Cross-modal Iterative Prompting (CIP)

• The paper introduces CIP as a method that uses both class names and support images to generate precise, grounded textual prompts in few-shot learning.
• It follows a structured reasoning pass with stages (Strategy, Perception, Refinement, Conclusion) to iteratively refine class semantics.
• CIP reduces ambiguity and hallucinations in class definitions, improving synthetic image generation and overall classifier performance.

Cross-modal Iterative Prompting (CIP) is a component of VT-FSL, a few-shot learning framework that bridges vision and text with LLMs by conditioning prompt generation on both class names and support images. In this usage, CIP generates precise, class-specific, visually grounded textual descriptions iteratively in a single structured reasoning pass, so that text is refined by actual visual evidence from the support set rather than produced from the label alone (Li et al., 29 Sep 2025). A recurrent source of confusion is acronym overlap: in a separate paper, CIP denotes “Causal Inference-based Input Enhancement” or a “Causal Inference Plugin,” a plug-and-play causal prompting framework for mitigating hallucinations under long-context noise, not Cross-modal Iterative Prompting (Ma et al., 12 Dec 2025).

1. Terminological scope and problem setting

Within VT-FSL, CIP is introduced to address a specific weakness of prior semantic augmentation methods for few-shot learning. The stated problem is that class names alone are too weak: they often carry little context and can be ambiguous. The paper further states that LLM-generated text from names alone can hallucinate semantics that contradict the visual evidence due to the lack of grounding in actual instances, resulting in noisy guidance and costly corrections (Li et al., 29 Sep 2025).

The motivation is therefore not generic prompt improvement, but the construction of precise cross-modal prompts conditioned on LLMs and support images. The central claim is that text should not be generated from the label alone, but should be refined by actual visual evidence from the support set. This places CIP in the space of grounded prompt construction for episodic few-shot learning rather than in long-context hallucination mitigation or retrieval-time causal control.

A useful distinction follows from the paper’s own contrasts with prior work. CIP is presented as different from simple prompt expansion or semantic augmentation methods that only use names, knowledge-base-driven methods such as WordNet-based expansion, attribute generation methods that still rely mostly on class names, and standard single-pass prompting that does not explicitly structure reasoning. This suggests that the defining feature of CIP is not merely semantic enrichment, but cross-modal grounding plus an explicitly structured reasoning format (Li et al., 29 Sep 2025).

2. Cross-modal conditioning and the rationale for iteration

CIP is described as both cross-modal and iterative. It is cross-modal because it uses both textual input, namely the class name, and visual input, namely the support images. It is iterative because the generation is guided through structured reasoning stages so the LLM gradually refines its description (Li et al., 29 Sep 2025).

The paper’s rationale for this design is direct. Prior prompting is characterized as too naive because earlier methods typically feed only the class name to the LLM and do not exploit the support images, so the text is not grounded in the episode’s actual data. By contrast, CIP conditions the LLM on both sources of evidence, and the support images help disambiguate vague or polysemous labels. The intended result is a class description that is more precise than the class name, grounded in the support images, suitable for synthetic image generation, and useful as a semantic prompt for the classifier.

The “iterative” aspect should not be conflated with multi-turn dialogue. The paper specifies a single structured inference pass. The iteration lies in internal refinement across tagged reasoning stages, especially in the step where the model compares the initial semantic interpretation with the observed image evidence, removes hallucinated or irrelevant attributes, adjusts the class meaning if needed, and strengthens the description with visually supported details. A plausible implication is that the method seeks the benefits of staged reasoning without the overhead of multi-turn interaction.

3. Structured reasoning pass

The paper states that CIP follows a single structured inference pass with four stages: Strategy, Perception, Refinement, and Conclusion. These stages are explicitly marked in the main text, while the appendix gives a semantically aligned prompt template using <SUMMARY>, <CAPTION>, <REASONING>, and <CONCLUSION> tags (Li et al., 29 Sep 2025).

Stage in main text	Template section in appendix	Function
Strategy	<SUMMARY>	Interpret the class name and plan revision
Perception	<CAPTION>	Describe common visual elements across images
Refinement	<REASONING>	Confirm or adjust class meaning from evidence
Conclusion	<CONCLUSION>	Produce the final concise definition

In the Strategy stage, the LLM acts like a “visual taxonomy expert.” It interprets the class name, considers possible meanings if the label is ambiguous, and establishes an initial semantic anchor. The prompt template summarizes the plan as follows: interpret the class name, analyze the images, and revise the description using both sources. This is the initial hypothesis before visual grounding is applied.

In the Perception stage, the LLM examines the support images and is asked to describe common visual elements across images, including shapes, colors, textures, background patterns, object parts, and distinctive appearance cues. The paper emphasizes shared traits across support examples as evidence for class definition. This stage grounds the class meaning in observed evidence and helps resolve ambiguity in the label.

In the Refinement stage, which the paper explicitly identifies as the key “iterative” part, the model aligns class semantics with observed visual evidence and eliminates hallucinated or irrelevant attributes. The description is revised so that it becomes visually grounded rather than purely text-based. In the Conclusion stage, the model outputs a final class description that should be concise, scientifically accurate, visually consistent, and non-redundant.

4. Prompt template, outputs, and enriched support construction

The appendix provides the operative template. It instructs the model as follows: “Given a class name and a set of representative images, your task is to generate a brief, scientifically accurate, and visually grounded definition. The definition should be primarily guided by the semantic meaning of the class name, and refined using the visual evidence from the images to resolve ambiguities and enhance relevance.” It then requires the structured reasoning format with <SUMMARY>, <CAPTION>, <REASONING>, and <CONCLUSION>, followed by Class Name: {class_label} and Images: [K image inputs attached here] (Li et al., 29 Sep 2025).

The direct output of CIP is a class description for each novel class in the episode. In the paper’s terminology, this output is a textual prompt $T_N$ for the $N$ classes. The description is then reused downstream in two ways. First, it functions as a semantic prompt. Second, it supports zero-shot synthesis of semantically consistent images, which are generated by a text-to-image generative model and then filtered by an LLM-based pairwise comparison strategy that retains the top-$K$ images per class.

The augmented support construction is formalized as

$$\hat{S}_{N \times (K+K)} = \{\text{Synthesize}({T}_N),\ S_{N \times K}\}.$$

Here, $T_N$ denotes generated textual descriptions for the $N$ novel classes, $S_{N \times K}$ denotes the original support images, $\text{Synthesize}(T_N)$ denotes synthetic images generated from the descriptions, and $\hat{S}_{N \times (K+K)}$ denotes the expanded support set combining real and synthetic samples (Li et al., 29 Sep 2025).

The paper characterizes the descriptions and synthetic images as complementary textual and visual prompts. The descriptions provide high-level class semantics, while the synthetic images provide low-level intra-class diversity to compensate for limited support data. In 1-shot settings, where real support data are extremely scarce, this dual use of CIP outputs is presented as especially consequential.

5. Position within VT-FSL and relation to multimodal alignment

CIP is only the first stage of VT-FSL, but the paper describes it as foundational. The pipeline is presented in the following order: CIP generates class descriptions from class names and support images; text-to-image synthesis uses those descriptions to create synthetic images; Cross-modal Geometric Alignment (CGA) aligns textual features $Z_t$, enhanced support features $N$0, and synthetic visual features $N$1; final class prototypes are computed from the fused prompts and support features; and query samples are classified using the learned prototypes (Li et al., 29 Sep 2025).

The paper states that CGA jointly aligns the fused textual, support, and synthetic visual representations by minimizing the kernelized volume of the 3-dimensional parallelotope they span. This captures global and nonlinear relationships among all representations, enabling structured and consistent multimodal integration. CIP therefore supplies the semantic material that makes the later geometric alignment meaningful: without grounded textual descriptions, the synthetic visual branch and the fused multimodal representation would be less reliable.

During testing, VT-FSL builds a textual prototype $N$2 from enhanced support features conditioned by text and a visual prototype $N$3 from the support set expanded by synthetic images. These are fused as

$N$4

where $N$5 is selected on validation data. CIP contributes indirectly to the final prototype $N$6 because it is responsible for generating the textual descriptions that support both the textual prompt branch and the synthetic image branch (Li et al., 29 Sep 2025).

The appendix also makes CIP explicit in the training algorithm: textual descriptions $N$7 are generated via an LLM with class name and $N$8 support images using the four structured reasoning stages, after which synthetic images are generated from $N$9 and top-$K$0 images are selected with LLM-based pairwise comparison. The paper characterizes this as an offline prompt-generation mechanism inside each episode, before feature extraction and alignment.

6. Empirical role, comparative claims, and common misconceptions

The paper attributes several benefits to CIP within VT-FSL. It says that CIP-generated text works better than class names and SemFew-style definitions because it combines the semantic prior of the class name and the visual evidence from the support images. This reduces ambiguity, hallucination, noisy guidance, and mismatch between text and visual samples. The ablations are summarized as showing that using both textual and visual prompts together gives the best results, and that the descriptions generated by CIP outperform simpler text generation schemes (Li et al., 29 Sep 2025).

At the system level, the proposed VT-FSL method is reported to establish new state-of-the-art performance across ten diverse benchmarks, including standard, cross-domain, and fine-grained few-shot learning scenarios. Since VT-FSL mainly consists of CIP and CGA, these results position CIP as the semantic engine of the framework rather than as an isolated classifier component. The paper also argues that CIP improves performance by disambiguating class names, grounding semantics in actual support images, producing better prompts for synthetic image generation, creating complementary textual and visual prompts, and supporting better geometric alignment later.

Two misconceptions are explicitly addressed by the source material. First, Cross-modal Iterative Prompting is not equivalent to naive LLM prompting for class-name definitions. Its defining property is that it prompts the LLM to reason over both modalities before producing the final class description. Second, it is not the same CIP acronym used in “CIP: A Plug-and-Play Causal Prompting Framework for Mitigating Hallucinations under Long-Context Noise,” where CIP stands for Causal Inference-based Input Enhancement or a Causal Inference Plugin and targets hallucination mitigation under long-context noise rather than few-shot multimodal grounding (Ma et al., 12 Dec 2025).

A plausible implication is that the significance of CIP lies in converting sparse few-shot supervision into usable multimodal guidance. In the paper’s own terms, it turns weak class labels into rich, grounded descriptions, enables zero-shot image synthesis, and provides the semantic basis for later alignment and prototype fusion.