Beyond Precision: Introducing GAMUT for Factual Completeness in Long-Form Generation

What Changed

For the past several years, the evaluation of large language models (LLMs) has been heavily skewed toward precision. Developers and researchers have relied on the 'decompose-search-verify' pipeline, a methodology that excels at identifying incorrect claims and hallucinations by breaking down generated text into individual propositions and verifying them against trusted sources. While this approach has significantly improved the reliability of AI outputs, it has created a blind spot: factual completeness. A model can be perfectly precise—making no false claims—while still failing to provide a comprehensive or useful answer.

The introduction of GAMUT (Grounded Assessment of Multimodal Factuality) marks a fundamental shift in how the industry approaches this problem. Instead of focusing solely on whether a model is lying, GAMUT evaluates whether a model is providing the full set of information required by a prompt. This benchmark recognizes that facts are rarely flat, independent entities; they are often part of complex, ordered processes or hierarchical relationships that traditional binary grading systems fail to capture. By introducing a two-level meta-rubric framework, the researchers behind GAMUT provide a pathway to measure the 'missing half' of factuality.

Technical Details

The core innovation of GAMUT lies in its two-level rubric representation. The framework separates the evaluation into two distinct phases: the meta-rubric and the compiled checklist.

At the first level, a structured meta-rubric is defined. This layer captures the high-level organization, logical flow, and relative importance of the content required to answer a prompt. It accounts for the nuance of open-ended questions where coverage is more important than a simple list of facts. This meta-rubric is then mechanically compiled into a second level: a flat checklist of binary, machine-gradable rubrics. This compilation process allows an LLM judge to score the output reliably, as it is no longer tasked with interpreting complex, subjective requirements, but rather with verifying specific, discrete points derived from the structured meta-rubric.

The benchmark itself is grounded in 1,813 questions derived from real-world wearable imagery across 10 diverse domains. By using wearable data, the researchers ensure that the models are being tested on complex, multi-modal scenarios that require deep contextual understanding. Each question is paired with an evidence-backed rubric that has been verified by expert human annotators, ensuring that the ground truth is robust. Furthermore, the framework is modality-agnostic, meaning the underlying logic of the meta-rubric can be applied to text-only tasks, and the researchers have released a text-only variant to support broader adoption.

Benchmark Analysis

The evaluation of 14 frontier and open-weight models using the GAMUT framework has revealed significant performance gaps across the current AI landscape. The benchmark proved to be highly discriminative, meaning it effectively separates models that can handle complex, multi-faceted reasoning from those that struggle with comprehensive information retrieval. The best-performing model, Gemini 3.1 Pro, achieved a score of 58.7%. This relatively low ceiling suggests that even the most advanced models currently available struggle with the nuance of factual completeness, indicating that there is significant room for improvement in how models prioritize and synthesize information.

Developer Implications

For developers and AI engineers, the implications of GAMUT are twofold. First, it provides a new, more rigorous standard for evaluating long-form generation. If your application involves complex reasoning, technical documentation, or multi-step instructions, relying on precision-based metrics alone is likely insufficient. Implementing a two-level rubric approach, as demonstrated by GAMUT, can help you build more robust evaluation pipelines that ensure your models are not just avoiding errors, but are actually providing the necessary depth of information.

Second, the modality-agnostic nature of the framework means that developers can adapt these techniques to their own specific domains. By creating structured meta-rubrics for your specific use cases—whether that be medical diagnostics, legal analysis, or complex coding tasks—you can move away from relying on generic 'correctness' scores and start measuring the actual utility of your model's output. This shift toward structured evaluation is likely to become a standard practice as the industry moves from simple chat interfaces to complex, agentic workflows where completeness is a prerequisite for success.

Bottom Line

Factual completeness is the missing half of the factuality equation. While the industry has made great strides in reducing hallucinations, the ability to provide a complete, well-organized, and comprehensive response remains a significant challenge for frontier models. GAMUT provides the necessary framework to quantify this capability, moving the field beyond simple binary checks and toward a more sophisticated understanding of how models synthesize information. As models continue to scale, the ability to measure and improve factual completeness will be a key differentiator for high-performance AI systems.