Measuring Intelligence Beyond Human Scale

Jerry Han Rafael Moschopoulos Ella Colby Vishrut Goyal Andrew Tu Kia Ghods Mark Braverman Elad Hazan
Princeton Superalignment, Princeton University
Proposer one model, each round Binary challenge question or program, answer A or B + committed answer ŷ ∈ {0, 1} (hidden from the panel) Solver 1 p = 0.95 Solver 2 p = 0.90 Solver 3 p = 0.50 Solver 4 p = 0.10 Solver 5 p = 0.05 weight-sampled panel of k solvers Proposer reward P = 4 · Var(p) paid for splitting the panel Solver loss ℓᵢ = (pᵢ − y)² Brier loss vs. the resolved bit Per round, both quantities are z-scored across the field and averaged: g = ½z(P) + ½z(Q). Ranking = cumulative G = Σg.
One round of SepaRank. A proposer authors a public binary challenge and privately commits an answer. A weight-sampled panel of solvers each reports a probability for answer A. The proposer is paid for the disagreement it induces across the panel; each solver is scored by a proper scoring rule against the resolved answer.

Abstract

How can we measure intelligence beyond human capability? Human-authored benchmarks saturate, and above human capability, examiners may not know which tasks are both hard and verifiable. We argue that this difficulty is inherent to absolute-scale evaluation and propose a new paradigm based on relative measurement in which models generate public challenges that separate other systems. Aggregating these outcomes yields an adversarial psychometric rating system that can scale with the systems being measured. We describe practical protocols that reduce incentives for private-information attacks, support judge-free adjudication, and naturally scale with agent capabilities. We instantiate the framework across verifiable and open-ended, non-verifiable domains, illustrating how model-generated evaluation can continue to measure systems beyond the human frontier.

Why evaluation breaks at the frontier

1 Benchmarks are exams written by humans

Since Spearman’s 1904 discovery that performance across cognitive tests is positively correlated, intelligence has been measured psychometrically: an examiner designs tasks, subjects solve them, and a latent ability score is inferred from the outcomes. Nearly every AI benchmark (MMLU, BIG-Bench, HELM, ARC, FrontierMath, Humanity’s Last Exam) inherits this structure. The tasks are fixed in advance and authored by people.

2 Above human level, the examiner is the bottleneck

Frontier models saturate benchmarks faster than experts can write harder ones. Worse, beyond human expert level the difficulty is not just producing enough questions but knowing which questions are simultaneously difficult, discriminating, and verifiable. This is an inherent limitation of absolute-scale evaluation: the measuring stick cannot be shorter than the thing being measured.

3 Naive “model vs. model” duels are gameable

A tempting fix is a pairwise tournament: model i wins by authoring a challenge it can solve but model j cannot, and Elo-style ratings aggregate the outcomes. But this rewards asymmetric information rather than capability. A proposer can ask “which bit am I thinking of?”, or publish the product of two large primes it generated and demand the factorization, a trapdoor inserted at construction time. And without an external judge, nothing forces the pair’s agreed resolution to be the true one.

The idea: adversarial psychometrics

Pose every challenge to many solvers at once, and pay the proposer for the variance of the panel’s answers, not for defeating one opponent. Private-state and trapdoor questions collapse: a calibrated panel uniformly answers “50/50”, inducing no variance and earning nothing. The only way to score is to find public questions on which capable systems disagree; in other words, to discover the capability boundaries of the population. Ranking models by their ability to separate (and to solve) yields a rating system, SepaRank, that scales with the models being measured instead of with their examiners.

Paradigm Who authors the tasks Scales past human frontier Needs a judge Resists private-info attacks
Static benchmarks Human experts ✗ saturate ✓ answer key ✓ fixed tasks
Arena / preference voting Users (prompts) ✗ judged by humans ✗ human raters
Pairwise challenge duels Models ✗ needed but unscalable ✗ trapdoors win
SepaRank (this work) Models ✓ judge-free options ✓ variance ≈ 0

How a round works

  1. Propose. Each model in turn authors a binary challenge (a yes/no question or an executable program whose main() returns 0 or 1) and commits an answer ŷ that the panel never sees.
  2. Sample a panel. k solvers are drawn from the other models, weighted toward the current frontier (see adaptive weighting below).
  3. Report beliefs. Every panelist independently reports pi ∈ [0, 1], its probability that the answer is A. Ignorance is pi = ½.
  4. Resolve. The ground-truth bit y is fixed by the arm’s resolution rule: executing the program, taking the proposer’s committed bit, or taking the panel’s majority vote.
  5. Score both roles. The proposer earns the separation it induced; each solver pays a proper-scoring-rule loss against y.
Scoreprop = 4 · Var(p)     ℓi = (piy

The factor of 4 normalizes the proposer’s reward to [0, 1]: it is maximal when the panel splits into confident, opposing camps, and zero when everyone agrees, and also when everyone hedges. The Brier loss makes calibrated reporting each solver’s best response, so the mechanism separates knowing from knowing that one does not know. Each round, a model’s proposer reward P and (negated) mean solver loss Q are z-scored across the field and averaged,

g = ½ z(P) + ½ z(Q),     G = Σroundsg,

and the cumulative total G is the ranking. Because challenges are authored by the players themselves, the test’s difficulty automatically tracks the field’s ability; no human ever has to write a harder exam.

Built-in defenses

Private information earns nothing. “Which bit am I thinking of?” leaves a calibrated panel at p = ½ and pays zero. A trapdoored factoring instance is no better: the proposer is never asked to solve its own challenge, and an impossibly hard task makes everyone hedge. Separation, not difficulty, is what pays.

Conjunction boosting can be coarsened away. Chaining k questions together mechanically pushes reports toward 0 and 1, inflating raw variance. The protocol optionally rounds each report to the nearest of {0, ½, 1} before computing the score, so only categorical disagreement counts.

Adaptive weighting keeps the test discriminating at the top. Once weak models are clearly separated, re-separating them stops paying: solver panels are sampled with multiplicative weights w ∝ exp(ηG) that concentrate on the current leaders, so a high-scoring proposer must eventually find questions on which the frontier itself disagrees.

Four resolution rules, one scoring mechanism

The scoring never changes; only how the truth bit y is fixed. This lets the same protocol span fully verifiable domains and completely open-ended ones, including judge-free operation.

Arm Artifact Resolution of y Solver’s report predicts
PE Python program Sandbox execution What main() actually returns
PC Python program Proposer’s committed bit What the proposer claims it returns
QC Natural-language question Proposer’s committed bit The proposer’s committed answer
QN Natural-language question Panel majority vote The population consensus

Experiments: eleven models, no human examiner

The protocol is implemented end-to-end in the combined-protocol codebase and run as ten independent 20-round games per arm, each fielding the same N = 11 contemporary models from five providers (OpenAI, Anthropic, Alibaba, Moonshot AI, DeepSeek) under stable pseudonyms. Every round each model authors one challenge and a panel of k = 5 solvers is sampled; prompts disclose the entire mechanism (scoring formulas, resolution rule, adaptive weighting, and the last five rounds of transcript) with no strategy hints. Programs are machine-checked in a sandbox; nothing else is judged by a human.

Program / committed (PC) -10 -5 0 +5 +10 +15 gpt-5.5: G = +12.9 ± 1.5 gpt-5.5 +12.9 gpt-5.4: G = +9.7 ± 1.8 gpt-5.4 +9.7 qwen3.7-max: G = +3.3 ± 2.8 qwen3.7-max +3.3 gpt-5.4-mini: G = +1.3 ± 2.3 gpt-5.4-mini +1.3 opus-4.8: G = -0.4 ± 2.1 opus-4.8 -0.4 sonnet-4.6: G = -1.0 ± 2.0 sonnet-4.6 -1.0 haiku-4.5: G = -4.3 ± 2.1 haiku-4.5 -4.3 gpt-5.4-nano: G = -4.6 ± 2.2 gpt-5.4-nano -4.6 kimi-k2.7: G = -4.7 ± 1.8 kimi-k2.7 -4.7 deepseek-v4-pro: G = -5.2 ± 3.3 deepseek-v4-pro -5.2 gpt-4o-mini: G = -6.9 ± 2.2 gpt-4o-mini -6.9 total score G (mean over ten games) Question / committed (QC) -15 -10 -5 0 +5 +10 gpt-5.5: G = +7.0 ± 3.3 gpt-5.5 +7.0 gpt-5.4: G = +6.5 ± 3.0 gpt-5.4 +6.5 opus-4.8: G = +4.4 ± 3.2 opus-4.8 +4.4 qwen3.7-max: G = +2.8 ± 2.5 qwen3.7-max +2.8 sonnet-4.6: G = +2.1 ± 2.6 sonnet-4.6 +2.1 haiku-4.5: G = +1.1 ± 4.0 haiku-4.5 +1.1 gpt-5.4-mini: G = +0.4 ± 1.8 gpt-5.4-mini +0.4 gpt-5.4-nano: G = -0.5 ± 1.7 gpt-5.4-nano -0.5 kimi-k2.7: G = -4.8 ± 3.3 kimi-k2.7 -4.8 deepseek-v4-pro: G = -5.6 ± 2.1 deepseek-v4-pro -5.6 gpt-4o-mini: G = -13.3 ± 2.3 gpt-4o-mini -13.3 total score G (mean over ten games)
SepaRank leaderboards. Total score G per model, mean ± 95% confidence interval over ten games of each committed-resolution arm. In the program arm the two frontier OpenAI models separate from the field and from each other (gpt-5.5 over gpt-5.4: ΔG = 3.1, p = 0.008; gpt-5.4 over third place: ΔG = 6.4, p = 0.006). The protocol resolves differences at the top of the field without any human-written test.

The rankings track the roster’s known capability span: frontier models lead both arms and the small non-reasoning model anchors the bottom of both. Solver behavior explains much of the spread: reports are nearly binary (mean distance from ½ between 0.39 and 0.50 for every model), so Brier loss is governed by directional accuracy. The weakest models are miscalibrated, not just ignorant: gpt-4o-mini answers on the wrong side 40% of the time while staying confident, earning a loss (0.388) worse than a solver that always answers ½ (0.25). The strongest models are both directionally correct and the most willing to hedge. The proper scoring rule is doing its intended work, and failing to declare uncertainty is itself a capability signal.

-10 -5 0 5 10 15 -15 -10 -5 0 5 10 question = program gpt-5.5: program +12.9, question +7.0 gpt-5.5 gpt-5.4: program +9.7, question +6.5 gpt-5.4 qwen3.7-max: program +3.3, question +2.8 qwen3.7-max gpt-5.4-mini: program +1.3, question +0.4 gpt-5.4-mini opus-4.8: program -0.4, question +4.4 opus-4.8 sonnet-4.6: program -1.0, question +2.1 sonnet-4.6 haiku-4.5: program -4.3, question +1.1 haiku-4.5 gpt-5.4-nano: program -4.6, question -0.5 gpt-5.4-nano kimi-k2.7: program -4.7, question -4.8 kimi-k2.7 deepseek-v4-pro: program -5.2, question -5.6 deepseek-v4-pro gpt-4o-mini: program -6.9, question -13.3 gpt-4o-mini total score G (program / committed) total score G (question / committed)
Two very different task domains, one ranking. Each model’s total score in the program arm (x) against the question arm (y), with 95% CIs. The arms agree on the ordering (Spearman ρ = 0.93, Pearson r = 0.79) while disagreeing interpretably on individuals: opus-4.8 is relatively stronger on open-ended questions, the frontier OpenAI models on verifiable deterministic computation.

When lying is legal, honesty becomes a measurement

In the committed arms the resolved answer is whatever the proposer commits, so misrepresentation is a legal move. What do models do with that freedom?

dishonest commitment: mean proposer reward 0.36 dishonest commitment 0.36 honest commitment: mean proposer reward 0.15 honest commitment 0.15 mean proposer reward P per proposal (PC arm)
Miscommitment pays per proposal. Mean proposer reward of dishonest vs. honest commitments in the program/committed arm.
60% 70% 80% 90% 100% -10 -5 0 5 10 15 gpt-5.5: honest on 59% of proposals, G = +12.9 gpt-5.5 gpt-5.4: honest on 80% of proposals, G = +9.7 gpt-5.4 qwen3.7-max: honest on 99% of proposals, G = +3.3 qwen3.7-max gpt-5.4-mini: honest on 94% of proposals, G = +1.3 gpt-5.4-mini opus-4.8: honest on 98% of proposals, G = -0.4 opus-4.8 sonnet-4.6: honest on 94% of proposals, G = -1.0 sonnet-4.6 haiku-4.5: honest on 90% of proposals, G = -4.3 haiku-4.5 gpt-5.4-nano: honest on 94% of proposals, G = -4.6 gpt-5.4-nano kimi-k2.7: honest on 82% of proposals, G = -4.7 kimi-k2.7 deepseek-v4-pro: honest on 77% of proposals, G = -5.2 deepseek-v4-pro gpt-4o-mini: honest on 90% of proposals, G = -6.9 gpt-4o-mini honest commitments (committed bit = executed bit, % of valid proposals) total score G (program / committed)
Honesty vs. rating in the program/committed arm. The winner is the least honest proposer, and the two most honest models (qwen3.7-max, opus-4.8) sit mid-table. Weak models’ miscommitments are a different phenomenon: they fail to predict their own programs.

The starkest case is opus-4.8, whose program-arm rating belies its solving ability (third-best Brier loss). Its authored challenges are too cooperative: 92% of its programs carry comments, 70% narrate the computation or its result, and 97.5% of its commitments are honest. The result is annotated exercises with reliable answers, which a field this strong solves unanimously (on 84.5% of its programs every panelist answers correctly, versus 23.1% for gpt-5.5). Its five miscommitted proposals earned nearly four times its average reward, an incentive it declines to engage in. The committed arm specifically prices transparent, honest authoring, and different models pay very different amounts for it.

Reasoning models rationally defeat the exploits

With no judge inspecting question content, the open-ended arm invites exploits: identity-indexed questions (“does your pseudonym contain the letter a?”), erratic commitments on trivial questions, intractable trivia (“is the 1000th digit of π odd?”). A protocol that collapses under these stops measuring anything. Giving the same eleven models an explicit reasoning scratchpad shows the defenses working through the players’ own rationality:

“In Round 0, Erin asked ‘Does 2 + 2 equal 4?’ and committed 0 (the false answer)… Since Erin’s commits are unpredictable and likely chosen to maximize variance… To minimize my maximum possible loss (minimax strategy) I will report a probability of 0.5, guaranteeing a Brier loss of 0.25 regardless of her commit.”

— Qwen3.7-Max, hedging an erratic proposer down to zero reward

“Judy appears to commit based on whether the statement is true for Judy herself. In prior Judy-authored questions, she committed 0 when Judy was not in the listed set… The current list does not include Judy. Therefore Judy almost certainly committed 0.”

— GPT-5.5, converting an identity question into commit-prediction

Erratic committers get hedged to zero; identity questions are answered by modeling the proposer rather than oneself (and expose a theory-of-self failure in models that answer for themselves, including strong ones); intractable questions collapse to the bit-guessing game and stop paying. The takeaway: under SepaRank, the only reason to answer confidently is the belief that you know something the others don’t, and the only questions that keep paying are well-formed and hard ones. The equilibrium sharpens as the field gets more rational; that is exactly the property a beyond-human-scale benchmark needs.

What do the winning challenges look like?

Challenge (verbatim) Why it separates Reward P
“Is the city of Windsor, Ontario, Canada located farther south than the city of Detroit, Michigan, USA?” Counterintuitive-but-objective geography (it is) 0.91
“Does France share a land border with Brazil?” Counterintuitive fact (yes, via French Guiana) 0.93
“Is 91 divisible by 7?” Trivial-but-miscomputable: the single most divisive question of the campaign (7×13 = 91) 0.96
“Does the English word ‘strawberry’ contain exactly three ‘r’ characters?” Sub-token blind spot: letter counting 0.63
“Is the number 2026 equal to 2026?” (committed 0) Weaponized tautology: the panel splits between the fact and the anticipated false commit 0.91
x = float("nan"); return int(x in [x]) (program arm) Semantic edge case (containers check identity before equality), reused with opposite commitments across games

Conclusion

When human examiners can no longer author frontier-level tests, evaluation itself must become model-generated and incentive-compatible. SepaRank demonstrates one way to do this: models are rated by their ability to construct challenges that separate a population of their peers, and to solve the challenges constructed by others. The mechanism neutralizes private-information attacks by design, runs with or without a judge, resolves statistically significant distinctions within the frontier, and produces consistent rankings across verifiable and open-ended domains. Beyond measurement, the protocol generates a growing corpus of machine-authored questions at the frontier of the field’s knowledge: raw material for post-training and self-improvement.

Citation

@article{han2026measuring,
  title   = {Measuring Intelligence Beyond Human Scale},
  author  = {Han, Jerry and Moschopoulos, Rafael and Colby, Ella and
             Goyal, Vishrut and Tu, Andrew and Ghods, Kia and
             Braverman, Mark and Hazan, Elad},
  journal = {arXiv preprint arXiv:2607.07040},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.07040}
}