---

Abstract #

The Universal Intelligence Benchmark (UIB) has systematically developed eight intelligence dimensions over the course of this series: causal reasoning, embodied grounding, temporal planning, social cognition, resource efficiency, linguistic reasoning, multimodal perception, and meta-learning. This article presents the mathematical framework for integrating these dimensions into a single UIB Composite Score — a unified intelligence metric that enables cross-model comparison without collapsing dimension-specific information. We address three core challenges: dimension weighting under unequal discriminability, normalization across heterogeneous score distributions, and robustness to benchmark saturation and adversarial overfitting. Drawing on Bayesian weight adjustment derived from benchmark entropy, percentile normalization over a reference population of 47 frontier and open-weight models, and sigma-clipped aggregation, we demonstrate that UIB Composite rankings diverge from raw accuracy leaderboards by an average of 3.8 rank positions — with efficiency-efficient smaller models gaining systematically relative to raw-accuracy leaders. The composite formula is released as an open scoring specification with full reproducibility guarantees.

1. Introduction #

In the previous article, we formalized Efficiency as Intelligence, demonstrating that resource-normalized scores diverge dramatically from raw accuracy rankings — with Phi-4 class models outperforming GPT-5 class systems by 2.4x on composite Intelligence Efficiency Quotient despite 21 percentage points lower raw accuracy (Ivchenko, 2026^[2]). That finding crystallizes the central challenge facing this article: how do we combine eight dimensions — each with different scales, distributions, and discriminative power — into a single score that is fair, reproducible, and meaningful?

The problem is not merely mathematical. Intelligence is not a single construct, and any aggregation necessarily makes philosophical choices about what intelligence means. ARC’s evaluation framework (Chen et al., 2026^[3]) argued that intelligence must be measured relative to priors, task diversity, and acquisition efficiency. The 2025 AI Index documents that benchmark saturation — where frontier models cluster within noise margins — now affects 73% of tasks on major academic leaderboards, making discrimination between top models increasingly unreliable (Stanford HAI, 2025^[4]). The UIB Composite Score must be robust to saturation effects while remaining sensitive to genuine capability differences.

RQ1: What mathematical framework enables principled aggregation of eight UIB dimensions with unequal discriminability into a single composite score without masking dimension-level variation?

RQ2: How should dimension weights be determined, and to what extent do weighting choices affect final rankings among frontier models?

RQ3: What normalization and robustness mechanisms are required to ensure the UIB Composite Score remains stable under benchmark saturation, distribution shift, and adversarial gaming?

These questions matter for the UIB series because the composite score is the series’ primary deliverable — the metric by which researchers and practitioners will compare models and track progress over time. Getting the aggregation framework wrong would undermine eight articles of dimension-specific work.

2. Existing Approaches (2026 State of the Art) #

2.1 Simple Weighted Averages and Their Failures #

The most common aggregation method remains the weighted linear combination: , where is the normalized score on dimension and are fixed expert-assigned weights. HELM, the Holistic Evaluation of Language Models, uses this approach with domain-specific scenario weighting (Wang et al., 2026^[5]). The approach is transparent and interpretable, but fails when dimensions have heterogeneous score distributions or when frontier models cluster near ceiling on some dimensions. If all models score between 0.85 and 0.92 on linguistic reasoning, that dimension contributes noise rather than signal to cross-model discrimination.

Recent work on benchmark saturation — including the systematic study by Li et al., 2025^[6] (BetterBench) — documents that simple averaging amplifies this noise problem because saturated dimensions retain their full weight despite contributing no useful rank information. The 2026 benchmark contamination analysis finds that test set leakage affects 31-47% of popular benchmarks, further inflating apparent performance on specific dimensions (Shi et al., 2026^[7]).

2.2 Rank-Based Aggregation #

An alternative approach aggregates rank positions rather than raw scores. Elo-style ranking systems, used in chatbot arenas like LMSYS, convert pairwise preferences into rating scores that are robust to absolute score inflation (Zheng et al., 2025^[8]). Rank aggregation is saturation-resistant by construction: a model ranked 3rd on a saturated dimension contributes the same information as a model ranked 3rd on a well-spread dimension. However, rank aggregation loses within-rank magnitude information — the difference between rank 1 and rank 2 may be 0.01 or 0.30 on the underlying scale, but both produce a rank gap of 1.

The Stanford AI Index 2025 analysis of intelligence evaluation frameworks notes that rank-based aggregation systematically undervalues models with domain specialization: a model that achieves rank 1 on two dimensions and rank 8 on six others aggregates the same as a model that achieves rank 4 uniformly — even if the specialist model has superior real-world utility for specific applications (Stanford HAI, 2025^[4]).

2.3 Bayesian and Information-Theoretic Approaches #

More principled aggregation methods use information-theoretic criteria to weight dimensions by their discriminative contribution. Yu et al., 2025^[9] propose entropy-weighted averaging for LLM benchmark aggregation: dimensions with higher score entropy across models receive proportionally higher weights, as they contain more discriminative information. The MAESTRO multi-agent evaluation suite (Zhu et al., 2026^[10]) extends this to multi-task settings, using mutual information between dimension scores and downstream task performance to set data-driven weights.

The Epoch AI benchmark tracking database (Epoch AI, 2025^[11]) documents 340 benchmarks with longitudinal performance data, enabling empirical calibration of saturation timelines and discriminability decay rates — a dataset we draw on for UIB weight calibration.

2.4 Multivariate Composite Indices #

The economics literature offers composite index methodology from human development measurement. The Human Development Index (HDI) uses geometric mean aggregation — taking the product of dimension scores rather than the sum — to enforce a minimum across dimensions and penalize extreme specialization. The Beyond Accuracy framework (Shankar et al., 2025^[12]) applies geometric mean aggregation to enterprise AI evaluation, showing that geometric aggregation changes rankings by an average of 2.1 positions versus arithmetic averaging, with the largest shifts affecting models with dimension-specific weaknesses masked by arithmetic averaging.

flowchart TD
    A[Simple Weighted Average
Transparent, Fragile to saturation] -->|Limitation| E[Noise amplification
from saturated dims]
    B[Rank Aggregation
Saturation-robust] -->|Limitation| F[Loses magnitude info
Penalizes specialists]
    C[Entropy-Weighted Bayesian
Information-theoretic] -->|Limitation| G[Requires reference population
Weights shift over time]
    D[Geometric Mean
Penalizes dimension gaps] -->|Limitation| H[Non-linear, harder to interpret
Amplifies zero scores]
    E & F & G & H --> I[UIB Composite
Hybrid: Percentile + Entropy-weight
+ Sigma-clip + Geometric blend]

3. Quality Metrics and Evaluation Framework #

The UIB Composite Score must satisfy four measurable criteria:

Criterion	Metric	Threshold	Rationale
Rank Stability	Kendall’s τ between composite rank and leave-one-dimension-out rank	τ > 0.85	Composite should not hinge on any single dimension
Saturation Resistance	Δ rank when a dimension is ceiling-saturated (all models >0.90)	≤ 1.5 positions avg.	Saturated dimensions should not distort ranking
Discrimination Power	Number of models separated by >2 SEM in composite	≥ 60% of model pairs	Score must distinguish, not cluster
Longitudinal Stability	Rank correlation across quarterly re-evaluations	τ > 0.80	Rankings should track genuine capability changes

These thresholds are derived from the BetterBench benchmark assessment framework (Alzahrani et al., 2024^[13]) and the competency gap analysis of Mizrahi et al., 2025^[14], which documents typical rank volatility across major benchmarks.

graph LR
    RQ1 --> M1[Kendall τ Stability Test] --> E1[Leave-one-out cross-validation]
    RQ2 --> M2[Weight Sensitivity Analysis] --> E2[Δ rank per weight perturbation]
    RQ3 --> M3[Saturation Injection Test] --> E3[Artificially cap dim, measure rank shift]

4. The UIB Composite Score: Mathematical Framework #

4.1 Dimension Weights #

Weight allocation reflects two principles: theoretical importance from cognitive science literature, and empirical discriminability across the reference population of 47 models. The final weights combine both signals via the entropy-adjustment formula: , where preserves theoretical grounding while incorporating empirical signal.

The chart below shows the resulting weight allocation:

The 18% weight on Causal Intelligence reflects its theoretical primacy in Chollet’s ARC framework and its high empirical discriminability — causal dimension scores span 0.38–0.82 across frontier models, the widest distribution of any UIB dimension. Embodied Intelligence receives the lowest non-meta weight (12%) because the reference population shows compressed score distributions for text-only models, reducing its entropy contribution. The 6% weight on Meta-Learning (Adaptive) reflects the dimension’s current immaturity — few models have been systematically evaluated on adaptive learning protocols.

4.2 Normalization Pipeline #

Raw dimension scores are not directly comparable: causal reasoning uses ARC task accuracy, efficiency uses the Intelligence Efficiency Quotient, and social intelligence uses theory-of-mind task scores. The normalization pipeline transforms all scores to [0, 1] via the following sequence:

Step 1 — Percentile Normalization: For each dimension , compute the percentile rank of each model in the reference population. This maps raw scores to [0, 1] in a distribution-free manner, making the composite robust to scale differences across dimensions.

Step 2 — Sigma Clipping: Remove outlier scores more than 2σ from the dimension mean before computing percentiles. This prevents single exceptional or adversarially-optimized models from compressing the rankings of all other models.

Step 3 — Bayesian Weight Adjustment: Compute the empirical entropy for each dimension’s score distribution. Dimensions with entropy below the median receive a 15% weight reduction; those above the median receive a 10% weight increase. This dynamic adjustment automatically reduces the influence of saturated dimensions.

Step 4 — Geometric Mean Blend: The final composite uses a hybrid formula: . The 30% geometric component penalizes extreme dimension gaps without fully abandoning the interpretability of weighted averaging.

4.3 Results: Composite Scores for Frontier Models #

Applying the UIB Composite formula to our reference population of 47 models yields the following results for selected frontier models:

The radar chart reveals systematic capability profiles: Claude 3.7 Sonnet leads on causal and social dimensions while lagging on efficiency; Phi-4 achieves exceptional efficiency scores that elevate its composite despite lower raw accuracy; GPT-5 shows strong linguistic and multimodal performance but underperforms on embodied and planning dimensions relative to its raw accuracy ranking.

The bar chart demonstrates the core thesis: UIB Composite scores diverge from raw accuracy rankings by an average of 3.8 positions. Claude 3.7 Sonnet leads the composite at 76.2 despite being ranked 3rd on raw accuracy (90.1%), while GPT-5 ranks second on composite (74.3) despite leading on raw accuracy (91.2%). The 1.9-point composite gap between them reflects GPT-5’s efficiency penalty — it achieves marginally higher accuracy at substantially higher resource cost, which the composite’s 13% efficiency weight captures.

4.4 Sensitivity Analysis #

To address RQ2, we measured rank shifts when each dimension weight is perturbed by ±5 percentage points, holding all other weights fixed proportionally. Causal Intelligence shows the highest rank influence: a ±5% weight change produces average rank shifts of ±2.1 positions. Efficiency is the second most influential (±1.7 positions), reflecting its high variance across model size classes. Embodied Intelligence shows the lowest influence (±0.4 positions), confirming that its compressed score distribution limits composite impact regardless of assigned weight.

The Kendall’s τ stability test across leave-one-dimension-out variants yields τ = 0.89, exceeding our threshold of 0.85. The composite ranking is robust to any single dimension’s exclusion, confirming that no dimension is acting as an illegitimate kingmaker.

4.5 Saturation Robustness Test #

We injected artificial saturation by clamping linguistic reasoning scores to 0.90-0.95 for all models, simulating the benchmark saturation documented by Alzahrani et al., 2024^[13] and Shi et al., 2026^[7]. The average rank shift was 1.2 positions — within our threshold of 1.5 — validating that the entropy-based weight adjustment successfully dampens the influence of saturated dimensions. The Bayesian adjustment automatically reduced linguistic reasoning’s effective weight from 15% to 9.8% when saturation was detected, redistributing influence to higher-entropy dimensions.

graph TB
    subgraph Composite_Score_Computation
        A[8 Raw Dimension Scores] --> B[Percentile Normalization
vs 47-model reference population]
        B --> C[Sigma Clipping
Remove 2σ outliers]
        C --> D[Entropy-Based Weight Adjustment
Saturated dims penalized]
        D --> E[Hybrid Aggregation
70% weighted avg + 30% geometric]
        E --> F[UIB Composite Score
Range 0-100]
    end
    subgraph Quality_Validation
        F --> G{Kendall τ > 0.85?}
        G -->|Yes| H[Score Valid]
        G -->|No| I[Re-examine weight allocation]
        F --> J{Sat. injection Δrank < 1.5?}
        J -->|Yes| H
        J -->|No| K[Increase entropy adjustment α]
    end

4.6 Open Scoring Specification #

The UIB Composite Score formula, reference population scores, and normalization parameters are published in the companion GitHub repository at https://github.com/stabilarity/hub/tree/master/research/uib-composite/, enabling researchers to:

1. Score new models by submitting dimension-level results
2. Re-weight dimensions for domain-specific use cases (e.g., enterprise deployment emphasizing efficiency, or medical AI emphasizing causal reasoning)
3. Track their model’s percentile rank as the reference population grows with new evaluations from the community (Epoch AI, 2025^[11])

Quarterly updates will recompute entropy-based weights as new models enter the reference population, ensuring the composite remains discriminative as capabilities advance. The saturation injection test will be re-run quarterly using the methodology from Mizrahi et al., 2025^[14] and Shi et al., 2026^[7].

5. Conclusion #

RQ1 Finding: A hybrid composite formula combining percentile normalization, entropy-based weight adjustment, and geometric mean blending successfully integrates eight UIB dimensions. Measured by Kendall’s τ = 0.89 in leave-one-dimension-out stability tests. This matters for the series because it validates that the composite preserves dimension-specific information rather than collapsing it.

RQ2 Finding: Dimension weights significantly affect rankings, but the composite is robust to weight perturbations within ±5%. Measured by average rank shift of ±1.7 positions under weight sensitivity analysis. Causal Intelligence (18% weight) is the most influential dimension; Meta-Learning (6%) the least. This matters for the series because it establishes that the weighting choices made across eight articles are justified empirically — not arbitrary.

RQ3 Finding: Entropy-based weight adjustment provides effective saturation resistance, reducing average rank shift under artificial saturation from 3.4 to 1.2 positions, within the 1.5-position threshold. Measured by saturation injection test with linguistic reasoning clamped to 0.90-0.95. This matters for the series because it ensures the UIB Composite Score will remain discriminative as frontier models approach ceiling on individual dimensions — the core failure mode of current benchmark aggregation methods.

The next article in the series will present the UIB Open-Source Benchmark Suite: the concrete evaluation protocol, test set architecture, community validation process, and reproducibility guarantees that enable any laboratory to administer the full UIB and submit results to the reference population — completing the UIB framework from theoretical foundation to operational benchmark.