Skip to content

Symbal: Detecting Systematic Misalignments in Model-Generated Captions

🕒 Published (v1): 2026-07-16 17:18 UTC · Source: Arxiv · Venue: ICML 2026 · link

Why this paper was selected

Stanford/Langlotz group; ICML 2026; systematic caption error detection for medical MLLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Multimodal LLMs generate image captions that contain recurring, bias-driven errors—termed systematic misalignments—where a specific textual error is consistently associated with a specific visual feature. Symbal is a two-stage, training-free pipeline that clusters and scores textual facts for misalignment, then identifies the correlated visual trigger. It achieves 63.8% end-to-end accuracy on the new SymbalBench benchmark, nearly 4× the closest baseline.

Problem

Existing image-caption alignment metrics (CLIPScore, CHAIR, GREEN, etc.) operate at the per-sample level and cannot identify global, systematic captioning errors—i.e., recurring textual errors that are statistically associated with a specific visual feature across an entire dataset. No benchmark existed to evaluate such global error detection, and ground-truth annotations at dataset scale are intractable to collect manually. This matters especially in safety-critical domains (e.g., radiology), where spurious correlations learned during MLLM training can produce systematically wrong yet plausible-sounding outputs.

Method

Symbal structures detection into two sequential stages, each with three subtasks (group → score → summarize), using off-the-shelf foundation models without any fine-tuning.

Stage 1 — Detect erroneous textual fact \(\hat{t}\): 1. Grouping: Split all captions into sentence-level textual facts; encode with a text embedding model (best: Qwen3-Embedding-8B); cluster with spherical K-Means, selecting \(K\) automatically via Silhouette distance. 2. Scoring: Score each cluster by mean image-text alignment. Three options: (a) embedding cosine similarity, (b) text-only LLM judge on generated image captions, (c) vision-language scorer feeding the image + fact directly to an MLLM (Qwen2.5-72B). Low score → high misalignment. 3. Summarizing: The lowest-scoring cluster \(C_{\text{text}}\) is summarized by an LLM into a single concept \(\hat{t}\) (e.g., "cardiomegaly").

Stage 2 — Detect associated visual feature \(\hat{v}\): 1. Grouping: Collect all images paired with facts in \(C_{\text{text}}\); encode with an image embedding model; cluster with spherical K-Means. 2. Scoring: Score each image cluster by mean misalignment with \(\hat{t}\) using the same scorer options. 3. Summarizing: The lowest-scoring cluster \(C_{\text{image}}\) is summarized into \(\hat{v}\) (e.g., "pacemaker") via either a text-only LLM operating on generated captions or a vision-language MLLM operating on the images directly.

SymbalBench is constructed by injecting predefined \((t, v)\) misalignments (controlled via Cramér's V association strength) into two base datasets—COCO (4,349 natural images) and MIMIC-CXR (2,233 chest X-rays)—yielding 420 evaluation settings and 1.7 M image-text pairs. Evaluation uses LLM-as-a-Judge (Llama3.3-70B) for semantic equivalence and reports Accuracy@K.

Key Contributions

  • Systematic misalignment detection task: formal definition of \(\mathcal{M}: \mathcal{D} \to (\hat{t}, \hat{v})\) for dataset-level captioning error discovery.
  • Symbal method: dual-stage, training-free pipeline with modular grouping/scoring/summarizing subtasks; supports both reference-free and reference-based settings.
  • SymbalBench: first benchmark for this task; 420 datasets, 1.7 M image-text pairs, two domains (natural + medical), automated injection with guaranteed ground-truth labels.
  • Real-world auditing: demonstrated on captions from four MLLMs and on existing public image-caption datasets in a black-box (no MLLM access) setting.

Results

  • End-to-end (SymbalBench): Symbal correctly identifies the systematic misalignment in 63.8% of 420 datasets—a nearly 4Ă— improvement over the closest baseline.
  • Stage 1, natural images (reference-free): Acc@5 = 94.2%; reference-based Acc@5 = 82.8% (best config: Qwen3-Embedding-8B + Qwen2.5-72B VL scorer + Qwen2.5-72B text summarizer).
  • Stage 1, medical images (reference-free): Acc@5 = 75.0%; reference-based Acc@5 = 95.0% (reference helps due to complexity of radiology text).
  • The vision-language alignment scorer substantially outperforms the text-only scorer under identical backbone (Qwen2.5-72B).
  • Reference-free performance exceeds reference-based for natural images (sparse COCO captions are unhelpful); the pattern reverses for medical images (complex domain benefits from reference text).

Limitations

  • SymbalBench uses injected, predefined misalignments—real-world MLLM errors may have different distributional properties or be harder to formalize.
  • The method predicts a single \((t, v)\) pair per run; multi-misalignment extension is discussed only in appendices and not the primary evaluation.
  • Spherical K-Means with automatic \(K\) selection may fail when misalignment clusters are diffuse or when dataset size is small.
  • No fine-tuning or model training means performance is bounded by the capability of off-the-shelf embedding and scoring models, which may underperform in highly specialized domains.
  • Reference-free performance degrades on medical images, limiting applicability in clinical settings where ground-truth reports may be unavailable.
  • Evaluation relies on LLM-as-a-Judge (Llama3.3-70B) for semantic equivalence, which may itself introduce errors.

Relevance to Vision-Language Models

Symbal directly targets a failure mode of MLLMs used for image captioning: spurious correlations encoded in model weights that produce systematic, plausible-but-wrong text conditioned on visual triggers. For VLM researchers, this work provides both a diagnostic tool and a benchmark to quantify how much a given VLM's caption generation is corrupted by dataset biases, without needing access to the model's internals. It bridges the gap between per-sample alignment metrics (CLIPScore, CHAIR) and the need for dataset-level auditing, which is increasingly critical as MLLM-generated captions are recycled as training data for next-generation VLMs. The medical imaging case study (pacemaker → cardiomegaly hallucination) illustrates a concrete, high-stakes failure mode that existing VLM evaluation pipelines would miss entirely.