Overview

The dataset combines three perspectives: academic historiometry (MIT Pantheon), public attention (Wikipedia pageviews), and model judgments (LLM rankings and explanations). The resulting list is not a claim to a final truth. It is a benchmark that shows how different sources agree or disagree, where omissions occur, and which figures are contested.

Goals

This project has two intertwined aims. First, to offer a public, readable reference for learning about history at scale. Second, to use the “top 1,000” task as a probe into model behavior: what do LLMs remember, overlook, or exaggerate when asked to weigh human impact across centuries?

Choosing the 1,000 most influential people is inherently value-laden and historically contingent. That is precisely why it is a useful test: the models must reveal (and justify) their implicit criteria in the descriptions shown in each figure’s detail panel.

Data sources

Core prompt (current)

Every list is generated with the same prompt to maintain comparability. This prompt is reused verbatim across models.

Role: You are a senior historian and data scientist specializing in "Historiometry"—the statistical analysis of historical data. Task: Generate a ranked list of the 1,000 most influential figures in world history. Ranking Criteria: "Importance" must be calculated based on the following three metrics: Breadth: The geographic extent of their influence (Global vs. Regional). Depth: The degree to which they fundamentally altered human behavior, thought, or the state of the world. Longevity: The duration of their impact across centuries. Strict Constraints to Prevent Clustering: No Categorical Grouping: Do not group figures by profession, era, or nationality. This is a singular, linear competition of impact. For example, if rank #450 is a scientist and #451 is a poet, it must be because the scientist’s total score marginally exceeds the poet’s, not because you are listing "famous scientists" and then "famous poets." Linear Degradation: The list must represent a true descending order of influence. Rank #1 must be demonstrably more influential than #100, and #500 more than #1000. Global Balance: Ensure the list reflects major figures based on what you determine to be their objective historical weight, not just fame in one culture or region. Output Format: Provide the data in a raw JSON array of objects. Each object must contain: {"rank": integer, "name": "string", "primary_contribution": "string"} Technical Instruction: Do not include introductory or concluding conversational text—output the JSON block only. output the FULL list with no duplicates.

Ranking method

Each model contributes multiple lists. We average ranks within a model, then compute a consensus rank across models. If a model omits a figure, that omission is treated as rank 1001 (just below the top 1,000 cutoff). This prevents two-model outliers from dominating and captures the implicit judgment that a missing figure is less important than any ranked entry.

Variance and disagreement

Variance is calculated across the padded model ranks (including 1001 for omissions). High variance indicates contested figures, sparse agreement, or strong model disagreement. These are often the most interesting cases, revealing how differently models weigh ideology, scientific innovation, or cultural legacy.

Name reconciliation

Model outputs vary in spelling, honorifics, and transliteration. Names are normalized and mapped to canonical figure IDs via alias tables (for example: “Siddhartha Gautama” → “Gautama Buddha”). Manual merges remove duplicates so each historical figure appears once.

Geography + metadata

Birthplaces, regions, and descriptions are drawn from Wikipedia and Wikidata. Regions are intentionally stable across eras to support cross-temporal comparisons and map-based views. Each figure has a birthplace coordinate for future globe visualizations.

Interpretation: history + model behavior

The list is not just a ranking; it is a lens on model reasoning. LLMs struggle with this task because "importance" is not a stable, objective quantity. By making the model explanations visible in the detail panel, HistoryRank treats the list as an interpretive argument rather than a scoreboard.

Over time, this allows comparative analysis: how do models diverge on religious figures, colonial leaders, or scientific innovators? Which traditions or regions are consistently underweighted?

The circularity problem

There is an obvious methodological concern: LLMs are trained on human texts. When we ask them to rank historical figures, are they doing anything more than reflecting back the existing historiographical consensus?

In one sense, yes—that's exactly what they're doing. An LLM has no independent access to the past. It cannot evaluate Caesar's military campaigns or Confucius's teachings except through the lens of what humans have written about them.

But the circularity is less damning than it first appears. The training corpora for modern LLMs include texts from many cultures, time periods, and perspectives—far more than any individual human could synthesize. When Claude and GPT disagree about the relative importance of a figure, that disagreement reflects genuine tensions in the historical record, weighted differently by different training approaches.

We treat this not as a bug but as a feature. The consensus ranking represents what the accumulated text of human civilization, filtered through multiple model architectures, collectively "believes" about historical importance. The patterns of agreement and disagreement are themselves the data.

Model quality and exclusions

Not all models can complete this task reliably. The "rank 1,000 most influential people in history" challenge turns out to be a surprisingly effective test of frontier model capabilities. Only a handful of current models can maintain coherent global reasoning across 1,000 entries without falling into failure modes.

The most common failure is what we call autoregressive pattern collapse: the model loses track of the task constraint ("rank by historical importance") and instead starts optimizing for local sequence prediction ("what comes next in this pattern"). By rank 500–700, weaker models often fall into "local minima" where they generate long sequences of thematically similar entries that have no business appearing on a list of history's most important figures.

{
  "rank": 615,
  "name": "Edgaras Ulanovas",
  "primary_contribution": "Lithuanian basketball player; small forward."
},
{
  "rank": 616,
  "name": "Rokas Jokubaitis",
  "primary_contribution": "Lithuanian basketball player; point guard."
},
{
  "rank": 617,
  "name": "Ignas Vaitiekūnas",
  "primary_contribution": "Lithuanian basketball player; power forward."
},
{
  "rank": 618,
  "name": "Karolis Lukošiūnas",
  "primary_contribution": "Lithuanian basketball player; center."
},
{
  "rank": 619,
  "name": "Paulius Sarpalis",
  "primary_contribution": "Lithuanian basketball player; center."
},
{
  "rank": 620,
  "name": "Dovydas Giedvila",
  "primary_contribution": "Lithuanian basketball player; shooting guard."
},
// ... continues for 15+ more entries, cycling through positions
{
  "rank": 633,
  "name": "Arvydas Sabonis",
  "primary_contribution": "Lithuanian basketball player; center; one of the best international players ever."
}

Limitations and bias

Every source here is partial. Wikipedia reflects modern attention. Pantheon reflects an academic framing. LLMs inherit their training data and may amplify dominant narratives. The project treats these biases as data: where they align, where they diverge, and where they are silent.

Roadmap

Future work

Our next phase extends the current benchmark in two directions: prompt sensitivity and cross-lingual comparison. We will run two prompt variants in parallel: the current prompt and a matched version that removes the global balance instruction. Each model will produce 10 independent lists under each prompt, letting us compare stability within a prompt and shifts between prompts.

We will also translate the prompt into multiple languages (Arabic, Farsi, Mandarin, Russian, French, Spanish, Hindi) while keeping the JSON keys in English. Each language will be sampled with the same number of lists per model to enable clean comparisons.

Planned analyses include rank correlations across prompts and languages, top-k overlap, and distribution shifts by era, region, and domain. These results will be reported alongside within-model stability metrics to clarify which differences reflect model priors versus prompt framing.

February 2026 update

We introduced a new “Weighted Avg v3” view that combines all three list pools (v1, v2, and v3) while adding robust weighting and quality filters.

• Weighted v3 aggregates lists from `data/raw`, `data/raw_v2`, and `data/raw_v3`.
• Model IDs are canonicalized (e.g., `gpt-5.2`/`gpt-5.2-thinking` and `gpt-5.3`/`gpt-5.3-thinking` are merged).
• Canonical model weights are derived from cross-model reliability analysis.
• Weights are coverage-aware: shorter lists contribute less via a √(coverage) penalty.
• Missing ranks are imputed as (list length + 1), so shorter lists do not inflate impact.
• Bayesian shrinkage pulls sparse-coverage figures toward a neutral prior.
• Alias + merge remapping collapses known duplicates before aggregation.
• Worst 10 lists (by combined quality + advanced name hygiene metrics) are excluded from v3 weighted:
• GLM 4.7 LIST 3 (January 14, 2026).txt
• Mistral Large 3 LIST 5 (January 18, 2026).txt
• Qwen3 235B A22B LIST 7 (January 14, 2026).txt
• GLM 4.7 LIST 2 (January 14, 2026).txt
• Grok 4 LIST 1 (January 14, 2026).txt
• Mistral Large 3 LIST 4 (January 17, 2026).txt
• Grok 4.1 Fast LIST 1 (January 14, 2026).txt
• Mistral Large 3 LIST 1 (January 15, 2026).txt
• Grok 4 LIST 3 (January 14, 2026).txt
• Claude Sonnet 4.5 LIST 3 (January 12, 2025).txt