Concept Matching with Agent for Out-of-Distribution Detection

TL;DR

This work tackles the challenge of robust OOD detection by moving beyond binary in/out boundaries and introducing a zero-shot, agent-based framework. Concept Matching with Agent (CMA) leverages neutral textual prompts as Agents to create a vector triangle relationship among ID labels, Agents, and data inputs within a CLIP-based vision–language space, enabling robust separation without additional training. Empirical results across large-scale and small-scale benchmarks show CMA outperforms both zero-shot and training-dependent baselines on AUROC and FPR95, and analyses reveal optimal agent counts and agent-specific effects. The approach demonstrates strong practical impact by offering a scalable, training-free method that adapts to diverse scenarios with potential for tailored Agent design.

Abstract

The remarkable achievements of Large Language Models (LLMs) have captivated the attention of both academia and industry, transcending their initial role in dialogue generation. To expand the usage scenarios of LLM, some works enhance the effectiveness and capabilities of the model by introducing more external information, which is called the agent paradigm. Based on this idea, we propose a new method that integrates the agent paradigm into out-of-distribution (OOD) detection task, aiming to improve its robustness and adaptability. Our proposed method, Concept Matching with Agent (CMA), employs neutral prompts as agents to augment the CLIP-based OOD detection process. These agents function as dynamic observers and communication hubs, interacting with both In-distribution (ID) labels and data inputs to form vector triangle relationships. This triangular framework offers a more nuanced approach than the traditional binary relationship, allowing for better separation and identification of ID and OOD inputs. Our extensive experimental results showcase the superior performance of CMA over both zero-shot and training-required methods in a diverse array of real-world scenarios.

Figures (10)

• Figure 1: In OOD detection, the Vector Triangle Relationship alters the traditional Binary Relationship by introducing Agents, thereby more effectively processing and distinguishing between ID data and OOD data.
• Figure 2: Overview of Concept Matching with Agent (CMA) framework. The input image $x$ undergoes Image Encoder $\mathcal{I}$ to produce an image embedding. The concatenation of the ID labels $\mathcal{Y}_{in}$ and Agents $\mathcal{Y}_{ntc}$ is then subjected to Text Encoder $\mathcal{T}$ to generate a text embedding. The similarity between the image and text embeddings is computed, with a higher result indicating a greater degree of similarity (darker shading denotes higher similarity). This is followed by the CMA operation, which computes the $\mathcal{S}_{\text{CMA}}$ for each image as the ultimate discriminative metric. Further details are provided in Section \ref{['sec:method']}.
• Figure 3: Heatmaps depicting the cosine similarity between image inputs and ID concept vectors. In Figure (A), the ID concept vectors consist of sentences containing the word "cat" of varying lengths. It is observed that images tend to align with longer sentences regardless of whether there is a matching ID concept. Concurrently, keywords such as "white" significantly influence image matching. In Figure (B), aside from "cat", no ID concept can be precisely matched with the given images. However, other than cat images, all images exhibit a preference for aligning with a long sentence devoid of tangible objects. Notably, cat image remains unaffected, aligning solely with the ID concept "cat". All the data in the figure is obtained from the practical use of CLIP (https://github.com/openai/CLIP).
• Figure 4: Impact curve of$k = \frac{\text{number of Agents}}{\text{number of ID Labels}}$on the performance of CMA.Left shows that as $k$ gradually increases, the FPR95 on various datasets generally decreases, with the fastest decline occurring in the range of $k$ less than 0.5, followed by a gradual slowdown, which is more evident on the average curve. Right shows that as $k$ gradually increases, the AUROC gradually increases, also with a rapid rise followed by a gradual slowdown.
• Figure 5: Comparison with different Agents.Left shows the performance of different agents on various datasets in terms of FPR95, while Right shows the performance in terms of AUROC. Clearly, the agent that performs best on average across all datasets does not perform best on every dataset. Moreover, even agents that perform poorly on average can still show decent performance on certain datasets.