([Xu et al\., 2025b](https://arxiv.org/html/2609.10177#bib.bib5);[Huang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib6)\)\. Here, surface level imitation refers to relying on superficial demonstration patterns without sufficiently grounding the response in query\-relevant multimodal evidence\. In standard multimodal ICL, MLLMs are conditioned on a sequence of input–output demonstrations\([Luo et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib39);[Wang et al\., 2024b](https://arxiv.org/html/2609.10177#bib.bib40)\)\. However, when faced with complex scenarios that require fine\-grained visual discrimination, MLLMs tend to rely on incorrect correlations\. For instance, as shown in Figure[1](https://arxiv.org/html/2609.10177#acmlabel1), if the retrieved demonstrations share visually similar global features, the MLLM is prone to copying the most frequent textual labels from the context\([Fei et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib41)\)\. By doing so, the MLLM bypasses the deeper visual analysis required for the task, matching only the output format while failing to ground its response in fine\-grained visual evidence\.

This issue reflects a fundamental limitation of standard demonstrations: they mainly align the final output, while overlooking the reasoning path that leads to the desired response\([Tang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib7);[Long et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib8);[Wu et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib9)\)\. When an MLLM produces a response influenced by surface similarities, the key challenge is not merely to specify the target output, but to achieve reasoning path alignment by aligning the generated reasoning path with a latent evidence\-grounded refinement relation under the multimodal input\. In complex multimodal tasks, this requires explicit guidance on which visual evidence is truly relevant and how it should support revising the current response\([Sun et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib42);[Zhou et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib43);[Zhang et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib44)\)\. Standard demonstrations provide little such guidance, because they present only the target output rather than the reasoning path from a suboptimal response to a better one\. Therefore, achieving reasoning path alignment requires demonstrations that make this refinement path explicit, enabling the MLLM to refine its response in a way that is better grounded in the underlying multimodal evidence\.

To move beyond surface imitation, we propose COMIL, a novel framework that promotes reasoning path alignment in MLLMs through contrastive modeling of demonstrations and the self refinement capability of MLLMs\([Baldassini et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib24);[Doveh et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib25)\)\. Rather than providing only a single target output, COMIL reformulates each standard ICL example under the same multimodal input as a structured demonstration tuple consisting of a suboptimal response, a better response, and a generated reasoning path describing how the response should be refined and what fine\-grained visual evidence supports that refinement\. By presenting both responses under the same input condition, this formulation makes explicit not only which response is better, but also how the current response should be improved and why that improvement is supported by the underlying multimodal evidence, thereby turning the demonstration into explicit guidance for reasoning path alignment\.

Furthermore, because reasoning path alignment depends on the MLLM’s current response\([Guo et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib28);[Lee et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib29)\), we couple these contrastive demonstrations with a retrieval mechanism guided by the current response\. Rather than retrieving examples only from input similarity, COMIL selects demonstrations by additionally considering the similarity between their suboptimal responses and the MLLM’s current response\. In this way, the retrieved reasoning paths provide more relevant guidance for refining the current response\. Combined with a lightweight alignment controller that predicts response quality and determines whether further refinement is needed, COMIL moves beyond passive output matching and instead guides generation toward more evidence\-grounded refinement\.

Extensive experiments across a diverse set of representative MLLMs verify the effectiveness of our framework\. On three reasoning intensive tasks, COMIL shows clear improvements\. For example, on VQAv2, COMIL improves the accuracy of Qwen3\.5\-9B to 81\.9%, achieving the best result among the compared methods\. On Flickr30k, it improves the CIDEr score of Gemma\-3\-27B to 0\.587\. COMIL also remains sample efficient, maintaining favorable performance even with a limited number of retrieved demonstrations\. In addition, the framework generalizes beyond open\-source settings and achieves strong results on closed\-source MLLMs\. Overall, these results show that COMIL effectively improves multimodal ICL by reducing surface level imitation and promoting reasoning path alignment\.

In summary, our main contributions are as follows:

- •A New Framework for Reasoning Path Alignment in Multimodal ICL:We propose COMIL, a new multimodal in\-context learning framework that promotes reasoning path alignment by providing explicit guidance on which visual evidence is truly relevant and how it should support revising the current response\.
- •Contrastive Demonstration Modeling with Response\-Conditioned Retrieval and Alignment Control:COMIL reformulates each standard ICL example as a structured contrastive demonstration tuple consisting of a suboptimal response, a better response, and the corresponding reasoning path\. It further combines this formulation with response\-conditioned retrieval and a lightweight alignment controller to guide the MLLM toward the desired reasoning path during inference\.
- •Effective Results on Reasoning\-Intensive Multimodal Tasks:Extensive experiments across multiple MLLMs and multimodal datasets show that COMIL consistently improves performance, with particularly clear gains on reasoning intensive tasks\. Additional reasoning\-path evaluation and results on closed\-source MLLMs further demonstrate the effectiveness and generality of the proposed framework\.

## 2\.Related Work

### 2\.1\.Multimodal Large Language Models

Multimodal Large Language Models \(MLLMs\) have recently achieved remarkable progress on a wide range of tasks, including visual question answering, image captioning, visual reasoning, and multimodal generation\([Erfani and Mansouri, 2026](https://arxiv.org/html/2609.10177#bib.bib10);[Yao et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib11);[Chen et al\., 2026d](https://arxiv.org/html/2609.10177#bib.bib12);[Jin et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib13);[Kang et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib64);[Chen et al\., 2026a](https://arxiv.org/html/2609.10177#bib.bib67);[Chen et al\., 2026b](https://arxiv.org/html/2609.10177#bib.bib66);[Chen et al\., 2026c](https://arxiv.org/html/2609.10177#bib.bib65)\)\. By integrating visual encoders with large language models, together with large\-scale vision\-language pretraining, instruction tuning, and alignment, MLLMs have developed strong cross\-modal understanding and generation capabilities, making them a key foundation for multimodal intelligence\. Their ability to process visual and textual information within a unified generative framework enables them to perform diverse tasks with a single model and has significantly broadened the scope of vision\-language applications\. Despite this progress, effective inference\-time adaptation remains important for improving MLLM performance on complex multimodal tasks\([Xu et al\., 2025c](https://arxiv.org/html/2609.10177#bib.bib14);[Fan et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib15)\)\. In particular, how to better leverage the multimodal understanding and generation capabilities of MLLMs on complex tasks without parameter updates remains an important challenge\([Wu et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib16);[Song et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib17)\)\. This has motivated growing interest in inference\-time adaptation paradigms such as multimodal in\-context learning\.

### 2\.2\.Multimodal In\-Context Learning

Multimodal In\-Context Learning \(multimodal ICL\) has become an important paradigm for adapting MLLMs to downstream tasks\([Huang et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib20);[Li et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib1);[Zhou et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib18);[Li et al\., 2025a](https://arxiv.org/html/2609.10177#bib.bib19);[Wang et al\., 2026c](https://arxiv.org/html/2609.10177#bib.bib68);[Wang et al\., 2026a](https://arxiv.org/html/2609.10177#bib.bib69);[Wang et al\., 2026b](https://arxiv.org/html/2609.10177#bib.bib70)\)\. By leveraging multimodal demonstrations composed of images, texts, and corresponding responses, multimodal ICL enables task adaptation at inference time without parameter updates\([Kumar, 2026](https://arxiv.org/html/2609.10177#bib.bib22);[Gao et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib23)\)\. This property makes multimodal ICL particularly attractive in practice, as it can flexibly adapt a pretrained MLLM to diverse tasks while avoiding the cost of additional training\. More importantly, by presenting a small number of task\-specific examples in context, multimodal ICL allows the MLLM to exploit its pretrained multimodal knowledge in a task\-aware manner, making it a simple yet effective paradigm for multimodal classification, generation, and reasoning tasks\. Existing studies have mainly improved multimodal ICL from the input side, such as demonstration retrieval, demonstration organization, prompt design, and instruction construction, showing that better demonstrations can substantially affect downstream performance\([Huang et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib20);[Chen et al\., 2025b](https://arxiv.org/html/2609.10177#bib.bib21)\)\. Despite this progress, most existing methods still emphasize providing better demonstrations or improving alignment with the demonstrated target output\. However, for complex multimodal tasks, recent observations suggest that MLLMs may still rely on superficial imitation of demonstrated responses rather than follow the reasoning path that leads to the correct result\([Nguyen et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib26);[Huang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib6)\)\. This limitation motivates the need to go beyond better demonstration selection alone and to consider how demonstrations can more explicitly guide the MLLM’s reasoning\.

## 3\.Method

![Pipeline of COMIL showing the multimodal input, initial response generation, response-conditioned retrieval of contrastive demonstrations, iterative refinement, and quality-based stopping by the lightweight alignment controller.](https://arxiv.org/html/2609.10177v1/method_1.png)Figure 2\.Overall framework of COMIL\.Given an inputx=\(v,u\)x=\(v,u\), the MLLM first generates an initial response\. The retrieval dataset contains contrastive demonstrationsdi=\(xi,yi−,yi\+,ri\)d\_\{i\}=\(x\_\{i\},y\_\{i\}^\{\-\},y\_\{i\}^\{\+\},r\_\{i\}\)\. At steptt, the current responseyty\_\{t\}is used to retrieve the top\-kkrelevant demonstrations to generateyt\+1y\_\{t\+1\}\. The lightweight alignment controller predictsqt=𝒫⁡\(x,yt\)q\_\{t\}=\\mathcal\{P\}\(x,y\_\{t\}\)and compares it withτ\\tauto determine whether to continue refinement untilTTor return the final responseyt∗y\_\{t^\{\*\}\}\.Pipeline of COMIL showing the multimodal input, initial response generation, response\-conditioned retrieval of contrastive demonstrations, iterative refinement, and quality\-based stopping by the lightweight alignment controller\.### 3\.1\.Problem Setup

Let the multimodal input be denoted byx=\(v,u\)x=\(v,u\), wherevvis the visual input anduuis the textual instruction or query\. Givenxx, the target MLLM generates a textual responseyyaccording toy∼pθ​\(y∣x\)y\\sim p\_\{\\theta\}\(y\\mid x\)\. In this work, reasoning path alignment refers to aligning the generated reasoning path with a latent evidence\-grounded refinement relation underxx: the path should diagnose missed or misused multimodal evidence and support revising the current response toward a better one\. We define a latent alignment functionQ⁡\(x,y\)∈ℝQ\(x,y\)\\in\\mathbb\{R\}to conceptually score the resulting response state, where a larger value indicates better alignment with this desired refinement relation\. SinceQQis not directly observable, it serves as a conceptual objective, while its practical proxy is introduced in the lightweight alignment control module\. Under this formulation, the ideal target response can be written conceptually asy∗=arg⁡maxy⁡Q⁡\(x,y\)y^\{\*\}=\\arg\\max\_\{y\}Q\(x,y\)\.

To this end, COMIL is built on three coupled components: contrastive demonstration modeling, response\-conditioned retrieval, and lightweight alignment control\. Together, these components reformulate demonstrations to make explicit the reasoning paths from suboptimal responses to better responses, retrieve the demonstrations that are most relevant to the current response, and guide the MLLM toward reasoning path alignment during inference\.

### 3\.2\.Contrastive Demonstration Modeling

To support reasoning path alignment, COMIL reformulates each demonstration through contrastive modeling\([Baldassini et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib24);[Doveh et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib25)\)\. Rather than using a single target response, we explicitly show the MLLM how a suboptimal response is refined into a better response together with the corresponding reasoning path\. Formally, we represent theiith demonstration as

\(1\)di=\(xi,yi−,yi\+,ri\),d\_\{i\}=\\bigl\(x\_\{i\},y\_\{i\}^\{\-\},y\_\{i\}^\{\+\},r\_\{i\}\\bigr\),wherexix\_\{i\}is the multimodal input,yi−y\_\{i\}^\{\-\}is a suboptimal response,yi\+y\_\{i\}^\{\+\}is a better response, andrir\_\{i\}is a generated reasoning path that describes howyi−y\_\{i\}^\{\-\}can be refined intoyi\+y\_\{i\}^\{\+\}underxix\_\{i\}\. Here,rir\_\{i\}provides explicit refinement guidance rather than a direct observation of the MLLM’s internal reasoning process\.

Under the fixed multimodal inputxix\_\{i\},yi−y\_\{i\}^\{\-\}andyi\+y\_\{i\}^\{\+\}define a task\-supervised refinement transition, whilerir\_\{i\}specifies how the former can be revised toward the latter, thereby providing explicit guidance for reasoning path alignment\. In the contrastive modeling,yi−y\_\{i\}^\{\-\}andyi\+y\_\{i\}^\{\+\}form the contrasted response pair\. The reasoning pathrir\_\{i\}identifies the deficiency inyi−y\_\{i\}^\{\-\}, highlights the multimodal evidence relevant to the revision, and shows how these cues support the transition toyi\+y\_\{i\}^\{\+\}\. As a result, the contrastive demonstration provides explicit guidance not only on the preferred response, but also on how the response should be refined toward it\. This supports reasoning path alignment at inference time\.

### 3\.3\.Response\-Conditioned Retrieval and Refinement

#### Construction of the Contrastive Retrieval Dataset\.

To support reasoning path alignment at inference time, we first construct a retrieval dataset of contrastive demonstrations from a selected subset of training examples\. For each training instancexix\_\{i\}, the same target MLLM used later at inference time is first prompted to generate an initial response, which serves as the suboptimal responseyi−y\_\{i\}^\{\-\}\. The better responseyi\+y\_\{i\}^\{\+\}is taken directly from the ground truth response in the dataset under the same multimodal input\. We then feedxix\_\{i\},yi−y\_\{i\}^\{\-\}, andyi\+y\_\{i\}^\{\+\}into the same target MLLM and prompt it to generate a reasoning pathrir\_\{i\}, which describes howyi−y\_\{i\}^\{\-\}can be refined intoyi\+y\_\{i\}^\{\+\}\. In this way, each training instance is transformed into a contrastive demonstrationdi=\(xi,yi−,yi\+,ri\)d\_\{i\}=\(x\_\{i\},y\_\{i\}^\{\-\},y\_\{i\}^\{\+\},r\_\{i\}\), which explicitly represents a generated refinement path from a suboptimal response state to a better one\([Nguyen et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib26);[Kang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib27)\)\. Repeating this process over the selected training subset yields the retrieval dataset𝒞=\{di\}i=1N\\mathcal\{C\}=\\\{d\_\{i\}\\\}\_\{i=1\}^\{N\}used in subsequent retrieval and refinement\.

#### Response\-Conditioned Demonstration Retrieval\.

Given the retrieval dataset𝒞=\{di\}i=1N\\mathcal\{C\}=\\\{d\_\{i\}\\\}\_\{i=1\}^\{N\}, the goal of retrieval is to select contrastive demonstrations whose reasoning paths are relevant to refining the current response\. Since refinement relevance depends on both the multimodal input and the current response, retrieval is conditioned onxxand, whenever available,yty\_\{t\}\([Guo et al\., 2026](https://arxiv.org/html/2609.10177#bib.bib28);[Lee et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib29)\)\. At the initial step, no response has been generated yet\. We therefore retrieve demonstrations only according to the multimodal input:

\(2\)𝒟0=arg⁡max⁡∑di∈𝒟𝒟⊂𝒞,\|𝒟\|=k⁡λx​sx​\(x,xi\),y0=𝒢⁡\(x,𝒟0\),\\mathcal\{D\}\_\{0\}=\\arg\\max\_\{\\mathcal\{D\}\\subset\\mathcal\{C\},\\,\|\\mathcal\{D\}\|=k\}\\sum\_\{d\_\{i\}\\in\\mathcal\{D\}\}\\lambda\_\{x\}\\,s\_\{x\}\(x,x\_\{i\}\),\\,y\_\{0\}=\\mathcal\{G\}\(x,\\mathcal\{D\}\_\{0\}\),where𝒟0⊂𝒞\\mathcal\{D\}\_\{0\}\\subset\\mathcal\{C\}is the retrieved demonstration set,𝒢\\mathcal\{G\}denotes the MLLM with the retrieved contrastive demonstrations as context, andsx​\(x,xi\)s\_\{x\}\(x,x\_\{i\}\)measures the relevance between the current multimodal inputxxand the demonstration inputxix\_\{i\}\. Once the current responseyty\_\{t\}is available, retrieval is performed according to a joint relevance score over the input and the current response:

\(3\)𝒟t\+1\\displaystyle\\mathcal\{D\}\_\{t\+1\}=arg⁡max⁡∑di∈𝒟𝒟⊂𝒞,\|𝒟\|=k⁡\[λx​sx​\(x,xi\)\+λy​sy​\(yt,yi−\)\],\\displaystyle=\\arg\\max\_\{\\mathcal\{D\}\\subset\\mathcal\{C\},\\,\|\\mathcal\{D\}\|=k\}\\sum\_\{d\_\{i\}\\in\\mathcal\{D\}\}\\left\[\\lambda\_\{x\}\\,s\_\{x\}\(x,x\_\{i\}\)\+\\lambda\_\{y\}\\,s\_\{y\}\(y\_\{t\},y\_\{i\}^\{\-\}\)\\right\],yt\+1\\displaystyle y\_\{t\+1\}=𝒢⁡\(x,yt,𝒟t\+1\),\\displaystyle=\\mathcal\{G\}\(x,y\_\{t\},\\mathcal\{D\}\_\{t\+1\}\),wheresy​\(yt,yi−\)s\_\{y\}\(y\_\{t\},y\_\{i\}^\{\-\}\)measures the similarity between the current responseyty\_\{t\}and the suboptimal responseyi−y\_\{i\}^\{\-\}in the demonstration, andλx,λy≥0\\lambda\_\{x\},\\lambda\_\{y\}\\geq 0balance the contributions of input relevance and response similarity\. We use response similarity as a practical proxy for refinement relevance\. Whenyty\_\{t\}is similar toyi−y\_\{i\}^\{\-\}, the associated reasoning pathrir\_\{i\}is more likely to describe a correction pattern relevant to the current response\. The retrieved demonstration therefore provides not only a better response but also potentially useful guidance for revision\. Response similarity does not guarantee the same error mode, but enables retrieval to incorporate the current response state in addition to input similarity\.

#### Refinement with Retrieved Demonstrations\.

Under this retrieval mechanism, iterative refinement aims to move the current response toward stronger reasoning path alignment, rather than guaranteeing monotonic improvement at every step\([Xu et al\., 2025a](https://arxiv.org/html/2609.10177#bib.bib30);[Chen et al\., 2025a](https://arxiv.org/html/2609.10177#bib.bib31);[Ke et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib32)\)\. In practice, the target MLLM uses the retrieved contrastive demonstrations to revise the current response according to refinement patterns relevant to its current state\. In this way, retrieval and refinement are tightly coupled: retrieval identifies relevant refinement guidance, while the MLLM uses this guidance to revise the current response toward the evidence\-grounded refinement relation defined in Section[3\.1](https://arxiv.org/html/2609.10177#S3.SS1)\.

![Training pipeline in which the lightweight predictor receives the multimodal input and a candidate response, predicts response quality, and is optimized with a regression loss against the observed task metric.](https://arxiv.org/html/2609.10177v1/method_2.png)Figure 3\.Training pipeline of the lightweight quality predictor\.The predictor takes an input and a candidate response as input, predicts the response evaluation metric, and is trained with a regression loss against the actual metricM⁡\(x,y\)M\(x,y\)\.Training pipeline in which the lightweight predictor receives the multimodal input and a candidate response, predicts response quality, and is optimized with a regression loss against the observed task metric\.

### 3\.4\.Lightweight Alignment Control

While reasoning path alignment is the target of refinement, it is not directly observable at inference time\. We therefore introduce a lightweight alignment controller that predicts response quality as a proxy for controlling the refinement trajectory\. During controller training, the task\-specific evaluation metric provides an observable supervision signal: under a fixed multimodal inputxx, a response that better follows the desired evidence\-grounded refinement relation is more likely to match the expected output and achieve a better task metric\([Madaan et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib33);[Paul et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib34)\)\. We therefore use the response\-level evaluation metric to supervise the controller\.

Formally, letM⁡\(x,y\)M\(x,y\)denote the actual evaluation metric of responseyyunder inputxx, where a larger value indicates better task performance\. As defined in Section[3\.1](https://arxiv.org/html/2609.10177#S3.SS1),Q⁡\(x,y\)Q\(x,y\)is a latent alignment property and cannot be directly observed\. We therefore useM⁡\(x,y\)M\(x,y\)as a practical supervision signal and train a lightweight predictor𝒫\\mathcal\{P\}to estimate the quality of each intermediate response,qt=𝒫⁡\(x,yt\)q\_\{t\}=\\mathcal\{P\}\(x,y\_\{t\}\), whereqtq\_\{t\}is the predicted quality score at refinement steptt\. The predictor is trained with supervision derived from the actual metricM⁡\(x,yt\)M\(x,y\_\{t\}\); thus,qtq\_\{t\}predicts response quality rather than directly measuring the latent alignment functionQQ\. For ease of exposition, we refer to this predictor as the*lightweight alignment controller*\. The relation between task performance and reasoning\-path alignment is further examined in Section[5\.1](https://arxiv.org/html/2609.10177#S5.SS1)\. This alignment control is helpful because iterative refinement does not always improve the response monotonically\. Although the retrieved demonstrations are selected to support reasoning path alignment, later refinement steps may still preserve, weaken, or distort the desired refinement direction\([Huang et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib35)\)\. The controller therefore evaluates the predicted quality of each intermediate response and prevents the final decision from depending only on the latest step\. In this way, it helps retain a higher\-quality response among the explored candidates\.

Based on the predicted quality score, we introduce an acceptance thresholdτ\\taufor early stopping\. In practice,τ\\tauis determined from the training data used for the alignment controller: we setτ\\tauto the evaluation metric corresponding to the top 25% of the training responses, which are ranked by the metric in descending order\. This choice encourages early stopping on high\-quality responses while avoiding an overly strict threshold that would make successful stopping difficult\. Ifqt≥τq\_\{t\}\\geq\\tau, the current response is regarded as sufficiently high\-quality and refinement stops\. Otherwise, refinement continues until the maximum refinement budgetTTis reached\. To unify early stopping and final selection, we define

\(4\)t∗=\{min⁡\{t∣qt≥τ\},if∃t∈\{0,…,T\},qt≥τ,arg⁡maxt∈\{0,…,T\}⁡qt,otherwise,t^\{\*\}=\\begin\{cases\}\\min\\\{t\\mid q\_\{t\}\\geq\\tau\\\},&\\text\{if \}\\exists\\,t\\in\\\{0,\\ldots,T\\\},\\,q\_\{t\}\\geq\\tau,\\\\ \\arg\\max\_\{t\\in\\\{0,\\ldots,T\\\}\}q\_\{t\},&\\text\{otherwise\},\\end\{cases\}and takeyt∗y\_\{t^\{\*\}\}as the final response\.

Through this design, the controller neither generates reasoning paths nor directly measures reasoning\-path alignment\. Instead, it uses the predicted response quality as a practical proxy to determine when refinement should stop or which intermediate response should be selected\. Together with contrastive demonstration modeling and response\-conditioned retrieval, it stabilizes the inference\-time refinement process toward reasoning path alignment\.

## 4\.Experiment

Table 1\.Main results on four multimodal datasets and four open\-source MLLMs, including recent multimodal ICL baselines\. Best and second\-best results are highlighted in bold and underlined, respectively; “–” denotes unavailable results\.### 4\.1\.Experimental Setup

#### MLLMs and Datasets\.

We evaluate COMIL on four representative open\-source multimodal large language models \(MLLMs\), including Gemma\-3\-27B\([Team, 2025](https://arxiv.org/html/2609.10177#bib.bib45)\), InternVL3\.5\-14B\([Wang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib46)\), Qwen3\.5\-9B\([Qwen Team, 2026](https://arxiv.org/html/2609.10177#bib.bib47)\), and Qwen3\.5\-4B\([Qwen Team, 2026](https://arxiv.org/html/2609.10177#bib.bib47)\)\. Following the main experimental setting, we conduct experiments on four datasets: CIFAR10\([Krizhevsky et al\., 2009](https://arxiv.org/html/2609.10177#bib.bib48)\)for image classification, Flickr30k\([Young et al\., 2014](https://arxiv.org/html/2609.10177#bib.bib49)\)for image captioning, and VQAv2\([Goyal et al\., 2017](https://arxiv.org/html/2609.10177#bib.bib51)\)and OKVQA\([Schwenk et al\., 2022](https://arxiv.org/html/2609.10177#bib.bib50)\)for visual question answering\.

#### Tasks and Metrics\.

We evaluate COMIL on three types of multimodal tasks with task\-specific metrics: image classification, image captioning, and visual question answering\. These tasks cover multimodal classification, generation, and reasoning, enabling evaluation across diverse multimodal settings\. For image classification, we report Accuracy\. For image captioning, we report CIDEr\([Vedantam et al\., 2015](https://arxiv.org/html/2609.10177#bib.bib52)\)\. For visual question answering, we report VQA Accuracy\([Antol et al\., 2015](https://arxiv.org/html/2609.10177#bib.bib53)\)\. For all metrics, higher values indicate better performance\.

#### Baselines\.

We compare COMIL with representative baselines from five categories, as shown in Table[1](https://arxiv.org/html/2609.10177#S4.T1)\. Specifically, we includeRetrieval\-based ICLmethods, including CLIPRE\([Radford et al\., 2021](https://arxiv.org/html/2609.10177#bib.bib63)\), KNN\([Guo et al\., 2003](https://arxiv.org/html/2609.10177#bib.bib54)\), and CR\([Liu and Croft, 2004](https://arxiv.org/html/2609.10177#bib.bib55)\);Train\-based ICLmethods, including LCL\([Tai et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib56)\)and MimIC\([Jiang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib57)\);CoTmethods, including Few\-shot CoT\([Kim et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib58)\)and Self\-Consistency CoT\([Wang et al\., 2022](https://arxiv.org/html/2609.10177#bib.bib59)\);Self\-Refinemethods, including Iteration, Self\-Refine\([Madaan et al\., 2023](https://arxiv.org/html/2609.10177#bib.bib33)\), and SC\-Captioner\([Zhang et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib60)\); and recentMultimodal ICLmethods, including AIM\([Gao et al\., 2025](https://arxiv.org/html/2609.10177#bib.bib23)\), TACO\([Li et al\., 2025b](https://arxiv.org/html/2609.10177#bib.bib4)\), and M2IV\([Li et al\., 2025a](https://arxiv.org/html/2609.10177#bib.bib19)\)\. These baselines cover retrieval\-based demonstration selection, training\-enhanced in\-context learning, reasoning\-oriented prompting, iterative refinement, and recent multimodal ICL approaches\.

#### Inference Details\.

For each test instance, COMIL generates an initial response and performs up to 3 refinement iterations using top\-k=10k=10retrieved contrastive demonstrations\. In our implementation, the lightweight alignment controller uses openai/clip\-vit\-base\-patch32\([Radford et al\., 2021](https://arxiv.org/html/2609.10177#bib.bib63)\)as the image encoder and google\-bert/bert\-base\-uncased\([Devlin et al\., 2019](https://arxiv.org/html/2609.10177#bib.bib62)\)as the text encoder\. For response\-conditioned retrieval, both input and response similarities are computed using cosine similarity, with the two retrieval weights set to 0\.5\. The initial response and each refinement round use a maximum generation length of 2048 tokens\. Both controller encoders are frozen, and their representations are fused and scored by a three\-layer MLP with a hidden dimension of 512 and dropout of 0\.1\. For controller training, we use part of the training split of each dataset\. The retrieval set is constructed from the last 500 samples of the corresponding training split and is kept disjoint from the controller training set\. The main experiments are conducted on the test split of each dataset\. Therefore, the controller training set, retrieval set, and evaluation set are mutually exclusive, and no data instance is reused across them\. We use the same experimental protocol across different target MLLMs and datasets, and compare all methods under consistent task settings for fair evaluation\.

### 4\.2\.Main Results

Table[1](https://arxiv.org/html/2609.10177#S4.T1)reports the main results across four multimodal datasets and four open\-source MLLMs, including comparisons with recent multimodal ICL methods\. Overall, COMIL achieves the best performance in 13 out of 16 MLLM–dataset settings and ranks second in the remaining three, demonstrating strong and consistent performance across multimodal classification, captioning, and visual question answering\. On CIFAR\-10, COMIL achieves the best result on three of the four MLLMs and ranks second on Qwen3\.5\-4B\. Its advantages are more consistent on multimodal generation and reasoning tasks\. On Flickr30k, COMIL achieves the best performance under all four MLLMs; for example, it reaches 0\.517 CIDEr on InternVL3\.5\-14B, outperforming the strongest baseline AIM by 0\.086\. On VQAv2, COMIL achieves the best result on three MLLMs and ranks second on InternVL3\.5\-14B\. In particular, it reaches 81\.9% accuracy on Qwen3\.5\-9B, exceeding TACO by 1\.0 percentage point\. On OKVQA, COMIL again achieves the best performance in three of the four settings and ranks second on Qwen3\.5\-4B\. These results show that COMIL remains competitive not only against conventional retrieval, CoT, and self\-refinement baselines, but also against recent multimodal ICL methods\. Another notable finding is that simple Self\-Refine based baselines perform poorly across most multimodal tasks\. This suggests that the gains of COMIL do not come from iterative refinement alone, but from the combination of contrastive demonstration modeling, response\-conditioned retrieval, and lightweight alignment control\.

### 4\.3\.Ablation Study

Table 2\.Ablation study of contrastive demonstration modeling in COMIL\. As more components of the contrastive four\-tuple are introduced, performance improves consistently across datasets and target MLLMs\.Table 3\.Ablation study of response\-conditioned retrieval in COMIL\. Response\-conditioned retrieval consistently outperforms random retrieval across datasets and target MLLMs\.Table 4\.Ablation study of the lightweight alignment control module in COMIL\. The alignment controller consistently improves final performance\.We further analyze how each component contributes to reasoning path alignment during inference\. Following the design of COMIL, we study three aspects: contrastive demonstration modeling, response\-conditioned retrieval, and lightweight alignment control\.

#### Effect of Contrastive Demonstration Modeling

Table[2](https://arxiv.org/html/2609.10177#S4.T2)evaluates the effect of the contrastive four\-tuple formulation\. When the demonstrations contain only the multimodal input\(vi,ui\)\(v\_\{i\},u\_\{i\}\)and the better responseyi\+y\_\{i\}^\{\+\}, performance drops substantially on both datasets and both target MLLMs\. Adding the suboptimal responseyi−y\_\{i\}^\{\-\}but removing the reasoning pathrir\_\{i\}results in a less severe performance drop\. A standard positive\-only demonstration mainly shows the MLLM a preferred response, but does not specify how the current response should be refined under the same input condition\. In contrast, introducingyi−y\_\{i\}^\{\-\}makes the contrast between suboptimal and better responses explicit, while addingrir\_\{i\}further clarifies how the response should be refined and what multimodal evidence supports that refinement\. This enables the demonstrations to provide more informative guidance for reasoning path alignment\.

#### Effect of Response\-Conditioned Retrieval

Table[3](https://arxiv.org/html/2609.10177#S4.T3)studies the retrieval strategy\. Replacing response\-conditioned retrieval with random retrieval consistently degrades performance across both tasks and both MLLMs\. This result shows that the effectiveness of retrieval in COMIL depends not simply on the availability of demonstrations, but on whether the retrieved demonstrations provide reasoning paths relevant to the current response under the same input condition\. By retrieving such demonstrations, response\-conditioned retrieval offers more effective guidance for refining the current response toward the desired reasoning path\.

#### Effect of Lightweight Alignment Control

Table[4](https://arxiv.org/html/2609.10177#S4.T4)examines the lightweight alignment controller\. Removing alignment control causes a consistent but smaller performance drop than removing the contrastive formulation or the retrieval mechanism\. This shows that alignment control is not the main source of gains, but it still contributes to the final performance\. Its role is better understood as stabilizing refinement rather than providing reasoning guidance by itself\. Contrastive demonstration modeling and response\-conditioned retrieval determine how the current response should be refined, while the controller predicts the quality of each intermediate response and determines when refinement should stop or which response should be selected\. In this way, the controller helps avoid unnecessary or harmful refinement steps and makes the overall refinement process more reliable across datasets and MLLMs\.

## 5\.Discussion

### 5\.1\.Direct Evaluation of Reasoning Path Alignment

Final task metrics do not directly measure whether the refinement paths are grounded in query\-relevant multimodal evidence\. We therefore use GPT\-5\.5 to evaluate each reasoning path along four 0–4 criteria: evidence mention, evidence correctness, task relevance, and response–evidence consistency\. The aggregated score is normalized to\[0,1\]\[0,1\]as the reasoning path alignment \(RPA\) score\. We also report the sample\-level Spearman correlationρ\\rhobetween RPA and the task metric\. As shown in Table[5](https://arxiv.org/html/2609.10177#S5.T5), COMIL achieves the highest RPA score in five of the six settings\. Moreover, its Spearmanρ\\rhoremains consistently positive, ranging from 0\.55 to 0\.68 with an average of 0\.63\. These results indicate that the task\-performance gains of COMIL are generally accompanied by refinement paths that are better grounded in multimodal evidence under the adopted evaluation protocol\. We note that RPA evaluates the generated refinement path and should not be interpreted as a direct measure of the model’s unobservable internal reasoning process\.

Table 5\.Direct reasoning\-path alignment evaluation\. RPA denotes the normalized reasoning\-path alignment score, andρ\\rhodenotes the sample\-level Spearman correlation with the task metric\.
### 5\.2\.Case Study

As shown in Figure[4](https://arxiv.org/html/2609.10177#acmlabel4), response\-conditioned retrieval can identify demonstrations with relevant refinement patterns\. In one VQA example, the query asks for the number of water bottles, for which the initial response overcounts 7 bottles as 8\. The retrieved demonstration asks for the number of glasses, but exhibits the same error: it predicts 8 instead of 7 because of rough counting\. Its reasoning path corrects this error through instance\-by\-instance enumeration, which similarly guides the query response from 8 to 7\. This case illustrates that response\-conditioned retrieval can retrieve transferable correction patterns even when the object semantics differ\. Nevertheless, response similarity serves as a proxy for refinement relevance rather than a guarantee of the same error mode\.

![A VQA case study where both the query and the retrieved demonstration overcount by one; instance-by-instance enumeration provides a transferable correction pattern that changes the response from 8 to 7.](https://arxiv.org/html/2609.10177v1/case_study2.png)Figure 4\.Case study of response\-conditioned retrieval\. Although the query and retrieved demonstration involve different objects, both exhibit rough\-counting overestimation that is corrected by instance\-by\-instance enumeration\.A VQA case study where both the query and the retrieved demonstration overcount by one; instance\-by\-instance enumeration provides a transferable correction pattern that changes the response from 8 to 7\.Figure 5\.Effect of retrieval set size on COMIL\. Performance generally improves as the retrieval set grows across Flickr30k, VQAv2, and OKVQA, with diminishing gains at larger scales\.Performance curves for Flickr30k, VQAv2, and OKVQA as the retrieval dataset size increases, showing generally improving performance with diminishing gains at larger retrieval sizes\.
### 5\.3\.Effect of Retrieval Dataset Size

We further study how the size of the contrastive retrieval dataset affects COMIL\. Specifically, we vary the retrieval dataset size from 20 to 500 and evaluate COMIL on Flickr30k, VQAv2, and OKVQA under three target MLLMs\. As shown in Figure[5](https://arxiv.org/html/2609.10177#acmlabel5), increasing the retrieval dataset size generally improves performance across tasks and MLLMs, suggesting that broader retrieval coverage increases the likelihood of finding demonstrations relevant to the current response\. The gains, however, gradually diminish as the retrieval dataset grows\. In several settings, performance is already close to its best value at around 300 examples, while further enlargement yields only marginal improvements or small fluctuations\. These results indicate that COMIL benefits from richer retrieval coverage but does not require a very large retrieval dataset to achieve strong performance\.

Table 6\.Results of COMIL on two closed\-source MLLMs\. COMIL achieves the best performance on CIFAR10 under both Claude Sonnet 4\.6 and GPT\-4o\.
### 5\.4\.Extension to Closed\-Source MLLMs

We further evaluate our method on closed\-source MLLMs to examine whether its effectiveness extends beyond open\-source settings\. As shown in Table[6](https://arxiv.org/html/2609.10177#S5.T6), our method consistently achieves the best performance on CIFAR10 under both Claude Sonnet 4\.6 and GPT\-4o\([Hurst et al\., 2024](https://arxiv.org/html/2609.10177#bib.bib61)\)\. Specifically, it improves the accuracy to 95\.3% on Claude Sonnet 4\.6 and 99\.8% on GPT\-4o, outperforming all retrieval\-based baselines in both cases\. These results suggest that the benefit of our framework does not depend on a specific open\-source architecture, but transfers well to stronger closed\-source MLLMs, further supporting the general applicability of the proposed framework\.

### 5\.5\.Cost Considerations

COMIL introduces additional inference cost due to iterative refinement and is therefore not the lowest\-cost ICL method\. Nevertheless, compared with heavier reasoning and refinement baselines, COMIL reduces average latency and token consumption by 51\.6% and 82\.7%, respectively, over SC\-CoT, and is 29\.2% faster than Self\-Refine\. The target MLLM remains frozen, with additional training confined to the lightweight controller\.

## 6\.Conclusion

In this paper, we present COMIL, a multimodal in\-context learning framework that moves beyond surface\-level imitation by promoting reasoning path alignment during inference\. COMIL reformulates demonstrations as contrastive tuples that describe how a suboptimal response can be refined into a better response, combines them with response\-conditioned retrieval, and introduces a lightweight alignment controller to guide refinement\. In this way, COMIL enables the MLLM not only to observe better responses in context, but also to refine the current response toward better alignment with the evidence\-grounded refinement relation\. Extensive experiments on multimodal tasks show that COMIL consistently improves performance, with particularly clear gains on reasoning\-intensive tasks\. Further reasoning\-path analyses and results on closed\-source MLLMs support the effectiveness and generality of the proposed framework across different model settings\.

###### Acknowledgements\.

This work was supported by the National Natural Science Foundation of China under Grant No\. 62502550 and the Shenzhen Science and Technology Program under Grant No\. KJZD20240903095700001\.

## References

- Antolet al\.\(2015\)S\. Antol, A\. Agrawal, J\. Lu, M\. Mitchell, D\. Batra, C\. L\. Zitnick, and D\. ParikhVqa: visual question answering\.InProceedings of the IEEE international conference on computer vision,pp\. 2425–2433\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px2.p1.1)\.
- Baldassiniet al\.\(2024\)F\. B\. Baldassini, M\. Shukor, M\. Cord, L\. Soulier, and B\. PiwowarskiWhat makes multimodal in\-context learning work?\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 1539–1550\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p4.1),[§3\.2](https://arxiv.org/html/2609.10177#S3.SS2.p1.1)\.
- Chenet al\.\(2026a\)P\. Chen, C\. Huang, Y\. Cao, C\. Liu, W\. Wang, W\. Wang, M\. Yang, L\. Shen, W\. Ren, and X\. CaoTowards explainable industrial anomaly detection via knowledge\-guided latent reasoning\.arXiv preprint arXiv:2602\.09850\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Chenet al\.\(2026b\)P\. Chen, F\. Huang, and C\. HuangDyC\-clip: dynamic context\-aware multi\-modal prompt learning for zero\-shot anomaly detection\.Pattern Recognition,pp\. 113215\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Chenet al\.\(2025a\)Q\. Chen, L\. Qin, J\. Liu, D\. Peng, J\. Guan, P\. Wang, M\. Hu, Y\. Zhou, T\. Gao, and W\. CheTowards reasoning era: a survey of long chain\-of\-thought for reasoning large language models\.arXiv preprint arXiv:2503\.09567\.Cited by:[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px3.p1.1)\.
- Chenet al\.\(2025b\)S\. Chen, Z\. Han, B\. He, J\. Liu, M\. Buckley, Y\. Qin, P\. Torr, V\. Tresp, and J\. GuCan multimodal large language models truly perform multimodal in\-context learning?\.In2025 IEEE/CVF Winter Conference on Applications of Computer Vision \(WACV\),pp\. 6000–6010\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Chenet al\.\(2026c\)Y\. Chen, W\. Wang, R\. Chen, J\. Wang, M\. Yang, Y\. Wang, W\. Wang, and X\. CaoDomain adaptive object detection via dual\-stream bilevel\-cycle optimization\.arXiv preprint arXiv:2606\.31373\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Chenet al\.\(2026d\)Y\. Chen, Y\. Ge, Y\. Ge, M\. Ding, B\. Li, R\. Wang, R\. Xu, Y\. Shan, and X\. LiuEgoplan\-bench: benchmarking multimodal large language models for human\-level planning\.International Journal of Computer Vision134\(3\),pp\. 118\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Devlinet al\.\(2019\)J\. Devlin, M\. Chang, K\. Lee, and K\. ToutanovaBert: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 \(long and short papers\),pp\. 4171–4186\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px4.p1.1)\.
- Dovehet al\.\(2024\)S\. Doveh, S\. Perek, M\. J\. Mirza, W\. Lin, A\. Alfassy, A\. Arbelle, S\. Ullman, and L\. KarlinskyTowards multimodal in\-context learning for vision and language models\.InEuropean Conference on Computer Vision,pp\. 250–267\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p4.1),[§3\.2](https://arxiv.org/html/2609.10177#S3.SS2.p1.1)\.
- Erfani and Mansouri \(2026\)A\. Erfani and A\. MansouriApplications of multimodal large language models in construction industry\.Advanced Engineering Informatics69,pp\. 103909\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Fanet al\.\(2025\)X\. Fan, X\. Chen, L\. Yang, C\. H\. Yap, R\. Qureshi, Q\. Dou, M\. H\. Yap, and M\. ShahTest\-time retrieval\-augmented adaptation for vision\-language models\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 8810–8819\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Feiet al\.\(2023\)Y\. Fei, Y\. Hou, Z\. Chen, and A\. BosselutMitigating label biases for in\-context learning\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 14014–14031\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p2.1)\.
- Gaoet al\.\(2025\)J\. Gao, Q\. Qiao, T\. Wu, Z\. Wang, Z\. Cao, and W\. LiAim: let any multimodal large language models embrace efficient in\-context learning\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 3077–3085\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Goyalet al\.\(2017\)Y\. Goyal, T\. Khot, D\. Summers\-Stay, D\. Batra, and D\. ParikhMaking the v in vqa matter: elevating the role of image understanding in visual question answering\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\. 6904–6913\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Guoet al\.\(2003\)G\. Guo, H\. Wang, D\. Bell, Y\. Bi, and K\. GreerKNN model\-based approach in classification\.InOTM Confederated International Conferences" On the Move to Meaningful Internet Systems",pp\. 986–996\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Guoet al\.\(2026\)Y\. Guo, H\. Tan, R\. Li, X\. Li, X\. Sun, P\. Qiang, and H\. ZhangProblem decomposition guided by reasoning utility for complex reasoning in llms\.Information Processing & Management63\(3\),pp\. 104509\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p5.1),[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px2.p1.1)\.
- Huanget al\.\(2025\)C\. Huang, Y\. Zhu, S\. Zhu, J\. Xiao, M\. Andrade, S\. Chopra, and Z\. KiraMimicking or reasoning: rethinking multi\-modal in\-context learning in vision\-language models\.arXiv preprint arXiv:2506\.07936\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p2.1),[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Huanget al\.\(2023\)Q\. Huang, L\. Shen, R\. Zhang, S\. Ding, B\. Wang, Z\. Zhou, and Y\. WangCrossgnn: confronting noisy multivariate time series via cross interaction refinement\.Advances in Neural Information Processing Systems36,pp\. 46885–46902\.Cited by:[§3\.4](https://arxiv.org/html/2609.10177#S3.SS4.p2.1)\.
- Huanget al\.\(2026\)Y\. Huang, K\. Roth, Q\. Bouniot, W\. Xu, and Z\. AkataDissecting multimodal in\-context learning: modality asymmetries and circuit dynamics in modern transformers\.arXiv preprint arXiv:2601\.20796\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Hurstet al\.\(2024\)A\. Hurst, A\. Lerer, A\. P\. Goucher, A\. Perelman, A\. Ramesh, A\. Clark, A\. Ostrow, A\. Welihinda, A\. Hayes, A\. Radford,et al\.Gpt\-4o system card\.arXiv preprint arXiv:2410\.21276\.Cited by:[§5\.4](https://arxiv.org/html/2609.10177#S5.SS4.p1.1)\.
- Jianget al\.\(2025\)Y\. Jiang, J\. Fu, C\. Hao, X\. Hu, Y\. Peng, X\. Geng, and X\. YangMimic in\-context learning for multimodal tasks\.InProceedings of the Computer Vision and Pattern Recognition Conference,pp\. 29825–29835\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Jinet al\.\(2025\)Y\. Jin, J\. Li, T\. Gu, Y\. Liu, B\. Zhao, J\. Lai, Z\. Gan, Y\. Wang, C\. Wang, X\. Tan,et al\.Efficient multimodal large language models: a survey\.Visual Intelligence3\(1\),pp\. 27\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Kanget al\.\(2025\)C\. Kang, X\. Liu, and F\. GuoRetrointext: a multimodal large language model enhanced framework for retrosynthetic planning via in\-context representation learning\.InThe Thirteenth International Conference on Learning Representations,Cited by:[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px1.p1.1)\.
- Kanget al\.\(2026\)Z\. Kang, J\. Gong, J\. Yan, W\. Xia, Y\. Wang, Z\. Cheng, W\. Cao, Z\. Wang, Z\. Feng, H\. Ding,et al\.Hssbench: benchmarking humanities and social sciences ability for multimodal large language models\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 74664–74719\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Keet al\.\(2025\)Z\. Ke, F\. Jiao, Y\. Ming, X\. Nguyen, A\. Xu, D\. X\. Long, M\. Li, C\. Qin, P\. Wang, S\. Savarese,et al\.A survey of frontiers in llm reasoning: inference scaling, learning to reason, and agentic systems\.arXiv preprint arXiv:2504\.09037\.Cited by:[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px3.p1.1)\.
- Kimet al\.\(2023\)S\. Kim, S\. Joo, D\. Kim, J\. Jang, S\. Ye, J\. Shin, and M\. SeoThe cot collection: improving zero\-shot and few\-shot learning of language models via chain\-of\-thought fine\-tuning\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 12685–12708\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Krizhevskyet al\.\(2009\)A\. Krizhevsky G\. Hintonet al\.Learning multiple layers of features from tiny images\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Kumar \(2026\)V\. KumarAct\-observe\-rewrite: multimodal coding agents as in\-context policy learners for robot manipulation\.arXiv preprint arXiv:2603\.04466\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Leeet al\.\(2025\)H\. Lee, S\. Oh, J\. Kim, J\. Shin, and J\. TackRevise: learning to refine at test\-time via intrinsic self\-verification\.arXiv preprint arXiv:2502\.14565\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p5.1),[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px2.p1.1)\.
- Liet al\.\(2025a\)Y\. Li, Y\. Cao, H\. He, Q\. Cheng, X\. Fu, X\. Xiao, T\. Wang, and R\. TangM2\{\}^\{2\}IV: towards efficient and fine\-grained multimodal in\-context learning via representation engineering\.arXiv preprint arXiv:2504\.04633\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Liet al\.\(2026\)Y\. Li, J\. Yang, Z\. Shen, L\. Han, H\. Xu, and R\. TangCatp: contextually adaptive token pruning for efficient and enhanced multimodal in\-context learning\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 6619–6627\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1),[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Liet al\.\(2025b\)Y\. Li, J\. Yang, T\. Yun, P\. Feng, J\. Huang, and R\. TangTaco: enhancing multimodal in\-context learning via task mapping\-guided sequence configuration\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 736–763\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1),[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Liaoet al\.\(2026\)W\. Liao, J\. Yu, Y\. Wang, Q\. Jiang, and X\. YangEnhancing visual in\-context learning by multi\-faceted fusion\.arXiv preprint arXiv:2601\.10107\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1)\.
- Liu and Croft \(2004\)X\. Liu and W\. B\. CroftCluster\-based retrieval using language models\.InProceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval,pp\. 186–193\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Longet al\.\(2024\)Q\. Long, Y\. Wu, W\. Wang, and S\. J\. PanDoes in\-context learning really learn? rethinking how large language models respond and solve tasks via in\-context learning\.arXiv preprint arXiv:2404\.07546\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.
- Luoet al\.\(2024\)M\. Luo, X\. Xu, Y\. Liu, P\. Pasupat, and M\. KazemiIn\-context learning with retrieved demonstrations for language models: a survey\.arXiv preprint arXiv:2401\.11624\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p2.1)\.
- Madaanet al\.\(2023\)A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang,et al\.Self\-refine: iterative refinement with self\-feedback\.Advances in neural information processing systems36,pp\. 46534–46594\.Cited by:[§3\.4](https://arxiv.org/html/2609.10177#S3.SS4.p1.1),[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Nguyenet al\.\(2026\)T\. Nguyen, W\. Yuan, S\. Wei, H\. Li, D\. Seita, and Y\. WangICLR: in\-context imitation learning with visual reasoning\.arXiv preprint arXiv:2603\.07530\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1),[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px1.p1.1)\.
- Paulet al\.\(2024\)D\. Paul, M\. Ismayilzada, M\. Peyrard, B\. Borges, A\. Bosselut, R\. West, and B\. FaltingsRefiner: reasoning feedback on intermediate representations\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 1100–1126\.Cited by:[§3\.4](https://arxiv.org/html/2609.10177#S3.SS4.p1.1)\.
- Qwen Team \(2026\)Qwen TeamQwen3\.5: towards native multimodal agents\.External Links:[Link](https://qwen.ai/blog?id=qwen3.5)Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Radfordet al\.\(2021\)A\. Radford, J\. W\. Kim, C\. Hallacy, A\. Ramesh, G\. Goh, S\. Agarwal, G\. Sastry, A\. Askell, P\. Mishkin, J\. Clark,et al\.Learning transferable visual models from natural language supervision\.InInternational conference on machine learning,pp\. 8748–8763\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1),[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px4.p1.1)\.
- Schwenket al\.\(2022\)D\. Schwenk, A\. Khandelwal, C\. Clark, K\. Marino, and R\. MottaghiA\-okvqa: a benchmark for visual question answering using world knowledge\.InEuropean conference on computer vision,pp\. 146–162\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Songet al\.\(2025\)S\. Song, X\. Li, S\. Li, S\. Zhao, J\. Yu, J\. Ma, X\. Mao, W\. Zhang, and M\. WangHow to bridge the gap between modalities: survey on multimodal large language model\.IEEE Transactions on Knowledge and Data Engineering37\(9\),pp\. 5311–5329\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Sunet al\.\(2025\)Y\. Sun, Q\. Chen, J\. Wang, J\. Wang, and Z\. LiExploring effective factors for improving visual in\-context learning\.IEEE Transactions on Image Processing\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.
- Taiet al\.\(2024\)Y\. Tai, W\. Fan, Z\. Zhang, and Z\. LiuLink\-context learning for multimodal llms\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 27176–27185\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Tanget al\.\(2025\)X\. Tang, X\. Wang, W\. X\. Zhao, and J\. WenDawn\-icl: strategic planning of problem\-solving trajectories for zero\-shot in\-context learning\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 1918–1934\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.
- Team \(2025\)G\. TeamGemma 3\.External Links:[Link](https://goo.gle/Gemma3Report)Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Vedantamet al\.\(2015\)R\. Vedantam, C\. Lawrence Zitnick, and D\. ParikhCider: consensus\-based image description evaluation\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\. 4566–4575\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px2.p1.1)\.
- Wanget al\.\(2024a\)J\. Wang, C\. Wang, C\. Tan, J\. Huang, and M\. GaoKnowledgeable in\-context tuning: exploring and exploiting factual knowledge for in\-context learning\.InFindings of the Association for Computational Linguistics: NAACL 2024,pp\. 3261–3280\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1)\.
- Wanget al\.\(2024b\)S\. Wang, Z\. Chen, C\. Shi, C\. Shen, and J\. LiMixture of demonstrations for in\-context learning\.Advances in Neural Information Processing Systems37,pp\. 88091–88116\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p2.1)\.
- Wanget al\.\(2025\)W\. Wang, Z\. Gao, L\. Gu, H\. Pu, L\. Cui, X\. Wei, Z\. Liu, L\. Jing, S\. Ye, J\. Shao,et al\.InternVL3\.5: advancing open\-source multimodal models in versatility, reasoning, and efficiency\.arXiv preprint arXiv:2508\.18265\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Wanget al\.\(2026a\)W\. Wang, P\. Chen, Y\. Xiao, Y\. Zhang, X\. Lu, J\. Huang, and X\. CaoTask\-related in\-context learning\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 43378–43400\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Wanget al\.\(2026b\)W\. Wang, X\. Yan, X\. Jia, Y\. Zhang, P\. Chen, B\. Zeng, and X\. CaoDynamickk\-shot in\-context learning\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Wanget al\.\(2026c\)W\. Wang, W\. Yujia, Y\. Xiao, Z\. Chen, Y\. Zhang, P\. Chen, M\. Yang, and X\. CaoIncomplete in\-context learning\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 41629–41650\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Wanget al\.\(2022\)X\. Wang, J\. Wei, D\. Schuurmans, Q\. Le, E\. Chi, S\. Narang, A\. Chowdhery, and D\. ZhouSelf\-consistency improves chain of thought reasoning in language models\.arXiv preprint arXiv:2203\.11171\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Wuet al\.\(2024\)J\. Wu, M\. Feng, S\. Zhang, F\. Che, Z\. Wen, C\. Liao, and J\. TaoBeyond examples: high\-level automated reasoning paradigm in in\-context learning via mcts\.arXiv preprint arXiv:2411\.18478\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.
- Wuet al\.\(2026\)J\. Wu, Y\. Jiang, C\. Ma, Y\. Liu, H\. Zhao, Z\. Yuan, S\. Bai, and X\. BaiLiquid: language models are scalable and unified multi\-modal generators\.International Journal of Computer Vision134\(1\),pp\. 39\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Xuet al\.\(2025a\)F\. Xu, Q\. Hao, C\. Shao, Z\. Zong, Y\. Li, J\. Wang, Y\. Zhang, J\. Wang, X\. Lan, J\. Gong,et al\.Toward large reasoning models: a survey of reinforced reasoning with large language models\.Patterns6\(10\)\.Cited by:[§3\.3](https://arxiv.org/html/2609.10177#S3.SS3.SSS0.Px3.p1.1)\.
- Xuet al\.\(2025b\)N\. Xu, F\. Wang, S\. Zhang, H\. Poon, and M\. ChenFrom introspection to best practices: principled analysis of demonstrations in multimodal in\-context learning\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 3299–3324\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p2.1)\.
- Xuet al\.\(2025c\)Z\. Xu, K\. D\. Nguyen, P\. Mukherjee, S\. Bagchi, S\. Chaterji, Y\. Liang, and Y\. LiLearning to inference adaptively for multimodal large language models\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 3552–3563\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Yanget al\.\(2023\)L\. Yang, S\. Zhang, Z\. Yu, G\. Bao, Y\. Wang, J\. Wang, R\. Xu, W\. Ye, X\. Xie, W\. Chen,et al\.Supervised knowledge makes large language models better in\-context learners\.arXiv preprint arXiv:2312\.15918\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1)\.
- Yaoet al\.\(2026\)L\. Yao, L\. Xing, Y\. Shi, S\. Li, Y\. Liu, Y\. Dong, Y\. Zhang, L\. Li, Q\. Dong, X\. Dong,et al\.Towards efficient multimodal large language models: a survey on token compression\.Authorea Preprints\.Cited by:[§2\.1](https://arxiv.org/html/2609.10177#S2.SS1.p1.1)\.
- Younget al\.\(2014\)P\. Young, A\. Lai, M\. Hodosh, and J\. HockenmaierFrom image descriptions to visual denotations: new similarity metrics for semantic inference over event descriptions\.Transactions of the association for computational linguistics2,pp\. 67–78\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px1.p1.1)\.
- Zhanet al\.\(2026\)Z\. Zhan, S\. Zhou, X\. Zhou, Y\. Xiao, J\. Wang, J\. Deng, H\. Zhu, Y\. Hou, Y\. Song, M\. Lin,et al\.Retrieval\-augmented in\-context learning for multimodal large language models in disease classification\.Journal of Biomedical Informatics,pp\. 105017\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1)\.
- Zhanget al\.\(2025\)L\. Zhang, X\. Zeng, K\. Li, G\. Yu, and T\. ChenSc\-captioner: improving image captioning with self\-correction by reinforcement learning\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 23145–23155\.Cited by:[§4\.1](https://arxiv.org/html/2609.10177#S4.SS1.SSS0.Px3.p1.1)\.
- Zhanget al\.\(2023\)Y\. Zhang, K\. Zhou, and Z\. LiuWhat makes good examples for visual in\-context learning?\.Advances in Neural Information Processing Systems36,pp\. 17773–17794\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.
- Zhenget al\.\(2023\)C\. Zheng, L\. Li, Q\. Dong, Y\. Fan, Z\. Wu, J\. Xu, and B\. ChangCan we edit factual knowledge by in\-context learning?\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 4862–4876\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p1.1)\.
- Zhouet al\.\(2026\)A\. Zhou, H\. Wu, K\. Peng, P\. Liu, and X\. LiImproving few\-shot multi\-modal aspect\-level sentiment classification with implicit in\-context learning\.InInternational Conference on Multimedia Modeling,pp\. 301–314\.Cited by:[§2\.2](https://arxiv.org/html/2609.10177#S2.SS2.p1.1)\.
- Zhouet al\.\(2024\)Y\. Zhou, X\. Li, Q\. Wang, and J\. ShenVisual in\-context learning for large vision\-language models\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 15890–15902\.Cited by:[§1](https://arxiv.org/html/2609.10177#S1.p3.1)\.