Interpretable Evaluation of AI-Generated Content with Language-Grounded Sparse Encoders

Yiming Tang, Arash Lagzian, Srinivas Anumasa, Qiran Zou, Trang Nguyen, Ehsan Adeli, Ching-Yu Cheng, Yilun Du, Dianbo LiuarXiv 2025 · DOI · arXiv

Introduction

The rapid advancement of generative AI has outpaced our ability to evaluate it rigorously. Current standard metrics, such as Fréchet Inception Distance (FID) or CLIP scores, offer holistic numerical values that summarize an entire image or dataset. However, these "black-box" scores are often blind to specific, high-stakes failures. An image might receive a high score for correctly depicting a "knight on a horse" even if the knight possesses anatomical impossibilities or the horse exhibits surreal physical distortions. This "vibes-based" evaluation lacks the granularity needed for safety and quality control.

To address this, researchers have introduced LanSE (Language-Grounded Sparse Encoders), a framework designed to "autopsy" AI-generated content by decomposing images into over 5,000 human-readable visual concepts.

Methodology: The LanSE Pipeline

LanSE leverages techniques from mechanistic interpretability—specifically Sparse Autoencoders (SAEs) and Transcoders—to identify "monosemantic" neurons. These are individual neurons within a network that fire exclusively for a single, specific concept.

The LanSE architecture decomposes inputs into labeled sparse neurons representing semantic meaning, visual realism, and physical plausibility.

The pipeline operates in three distinct stages:

  1. Pattern Discovery: SAEs are used to reconstruct CLIP embeddings, forcing the model to utilize only a small number of active neurons at any given time. This sparsity ensures that each neuron corresponds to a distinct visual feature.
  2. LMM Interpretation: To translate these neural activations into human-understandable terms, a Large Multimodal Model (Claude-3.5-Haiku) analyzes the images that most strongly activate a specific neuron. It then generates a natural language label, such as "Structures with unrealistic proportions" or "Dogs in outdoor environments."
  3. Modality Distillation: These patterns are distilled into single-modality encoders, allowing LanSE to analyze images independently of their original text prompts.

Datasets and Domain Adaptation

LanSE was built and validated using a diverse array of data:

  • Natural Images: Utilizing MS-COCO, Flickr8k/30k, and TextCaps to establish a broad vocabulary of everyday concepts.
  • Medical Imaging: The framework was adapted into CXR-LanSE using the MIMIC-CXR dataset, enabling clinical auditing of chest X-rays for specific medical patterns.
  • AI-Generated Benchmarking: A test set of 48,000 images from eight state-of-the-art models (including DALL-E 3, FLUX.1-dev, and SDXL) was used to evaluate how different architectures handle specific visual concepts.

Key Insights: Convergence and Plausibility

One of the most significant findings is the phenomenon of Semantic Convergence vs. Error Divergence. While top-tier models like FLUX and SDXL have largely converged on what objects should look like (showing high correlation in semantic neurons), they vary significantly in how they fail. Each model possesses a unique "fingerprint" of artifacts and physical distortions.

LanSE identifies a wide variety of semantic concepts across different activities, such as soccer, equestrian sports, and skiing.

Furthermore, the research highlights that while modern models excel at Semantic Loyalty (following prompts accurately), they consistently struggle with Physical Plausibility. LanSE is the first tool to systematically evaluate this dimension, identifying errors like extra limbs or impossible lighting that holistic metrics miss.

Results and Performance

LanSE demonstrates exceptional alignment with human judgment and outperforms general-purpose models in error detection:

  • High Human Alignment: LanSE identified 5,309 reliable patterns in natural images with 93.2% human agreement. In the medical domain, CXR-LanSE achieved 74% agreement with radiologists.
  • Outperforming GPT-4o: In detecting physical plausibility errors (e.g., anatomical glitches), LanSE achieved 76.6% accuracy, surpassing GPT-4o (72.3%) and significantly outperforming Gemini Pro 1.5 (33.2%) and Claude-3-Opus (27.7%).
  • Model Leaderboard: Evaluation showed FLUX.1-dev as the leader in physical plausibility, while SDXL-medium performed best in photorealism and content diversity.

Limitations and Future Outlook

Despite its success, LanSE faces several challenges. It is currently limited by a language bottleneck, where it can only detect patterns that can be described in natural language. It also inherits any inherent biases present in its base CLIP embeddings. Additionally, the current vocabulary of 5,000+ patterns is static and will require updates as generative models evolve. Finally, the computational cost of training SAEs remains a barrier to real-time application.

By shifting mechanistic interpretability from an internal diagnostic tool to an external "microscope" for content quality, LanSE provides a scalable path toward more transparent and reliable AI evaluation.

Summary updated

Abstract

The rapid development of generative AI has transformed content creation, communication, and human development. However, this technology raises profound concerns in high-stakes domains, demanding rigorous methods to analyze and evaluate AI-generated content. While existing analytic methods often treat images as indivisible wholes, real-world AI failures generally manifest as specific visual patterns that can evade holistic detection and suit more granular and decomposed analysis. Here we introduce a content analysis tool, Language-Grounded Sparse Encoders (LanSE), which decompose images into interpretable visual patterns with natural language descriptions. Utilizing interpretability modules and large multimodal models, LanSE can automatically identify visual patterns within data modalities. Our method discovers more than 5,000 visual patterns with 93\% human agreement, provides decomposed evaluation outperforming existing methods, establishes the first systematic evaluation of physical plausibility, and extends to medical imaging settings. Our method's capability to extract language-grounded patterns can be naturally adapted to numerous fields, including biology and geography, as well as other data modalities such as protein structures and time series, thereby advancing content analysis for generative AI.
BibTeX
@misc{tang2025interpretable,
  author = {Yiming Tang and Arash Lagzian and Srinivas Anumasa and Qiran Zou and Trang Nguyen and Ehsan Adeli and Ching-Yu Cheng and Yilun Du and Dianbo Liu},
  title = {{Interpretable Evaluation of AI-Generated Content with Language-Grounded Sparse Encoders}},
  year = {2025},
  eprint = {2508.18236},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.18236},
}
Figure