Multimodal Large Language Models (MLLMs) are increasingly capable of generating "hyper-detailed" image captions, moving well beyond simple descriptions like "a cat on a mat" to produce paragraph-length scene analyses. However, as these descriptions grow in length, models tend to hallucinate more frequently.
To address this, researchers have introduced CapMAS (Caption factuality enhancing MultiAgent System), a multiagent pipeline that decomposes, verifies, and corrects long captions at an atomic level. Alongside this system, the paper introduces a new evaluation framework that independently measures a caption's Factuality (truthfulness) and Coverage (detail).
The Challenge of "Long-Form" Hallucinations
While modern MLLMs like GPT-4V or LLaVA excel at generating detailed text, they suffer from a "long-form hallucination" problem. The authors' analysis reveals a critical failure point: as a caption grows longer, the model begins to rely more heavily on its own generated text rather than the actual input image. This "contextual drift" is a major driver of fabricated details.
Empirical data demonstrates that traditional hallucination detection methods—which often rely on token-level confidence or consistency—lose their effectiveness for objects mentioned late in a long caption. Specifically, these metrics tend to fail once a caption exceeds roughly the 192-token threshold.
Furthermore, standard benchmarks like Visual Question Answering (VQA) do not necessarily reflect a model's ability to write accurate, detailed captions. A model might rank highly on a VQA benchmark by providing short, concise answers, but still struggle to describe a complex scene comprehensively without making things up.
CapMAS: A Multiagent Solution
To combat contextual drift, the researchers propose CapMAS, a "plug-and-play" multiagent pipeline that requires no additional training or fine-tuning of the underlying models. It operates in three distinct phases:
- Decomposition: A Large Language Model (LLM) acts as a "Decomposer." It breaks a long, complex caption into "atomic propositions"—simple, standalone claims that are definitively true or false.
- Verification: An MLLM acts as a "Fact-checker." It takes each atomic unit in isolation and compares it against the original image to assign a hallucination score. By isolating the facts (e.g., "Is there a red ball?"), the model avoids the distraction of the surrounding long-form text.
- Correction: A final LLM acts as a "Corrector." It receives the original caption alongside the list of verified and refuted facts. It then rewrites the caption to remove or fix the hallucinations while preserving the original flow and level of detail.
Redefining Evaluation: Factuality vs. Coverage
The authors argue that evaluating detailed captions requires a dual-axis approach:
- Factuality: The ratio of true statements to false statements.
- Coverage: How much of the image's actual information is captured in the text.
To support this, the study utilizes the DOCCI dataset (human-annotated, highly detailed image captions) to validate their new factuality metrics. They also built a new Coverage Benchmark using a subset called IIW-400. For this benchmark, the authors generated over 19,000 multiple-choice questions. To measure "Coverage," an LLM must answer these questions using only the generated caption, without seeing the image. If the caption is sufficiently detailed, the questions should be answerable.
Key Findings and Results
- Universal Improvement: CapMAS improved the factuality of every model tested, including state-of-the-art closed models like GPT-4V.
- Better Human Alignment: The proposed GPT-4o-based evaluation metric achieved a Spearman’s correlation of 70.2 with human judgments, significantly outperforming existing metrics like FactScore (67.9).
- The Factuality/Coverage Trade-off: CapMAS features a tunable hyperparameter ($\pi$). By making the threshold for "truth" stricter, users can increase factuality, though this sometimes results in a slight decrease in the total amount of detail (coverage).
- Enhanced Fluency: Qualitative and LLM-based evaluations showed that the final "corrected" captions were often more fluent and readable than the raw outputs from the base models.
Limitations and Future Directions
While highly effective, CapMAS is an inference-time strategy that comes with a computational cost. Because it requires multiple calls to LLMs and MLLMs for decomposition, verification, and correction, it is significantly slower and more expensive than a single-pass generation.
Additionally, the system's success depends heavily on the Decomposer LLM's ability to perfectly isolate atomic units. If a proposition remains too complex, the fact-checker might still struggle. Future work may involve using the high-quality data generated by CapMAS to fine-tune base models, allowing them to generate factual, hyper-detailed captions natively without the need for a multiagent pipeline.
Summary updated