THRONE: Ancient Authority & AI Benchmark

• THRONE is a concept embodying both the material and symbolic authority in ancient Egyptian funerary architecture and modern AI model evaluation.
• In archaeological contexts, the iron throne features complex construction from meteoritic iron and cedar, linked to ritual cosmic journeys and advanced artisan techniques.
• In vision-language models, the THRONE benchmark rigorously identifies hallucination errors by combining free-form responses and structured existence matrices.

A throne is an object of emblematic and functional significance found in multiple contexts, ranging from the apparatus of royal authority in ancient civilizations to the locus of error diagnosis in modern vision-language AI systems. In both cases, its definition is bound not only by materiality and construction, but by its symbolic, procedural, and methodological roles.

1. Thrones in Ancient Egyptian Funerary Architecture

Within Egyptian Fourth Dynasty mortuary complexes, the "iron throne" figures centrally in both ritual literature and speculative archaeological interpretation. Magli (2018) contextualizes the throne as a key funerary equipment item, hypothesizing its presence within the sealed void recently detected by muon radiography in Khufu’s pyramid at Giza (Magli, 2017).

In the Pyramid Texts, notably utterance PT536, the deceased king is enjoined to "sit on this your iron throne" after passing through metaphorical doors of the sky, with resurrection texts consistently linking the throne to a journey across northern sky doors and a stairway. This union of celestial symbolism and architectural geometry underpins the speculative identification of the Grand Gallery-mirroring void as a ritual "chamber" to house the throne, which is interpreted as a material correlate of the heavenly ascent.

2. Material Design and Technological Considerations

The hypothesized iron throne is inferred to follow the typology of the early 4th-Dynasty throne of Hetepheres, characterized by a low cedar frame with inlays and precious metal decoration. Unlike conventional iron objects, the metallic component in this case is proposed to be thin sheets of meteoritic iron, archaeologically discernible by a high Ni/Fe ratio as in Tutankhamen's dagger.

Observational constraints include an absence of specific dimensions or construction details; the throne is assumed to be comparable in scale to known dynastic analogues. Assemblage would involve cold working and hammering meteorite iron into laminas, subsequently affixed to the cedar support structure in a manner similar to gold leaf application. Evidence for indigenous iron-smelting is absent for this period, supporting the identification of meteoritic origin as the exclusive iron source.

3. Structure and Symbolic Integration within the Pyramid

Magli’s scenario rests on the architectural and geometrical congruence between the newly detected void and the Grand Gallery. The void, with a cross-section and length approximately equivalent to that of the Gallery (length ≥ 30 m), is theoretically written as $$V_{\text{void}} \simeq V_{\text{GG}} = L_{\text{GG}} \times W_{\text{GG}} \times H_{\text{GG}}$$, although no explicit metric values are given.

The architectural narrative posits a sealed “stairway” culminating at the location of the throne, accessed conceptually through the lower north shaft, whose terminus is mapped to a zone directly above the start of the Grand Gallery. The queen’s chamber, with its own shaft and doors, is interpreted as the locus for the Opening-of-the-Mouth ceremony, initiating the sequence of posthumous transformation culminating at the throne.

4. Provenance, Comparison, and Archaeometallurgical Evidence

Meteoritic iron was a rare but confirmed medium in pre-dynastic and Old Kingdom Egypt, with notable artefactual precedents in beads and high-nickel objects, including Tutankhamen’s blade. Analytical studies confirm that a previously recovered iron plate from Khufu’s pyramid is not meteoritic and is likely intrusive. The acceptance of the iron throne hypothesis would indicate a sophisticated level of meteorite iron cold-working and symbolic installation unprecedented in Fourth Dynasty craftsmanship.

5. Thrones in Vision-LLM Hallucination Auditing: The THRONE Benchmark

THRONE is also the designation for an object-based benchmark system designed to diagnose and quantify free-form hallucination errors in large vision–LLMs (LVLMs) (Kaul et al., 2024). Within this context, “throne” no longer denotes a physical seat but a diagnostic architecture for rigorously evaluating the propensity of LVLMs to invent non-existent objects in unconstrained text outputs.

Hallucinations are delineated into:

• Type I Hallucinations: Errors arising when an LVLM, in response to an open-ended prompt (e.g., “Describe this image in detail.”), inserts objects not present in the image.
• Type II Hallucinations: False assertions about object existence made in narrow, structured tasks (e.g., incorrect “yes” to “Is there a traffic light in this image?”).

Empirical evidence shows that improvement in Type II error rates does not reliably reduce Type I errors. Type I and Type II hallucinations may even be anti-correlated.

6. Methodology and Metrics in the THRONE Framework

The THRONE protocol evaluates LVLMs by prompting free-form responses per image and probing for object hallucinations across a fixed vocabulary $C$ (e.g., COCO80). For each (image, LVLM response, object $j$), an open-source LLM (FLAN-T5 Base, Large, XL) is tasked with abstractive QA components over three prompt variants, resulting in an existence matrix:

$\hat Y \in \{0,1\}^{|I|\times|C|},$

where $\hat Y_{i,j}=1$ only if all nine (three LMs × three prompts) answers are “Yes”, $0$ for unanimous “No”, else “ignored”. Fewer than 3% of entries are typically ignored.

Computed metrics include:

• Overall precision:

$$P_{\mathrm{ALL}} = \frac{\sum_{i,j}[\hat Y_{i,j}=1\,\land\,Y_{i,j}=1]}{\sum_{i,j}[\hat Y_{i,j}=1]}$$

• Overall recall:

$$R_{\mathrm{ALL}} = \frac{\sum_{i,j}[\hat Y_{i,j}=1\,\land\,Y_{i,j}=1]}{\sum_{i,j}[Y_{i,j}=1]}$$

• Class-wise precision:

$$P_{\mathrm{CLS}} = \frac1{|C|}\sum_{j=1}^{|C|}\frac{\sum_i[\hat Y_{i,j}=1\land Y_{i,j}=1]}{\sum_i[\hat Y_{i,j}=1]}$$

• Class-wise recall:

$$R_{\mathrm{CLS}} = \frac1{|C|}\sum_{j=1}^{|C|}\frac{\sum_i[\hat Y_{i,j}=1\land Y_{i,j}=1]}{\sum_i[Y_{i,j}=1]}$$

Principal model ranking is by class-wise $C$0-score with $C$1:

$C$2

to emphasize precision.

7. Empirical Findings, Baseline Comparisons, and Error Mitigations

Evaluation of ∼7B-parameter LVLMs (Adapter-v2, InstructBLIP, MiniGPT-4/2, LLaVA-v1.3/1.5/Mistral, Otter-Image, LRV-Instruction-v2, mPLUG-Owl) on COCO 2017 and Objects365 demonstrates persistent Type I hallucination rates: even top-performing systems achieve only 76% class-wise $C$3, corresponding to significant precision deficits. POPE, a baseline focused on Type II hallucinations with three positives/negatives per image, underestimates error frequency, with “POPE-C” (exhaustive class querying) revealing 20–50 point precision drops. CHAIR, a caption-matching-based baseline, has a false judgement rate nearly double that of THRONE in human scoring, primarily due to abstract term misclassification.

A simple, effective mitigation involves augmenting model training with explicit object enumeration tasks, whereby models are trained to list category occurrences and locations prior to free-form description. In LLaVA-v1.5, this raises class-wise $C$4 from 66.8% to 84.1% and increases Type II (POPE-C) precision by 4–6 points.

8. Conclusion and Thematic Synthesis

The concept of the throne illustrates the intersection of material construction, symbolic authority, ritual trajectory, and, in new semantic domains, model fidelity and diagnostic benchmarking. In funerary context, it actualizes the king’s celestial journey; in the assessment of artificial intelligence, it systematizes the auditing of model veracity in multimodal free-form generations. Both usages demand precision: of engineering and placement in the physical past, of metric and methodology in the computational present. The ongoing exploration of the Khufu pyramid’s void and the continued refinement of THRONE for LVLM evaluation share a common impulse—rigorous, evidenced inquiry into what is manifest, hypothesized, or hallucinatory (Magli, 2017, Kaul et al., 2024).